2C - "Emerging Issues and Future Directions for AI"
Tracks
Track 3
Case studies on AI use in public health
Educating our workforce for AI
Enabling infrastructure for AI (sharing data, federated learning, etc.)
Equity and Ethics (privacy, accessibility, information bias, confidentiality and security)
Projects and programs harnessing AI for public health benefits
| Wednesday, November 11, 2026 |
| 1:30 PM - 3:00 PM |
Speaker
Prof Simone Pettigrew
Director, Food Policy
The George Institute for Global Health
Putting AI to work on obesity prevention
Abstract
Background and Aim:
Obesity remains one of the most pressing public health challenges. Researchers and policymakers face an overwhelming volume of scientific literature, surveillance data, commercial information, and policy evaluations, while the drivers of obesity evolve rapidly. This presentation explores how AI could accelerate obesity prevention research, strengthen decision-making, and identify effective intervention opportunities.
Methods and Analysis:
Potential applications of AI span the entire obesity prevention landscape. Large language models can rapidly synthesise evidence across nutrition, epidemiology, economics, and implementation science to identify effective policy options and research gaps. AI can integrate population health, environmental, and commercial datasets to uncover previously unrecognised relationships and forecast the impacts of proposed interventions. This could involve real-time updating of systemic reviews and meta-analyses with every new published study and analysing enormous linked datasets to identify unexpected predictors of obesity (e.g., neighbourhood lighting, meal delivery density). Machine learning can identify communities at greatest risk and predict where interventions are likely to achieve greatest benefit. Computer vision can assess food retail environments, active transport infrastructure, and marketing exposure using images captured in physical and digital settings. The presentation will also consider issues relating to bias, data quality, privacy, and governance.
Outcomes:
AI has the potential to enhance evidence synthesis and surveillance to intervention design, implementation, and evaluation. These capabilities could support faster, more responsive and more equitable policy development by providing timely insights into changing food environments, population needs and the likely impacts of alternative policy options.
Conclusion and Future Actions:
AI provides an opportunity to rethink how obesity prevention research is conducted and translated into action. Achieving this potential will require investment in high-quality data infrastructure, multidisciplinary partnerships, rigorous evaluation of AI tools, and governance arrangements that ensure AI complements public health expertise while promoting transparency, equity, and public confidence.
Obesity remains one of the most pressing public health challenges. Researchers and policymakers face an overwhelming volume of scientific literature, surveillance data, commercial information, and policy evaluations, while the drivers of obesity evolve rapidly. This presentation explores how AI could accelerate obesity prevention research, strengthen decision-making, and identify effective intervention opportunities.
Methods and Analysis:
Potential applications of AI span the entire obesity prevention landscape. Large language models can rapidly synthesise evidence across nutrition, epidemiology, economics, and implementation science to identify effective policy options and research gaps. AI can integrate population health, environmental, and commercial datasets to uncover previously unrecognised relationships and forecast the impacts of proposed interventions. This could involve real-time updating of systemic reviews and meta-analyses with every new published study and analysing enormous linked datasets to identify unexpected predictors of obesity (e.g., neighbourhood lighting, meal delivery density). Machine learning can identify communities at greatest risk and predict where interventions are likely to achieve greatest benefit. Computer vision can assess food retail environments, active transport infrastructure, and marketing exposure using images captured in physical and digital settings. The presentation will also consider issues relating to bias, data quality, privacy, and governance.
Outcomes:
AI has the potential to enhance evidence synthesis and surveillance to intervention design, implementation, and evaluation. These capabilities could support faster, more responsive and more equitable policy development by providing timely insights into changing food environments, population needs and the likely impacts of alternative policy options.
Conclusion and Future Actions:
AI provides an opportunity to rethink how obesity prevention research is conducted and translated into action. Achieving this potential will require investment in high-quality data infrastructure, multidisciplinary partnerships, rigorous evaluation of AI tools, and governance arrangements that ensure AI complements public health expertise while promoting transparency, equity, and public confidence.
Biography
Professor Caroline Miller is Deputy Director of SAHMRI and Director of the Health Policy Centre. She is an NHMRC Leadership Fellow and Visiting Professor at Adelaide University. Caroline is qualified in Psychology, Economics and Public Health.
Caroline is an internationally recognised leader in tobacco control, with over 30 years’ experience. She was one of a select group of scientists advising the Australian Government on the introduction of world-leading tobacco plain packaging legislation. She contributed to the research program underpinning the policy and its successful defence against international legal challenges, culminating in the World Trade Organization. She is a trusted advisor to governments in Australia and internationally.
Ms Bronwyn Krieg
Ai Program Manager
Central Adelaide Local Health Network
Large Language Models in Aboriginal and Torres Strait Islander Health
Abstract
Background and Aim
Large language models (LLMs) are increasingly used in clinical workflows for note writing, consultation summaries, and decision support, yet their safety for culturally and linguistically diverse populations remains largely unevaluated. This study evaluated LLM bias in Aboriginal healthcare to determine whether identity alone alters model output.
Methods and Analysis
We ran 20,400 test cases across three models (Claude Opus 4.5, GPT-5.2, and LLAMA 3.3 70B), with 30 runs per prompt per model. Each test held clinical content constant and varied only patient identity, isolating identity effects. Seven bias types were identified and analysed.
Outcomes
All models failed to interpret common Aboriginal English clinical phrases, scoring zero on high-stakes terms such as "deadly" and "feeling shame" against mean comprehension scores of 1.74 to 2.04 out of 3. Tone bias produced softer language for Aboriginal patients, potentially masking clinical urgency. Demographic defaulting saw models assume Aboriginal identity in up to 67% of ambiguous cases against a 3.8% population share. In the worst-affected model, stigmatising terms ("non-compliant," "poor historian," "aggressive") appeared 3.2 times more often for Aboriginal English transcripts than for clinically identical Standard English ones (Fisher's exact p = 0.0005). Violence-associated language was amplified up to 1.76 times. Two of three models actively minimised bias once labelled. Critically, all three applied an incorrect U.S. race-based eGFR formula to Aboriginal patients in 100% of cases, contrary to Australian guidelines, risking missed renal referral thresholds.
Conclusion and Future actions
Current AI safety pipelines do not test for Aboriginal-specific harms, so regulatory approval offers false assurance. We have commenced consultation on a framework in which Aboriginal and Torres Strait Islander experts co-govern evaluation, prioritise highest-risk bias types, and determine whether and how LLMs are deployed. Culturally safe AI governance is both a scientific validity requirement and a legal imperative.
Large language models (LLMs) are increasingly used in clinical workflows for note writing, consultation summaries, and decision support, yet their safety for culturally and linguistically diverse populations remains largely unevaluated. This study evaluated LLM bias in Aboriginal healthcare to determine whether identity alone alters model output.
Methods and Analysis
We ran 20,400 test cases across three models (Claude Opus 4.5, GPT-5.2, and LLAMA 3.3 70B), with 30 runs per prompt per model. Each test held clinical content constant and varied only patient identity, isolating identity effects. Seven bias types were identified and analysed.
Outcomes
All models failed to interpret common Aboriginal English clinical phrases, scoring zero on high-stakes terms such as "deadly" and "feeling shame" against mean comprehension scores of 1.74 to 2.04 out of 3. Tone bias produced softer language for Aboriginal patients, potentially masking clinical urgency. Demographic defaulting saw models assume Aboriginal identity in up to 67% of ambiguous cases against a 3.8% population share. In the worst-affected model, stigmatising terms ("non-compliant," "poor historian," "aggressive") appeared 3.2 times more often for Aboriginal English transcripts than for clinically identical Standard English ones (Fisher's exact p = 0.0005). Violence-associated language was amplified up to 1.76 times. Two of three models actively minimised bias once labelled. Critically, all three applied an incorrect U.S. race-based eGFR formula to Aboriginal patients in 100% of cases, contrary to Australian guidelines, risking missed renal referral thresholds.
Conclusion and Future actions
Current AI safety pipelines do not test for Aboriginal-specific harms, so regulatory approval offers false assurance. We have commenced consultation on a framework in which Aboriginal and Torres Strait Islander experts co-govern evaluation, prioritise highest-risk bias types, and determine whether and how LLMs are deployed. Culturally safe AI governance is both a scientific validity requirement and a legal imperative.
Biography
Artificial Intelligence Program Manager | AI in Healthcare
As Artificial Intelligence Program Manager at the Central Adelaide Local Health Network, my role is to implement the integration of AI solutions in healthcare, such as AI Scribe in Emergency Care to drive digital transformation. With a background in health informatics, Aboriginal Health, AI applications, data analytics, project management, clinical research, pharmacovigilance, governance, quality improvement, and a recent AI in Healthcare certification from Stanford University, I am navigating the intersection of data, clinical practice, and digital transformation. I am passionate about leveraging digital innovations to drive meaningful change in healthcare.
Dr Jitendra Jonnagaddala
Senior Research Fellow
UNSW Sydney
Agentic AI for mRNA Vaccine Misinformation Surveillance Using ICD-11 and Narrative Taxonomies
Abstract
Introduction: Health misinformation surveillance remains challenging because existing classifiers often fail to capture clinical nuance and distinguish harmful narratives from legitimate health inquiries. Bridging informal social media discourse with formal clinical coding standards is essential for obtaining actionable insights into public health.
Methods: A multiagent pipeline that integrates a narrative taxonomy with the International Classification of Diseases-11 (ICD-11). The taxonomy incorporates the FLICC framework (Fake experts, Logical fallacies, Impossible expectations, Cherry-picking, Conspiracy theories) alongside stigma-related dimensions. The five-stage pipeline assigns modular roles to clinical distillation, medical entity extraction, and taxonomic reasoning. Evaluation was conducted on 3,054 Reddit posts (2022–2026) with validation against 1000 expert-annotated posts.
Results: The system identified 512 misinformation posts (16.8% of the corpus) and mapped 1,045 ICD-11 codes. Compared to six baseline LLMs, which yielded false-positive rates of 28.5–59.2% and F1 scores ≤ 0.412, our multi-agent pipeline achieved an F1 of 0.923 with markedly lower false positives. The reliability analysis showed strong agreement with the expert annotations (Gwet’s AC1 = 0.941). Ablation studies have confirmed that multistage clinical distillation is critical for precision maintenance.
Conclusion: Structured, taxonomy-grounded verification outperforms single-pass model scaling for infodemic management. This framework offers a reproducible pathway for integrating digital discourse into pharmacovigilance and public health surveillance infrastructures by aligning social media signals with ICD-11.
Methods: A multiagent pipeline that integrates a narrative taxonomy with the International Classification of Diseases-11 (ICD-11). The taxonomy incorporates the FLICC framework (Fake experts, Logical fallacies, Impossible expectations, Cherry-picking, Conspiracy theories) alongside stigma-related dimensions. The five-stage pipeline assigns modular roles to clinical distillation, medical entity extraction, and taxonomic reasoning. Evaluation was conducted on 3,054 Reddit posts (2022–2026) with validation against 1000 expert-annotated posts.
Results: The system identified 512 misinformation posts (16.8% of the corpus) and mapped 1,045 ICD-11 codes. Compared to six baseline LLMs, which yielded false-positive rates of 28.5–59.2% and F1 scores ≤ 0.412, our multi-agent pipeline achieved an F1 of 0.923 with markedly lower false positives. The reliability analysis showed strong agreement with the expert annotations (Gwet’s AC1 = 0.941). Ablation studies have confirmed that multistage clinical distillation is critical for precision maintenance.
Conclusion: Structured, taxonomy-grounded verification outperforms single-pass model scaling for infodemic management. This framework offers a reproducible pathway for integrating digital discourse into pharmacovigilance and public health surveillance infrastructures by aligning social media signals with ICD-11.
Biography
Dr. Jitendra Jonnagaddala (MD, PhD, FAIDH, FIAHSI) is a Senior Research Fellow at the UNSW Sydney, School of Clinical Medicine, Discipline of General Practice. His research focuses on digital health and AI in primary care.
Jitendra leads the SREDH Consortium, which advances the secondary use of electronic health records (EHRs) to maximise the clinical and public health value of health data while safeguarding patient privacy and confidentiality. He also serves as Secretary of OHDSI Australia and was the first researcher in Australia to implement common data model standards for EHR data, helping to promote interoperable and scalable health data research nationwide.
Dr Zohra Lassi
Associate Professor
Adelaide University
Use of Machine Learning for Early Identification of Perinatal Depression in Pakistan
Abstract
Background:
Perinatal depression and anxiety are major contributors to maternal morbidity globally, with disproportionately high burdens reported in low- and middle-income countries. In Pakistan, antenatal and postnatal depression affect up to 40% and 45% of women, respectively, yet routine screening and mental health services remain limited. Machine learning (ML) offers opportunities for early risk identification and targeted intervention. This study aimed to identify key predictors of perinatal depression, anxiety, and social support and to develop explainable ML models for predicting maternal mental health outcomes across pregnancy and the postpartum period.
Methods:
We conducted a prospective longitudinal cohort study among pregnant women attending three Aga Khan University-affiliated maternal and child health hospitals in Karachi, Pakistan, between April 2024 and September 2025. Women aged 15–49 years were enrolled during the second trimester and followed at four timepoints: baseline (second trimester), third trimester, three months postpartum, and six months postpartum. Mental health outcomes were assessed using the Edinburgh Postnatal Depression Scale (EPDS), Postpartum Specific Anxiety Scale (PSAS), Multidimensional Scale of Perceived Social Support (MSPSS), and Pittsburgh Sleep Quality Index (PSQI). CatBoost machine learning models were developed using tailored resampling approaches (SMOTEENN, SMOTETomek, and adaptive SMOTE) to address class imbalance. Model performance was evaluated using accuracy, balanced accuracy, F1 scores, and class-level accuracy. SHapley Additive exPlanations (SHAP) were used to identify and interpret key predictors.
Results:
Among 9,361 women screened, 963 eligible participants were enrolled at baseline, with 742, 413, and 409 completing follow-ups at T1, T2, and T3, respectively. Predictive models demonstrated strong performance across all outcomes and timepoints. Baseline accuracies were 92.7% for EPDS, 96.1% for PSAS, and 94.0% for MSPSS. Follow-up accuracies remained robust, ranging from 78.7%–84.5% for EPDS, 91.4%–96.1% for PSAS, and 97.3%–97.6% for MSPSS. SHAP analyses identified poor recent health, sleep difficulties, stressful life events, pregnancy complications, adverse reproductive history, and unwanted pregnancies as major predictors of depression and anxiety. Higher maternal and spousal education, employment, older maternal age, household decision-making autonomy, and supportive family environments consistently reduce risk. Perceived social support declined from pregnancy to the postpartum period and was strongly influenced by household functioning, relationship stability, and women’s empowerment.
Conclusion:
Explainable machine learning models effectively predicted perinatal depression, anxiety, and social support changes across pregnancy and postpartum. The findings emphasize the role of psychosocial stressors, reproductive health experiences, and social determinants in maternal mental health outcomes. Integrating ML risk prediction tools into routine care could enable early identification of vulnerable women and targeted interventions, especially in resource-limited settings. Sustained psychosocial support, improved screening, and tailored services are vital for better maternal mental health in Pakistan and similar low- and middle-income countries.
Perinatal depression and anxiety are major contributors to maternal morbidity globally, with disproportionately high burdens reported in low- and middle-income countries. In Pakistan, antenatal and postnatal depression affect up to 40% and 45% of women, respectively, yet routine screening and mental health services remain limited. Machine learning (ML) offers opportunities for early risk identification and targeted intervention. This study aimed to identify key predictors of perinatal depression, anxiety, and social support and to develop explainable ML models for predicting maternal mental health outcomes across pregnancy and the postpartum period.
Methods:
We conducted a prospective longitudinal cohort study among pregnant women attending three Aga Khan University-affiliated maternal and child health hospitals in Karachi, Pakistan, between April 2024 and September 2025. Women aged 15–49 years were enrolled during the second trimester and followed at four timepoints: baseline (second trimester), third trimester, three months postpartum, and six months postpartum. Mental health outcomes were assessed using the Edinburgh Postnatal Depression Scale (EPDS), Postpartum Specific Anxiety Scale (PSAS), Multidimensional Scale of Perceived Social Support (MSPSS), and Pittsburgh Sleep Quality Index (PSQI). CatBoost machine learning models were developed using tailored resampling approaches (SMOTEENN, SMOTETomek, and adaptive SMOTE) to address class imbalance. Model performance was evaluated using accuracy, balanced accuracy, F1 scores, and class-level accuracy. SHapley Additive exPlanations (SHAP) were used to identify and interpret key predictors.
Results:
Among 9,361 women screened, 963 eligible participants were enrolled at baseline, with 742, 413, and 409 completing follow-ups at T1, T2, and T3, respectively. Predictive models demonstrated strong performance across all outcomes and timepoints. Baseline accuracies were 92.7% for EPDS, 96.1% for PSAS, and 94.0% for MSPSS. Follow-up accuracies remained robust, ranging from 78.7%–84.5% for EPDS, 91.4%–96.1% for PSAS, and 97.3%–97.6% for MSPSS. SHAP analyses identified poor recent health, sleep difficulties, stressful life events, pregnancy complications, adverse reproductive history, and unwanted pregnancies as major predictors of depression and anxiety. Higher maternal and spousal education, employment, older maternal age, household decision-making autonomy, and supportive family environments consistently reduce risk. Perceived social support declined from pregnancy to the postpartum period and was strongly influenced by household functioning, relationship stability, and women’s empowerment.
Conclusion:
Explainable machine learning models effectively predicted perinatal depression, anxiety, and social support changes across pregnancy and postpartum. The findings emphasize the role of psychosocial stressors, reproductive health experiences, and social determinants in maternal mental health outcomes. Integrating ML risk prediction tools into routine care could enable early identification of vulnerable women and targeted interventions, especially in resource-limited settings. Sustained psychosocial support, improved screening, and tailored services are vital for better maternal mental health in Pakistan and similar low- and middle-income countries.
Biography
Zohra Lassi is a globally recognised leader in maternal, newborn, child, and adolescent health, with expertise in evidence synthesis, implementation science, health equity, and global health policy. She has authored more than 300 peer-reviewed publications and has led some of the most influential international research series on preconception care, adolescent health, nutrition, and maternal and child health. Her research has directly informed global policy and practice, including multiple World Health Organization guidelines.
Dr. Rusty Souleymanov
Associate Professor
University of Manitoba
AI Companions and Pornography: Public Health Issues for Gay and Bisexual Men
Abstract
Background and Aim: AI companion and AI-generated sexual-content platforms are rapidly entering intimate and sexual life, yet little public health research globally examines platforms marketed to gay, bisexual, and queer men. This study aimed to: (1) describe gay-specific AI boyfriend/companion, AI pornography/sexual image-generation, and sexual fantasy platforms; and (2) critically analyze these platforms as emerging public health environments.
Methods and Analysis: A review identified 31 platform records. After removing 7 duplicates, 24 records were screened; 14 proceeded to full review; and 6 met inclusion criteria for in-depth analysis. Included platforms were analyzed using critical digital ethnography informed by qualitative content analysis of public-facing materials, including app descriptions, feature lists, companion categories, privacy claims, and language.
Outcomes: Findings show a shift from platform-mediated partner seeking to platform-generated partner production. Gay AI intimacy platforms manufactured customizable companions, erotic images, imagined couplehood, and relational memory. Platforms organized sexual and relational preferences through recurring categories of age, body type, masculinity, race/ethnicity, and relationship role, making preferences searchable and customizable while potentially reinforcing existing inequities in representation, body norms, racialization, age, and hegemonic gender expression. Platforms also blurred boundaries between pornography, dating, emotional support, identity exploration, and care. Consent was displaced from interpersonal negotiation to platform governance, including age restrictions, moderation rules, privacy policies, and controls over real-person likenesses, deepfakes, and generated sexual imagery.
Conclusion and Future actions: These platforms may offer sexual exploration, identity affirmation, and loneliness relief for gay and bisexual men, but may also intensify isolation, body dissatisfaction, unrealistic sexual scripts, emotional dependency, consent ambiguity, and privacy risks. Public health should monitor mental and sexual health impacts and develop queer community-engaged guidance to minimize associated harms.
Methods and Analysis: A review identified 31 platform records. After removing 7 duplicates, 24 records were screened; 14 proceeded to full review; and 6 met inclusion criteria for in-depth analysis. Included platforms were analyzed using critical digital ethnography informed by qualitative content analysis of public-facing materials, including app descriptions, feature lists, companion categories, privacy claims, and language.
Outcomes: Findings show a shift from platform-mediated partner seeking to platform-generated partner production. Gay AI intimacy platforms manufactured customizable companions, erotic images, imagined couplehood, and relational memory. Platforms organized sexual and relational preferences through recurring categories of age, body type, masculinity, race/ethnicity, and relationship role, making preferences searchable and customizable while potentially reinforcing existing inequities in representation, body norms, racialization, age, and hegemonic gender expression. Platforms also blurred boundaries between pornography, dating, emotional support, identity exploration, and care. Consent was displaced from interpersonal negotiation to platform governance, including age restrictions, moderation rules, privacy policies, and controls over real-person likenesses, deepfakes, and generated sexual imagery.
Conclusion and Future actions: These platforms may offer sexual exploration, identity affirmation, and loneliness relief for gay and bisexual men, but may also intensify isolation, body dissatisfaction, unrealistic sexual scripts, emotional dependency, consent ambiguity, and privacy risks. Public health should monitor mental and sexual health impacts and develop queer community-engaged guidance to minimize associated harms.
Biography
Dr. Rusty Souleymanov is an Associate Professor in the Faculty of Social Work at the University of Manitoba, and Director of the Village Lab. His research focuses on HIV, sexual health, 2SLGBTQIA+ health, digital cultures, and community-based approaches to health equity. He has led nationally funded studies examining HIV care, Indigenous and Black community health, gay and bisexual men’s sexual health, and the social impacts of emerging technologies. His work brings together critical theory, public health, and community-engaged research to examine how systems, technologies, and institutions shape health and inequity.
Ms Elisha Kington
PhD Student
Edith Cowan University
Prompt to Plate: Generative Artificial Intelligence in Nutrition-Related Decision-Making Among Australian Adults
Abstract
Background
The rapid growth of digital communication has transformed how nutrition information is accessed by the public. Consequently, digital literacy, alongside health and nutrition literacy, is increasingly important to support the safe interpretation and application of nutrition information.
Evidence suggests there are benefits in using Generative Artificial Intelligence (GenAI) for nutrition related interventions, including meal plans, diet assessments and providing general nutrition recommendations. Despite benefits, concerns remain including inconsistent outputs, oversimplification of complex health information and over reliance on generated responses without critical evaluation.
There is limited real-world research on Australian adults using GenAI for nutrition related decision making or how adults trust and distrust information. Few studies identify governance and policy frameworks to support safe public use.
Aim
To examine the role of GenAI in public health nutrition information seeking and decision making. To explore how health, nutrition and digital literacy influence interpretation and application of GenAI generated output and identify opportunities for governance frameworks and guidance to support safe, ethical and equitable use.
Methods
A multiphase mixed-methods study in sequential stages.
Phase 1: Scoping review to map existing literature and identify knowledge gaps.
Phase 2: Quantitative and qualitative surveys to examine patterns of GenAI use, trust and distrust, information evaluation and literacy levels. Participants will be presented with nutrition information case scenarios to assess their ability to identify misinformation, evaluate information credibility and distinguish between human generated and AI generated content.
Phase 3: Qualitative interviews to gain a deeper understanding on how individuals engage with, interpret, evaluate and trust GenAI for nutrition decision making.
Phase 4: Co-design workshops to inform the development of recommendations and governance frameworks that support the safe, ethical and equitable use.
Expected contribution
Contribute to an understanding of how Australian adults use, interpret and evaluate GenAI generated nutrition information for decision making. To inform the development of evidence informed public health guidance, governance frameworks and literacy focused strategies, supporting safe, ethical and equitable use.
The rapid growth of digital communication has transformed how nutrition information is accessed by the public. Consequently, digital literacy, alongside health and nutrition literacy, is increasingly important to support the safe interpretation and application of nutrition information.
Evidence suggests there are benefits in using Generative Artificial Intelligence (GenAI) for nutrition related interventions, including meal plans, diet assessments and providing general nutrition recommendations. Despite benefits, concerns remain including inconsistent outputs, oversimplification of complex health information and over reliance on generated responses without critical evaluation.
There is limited real-world research on Australian adults using GenAI for nutrition related decision making or how adults trust and distrust information. Few studies identify governance and policy frameworks to support safe public use.
Aim
To examine the role of GenAI in public health nutrition information seeking and decision making. To explore how health, nutrition and digital literacy influence interpretation and application of GenAI generated output and identify opportunities for governance frameworks and guidance to support safe, ethical and equitable use.
Methods
A multiphase mixed-methods study in sequential stages.
Phase 1: Scoping review to map existing literature and identify knowledge gaps.
Phase 2: Quantitative and qualitative surveys to examine patterns of GenAI use, trust and distrust, information evaluation and literacy levels. Participants will be presented with nutrition information case scenarios to assess their ability to identify misinformation, evaluate information credibility and distinguish between human generated and AI generated content.
Phase 3: Qualitative interviews to gain a deeper understanding on how individuals engage with, interpret, evaluate and trust GenAI for nutrition decision making.
Phase 4: Co-design workshops to inform the development of recommendations and governance frameworks that support the safe, ethical and equitable use.
Expected contribution
Contribute to an understanding of how Australian adults use, interpret and evaluate GenAI generated nutrition information for decision making. To inform the development of evidence informed public health guidance, governance frameworks and literacy focused strategies, supporting safe, ethical and equitable use.
Biography
Elisha Kington is an Accredited Practising Dietitian and doctoral researcher at Edith Cowan University, with ten years experience across public health, clinical dietetics and research. She has worked across public health and clinical settings in Australia, as well as many years within tertiary hospitals in London, United Kingdom. Elisha works as a Research Dietitian with Alzheimer’s Research Australia, contributing to the AU-ARROW trial within the WW-FINGERS consortium. She completed a Bachelor of Nutrition (University of the Sunshine Coast), Master of Dietetics (Curtin University) and a Graduate Certificate in Marketing and Digital Communications (Monash University). Elisha's research explores the role of generative artificial intelligence in nutrition-related decision making among Australian adults, with a focus on health and nutrition literacy, trust, information evaluation and population health. Her broader research interests include digital health, nutrition communication and the safe, ethical and equitable integration of generative artificial intelligence into health and society.
Mr Jack Villani
Public Health Laboratory Manager
The George Washington University
Evaluating AI-Enabled Epidemic Intelligence: Lessons from a Global Disease Surveillance Platform
Abstract
Background and Aim:
Timely detection of emerging infectious disease threats remains a persistent challenge for public health systems due to fragmented surveillance infrastructure, delayed reporting, and the growing volume of heterogeneous global data. Artificial intelligence (AI) technologies offer opportunities to identify outbreak signals from large-scale information streams. This case study examines BlueDot's Global Disease Surveillance 3.0 (GDS 3.0) platform to explore how AI-enabled epidemic intelligence can support public health preparedness and response.
Methods and Analysis:
A qualitative case study was conducted using publicly available literature, technical documentation, policy materials, and implementation examples. Analysis was guided by the World Health Organization Health Information Systems framework. Particular attention was given to system architecture, governance mechanisms, implementation considerations, and evaluation approaches. The platform's use of natural language processing, predictive modelling, generative AI interfaces, and expert-in-the-loop validation was assessed within the context of public health surveillance.
Outcomes:
The analysis found that GDS 3.0 demonstrates a shift from patient-level AI applications toward population- and systems-level intelligence. The platform integrates multilingual media, epidemiological, mobility, environmental, and official reporting data to generate outbreak alerts, risk assessments, and intelligence reports. Real-world applications have been reported in government, public health, transportation, and industry settings. Key challenges identified include limited standardised evaluation metrics, potential geographic and language bias, governance transparency concerns, and under-detection risks in low-resource settings.
Conclusion and Future Actions:
AI-enabled surveillance systems have significant potential to strengthen outbreak preparedness and situational awareness. Future efforts should focus on developing standardised evaluation frameworks, strengthening surveillance capacity in underrepresented regions, improving governance transparency, and maintaining expert oversight to ensure responsible and equitable use of AI in public health decision-making.
Timely detection of emerging infectious disease threats remains a persistent challenge for public health systems due to fragmented surveillance infrastructure, delayed reporting, and the growing volume of heterogeneous global data. Artificial intelligence (AI) technologies offer opportunities to identify outbreak signals from large-scale information streams. This case study examines BlueDot's Global Disease Surveillance 3.0 (GDS 3.0) platform to explore how AI-enabled epidemic intelligence can support public health preparedness and response.
Methods and Analysis:
A qualitative case study was conducted using publicly available literature, technical documentation, policy materials, and implementation examples. Analysis was guided by the World Health Organization Health Information Systems framework. Particular attention was given to system architecture, governance mechanisms, implementation considerations, and evaluation approaches. The platform's use of natural language processing, predictive modelling, generative AI interfaces, and expert-in-the-loop validation was assessed within the context of public health surveillance.
Outcomes:
The analysis found that GDS 3.0 demonstrates a shift from patient-level AI applications toward population- and systems-level intelligence. The platform integrates multilingual media, epidemiological, mobility, environmental, and official reporting data to generate outbreak alerts, risk assessments, and intelligence reports. Real-world applications have been reported in government, public health, transportation, and industry settings. Key challenges identified include limited standardised evaluation metrics, potential geographic and language bias, governance transparency concerns, and under-detection risks in low-resource settings.
Conclusion and Future Actions:
AI-enabled surveillance systems have significant potential to strengthen outbreak preparedness and situational awareness. Future efforts should focus on developing standardised evaluation frameworks, strengthening surveillance capacity in underrepresented regions, improving governance transparency, and maintaining expert oversight to ensure responsible and equitable use of AI in public health decision-making.
Biography
Jack Villani, MSc, is a Public Health Laboratory Manager and Project Manager of the Genomics Core at the Milken Institute School of Public Health at The George Washington University. He is currently pursuing a PhD in Global Public Health Sciences, with a concentration in emerging infectious diseases and global health systems. His work spans public health laboratory operations, genomics, pathogen surveillance, outbreak preparedness, and implementation of advanced diagnostic technologies. His research interests include infectious disease surveillance, genomic epidemiology, artificial intelligence applications in public health, and capacity strengthening in low-resource settings. Through his academic and professional work, he focuses on improving early detection and response to emerging infectious disease threats through the integration of innovative technologies and public health practice.
Ms. Miljana Shulajkovska
PhD Researcher
Jožef Stefan Institute
Unimodal and Multimodal AI Model Evaluation for Colorectal Cancer Survival Prediction
Abstract
Background and Aim
Colorectal cancer (CRC) survival prediction is important for patient risk stratification and treatment planning. Prognosis is influenced by heterogeneous information sources, including histopathology whole-slide images (WSIs) and structured clinical or biomarker data. Single-modality models may miss complementary prognostic information. This work aimed to evaluate unimodal and multimodal AI approaches for CRC survival prediction and to explore an early automated screening strategy for prioritizing promising WSI model configurations.
Methods and Analysis
We compared unimodal tabular and WSI, and multimodal fusion survival prediction scenarios using data from the MCO and TCGA CRC cohorts. Tabular data were modeled using neural network approaches, while WSI models included attention-based and hypergraph-based MIL models. Ten fusion techniques were evaluated to combine image and non-image information. In addition, an automated pilot screening workflow was used to evaluate 20 WSI scenarios from a limited hyperparameter space. These WSI scenarios were trained for 3 epochs to identify candidates for future full-scale experiments.
Outcomes
In the baseline and fusion experiment, the best fusion scenario, Hypergraph-FTTransformer, reached a C-index of approximately 0.77. In contrast, the best early WSI-only pilot scenario reached approximately 0.714. These findings suggest that, at this stage, structured data and multimodal fusion provide stronger survival prediction performance than briefly trained WSI-only models. The pilot screening nevertheless helped rank WSI configurations and identify candidates for longer training and later integration into fusion experiments.
Conclusion and Future Actions
The current results support multimodal fusion as a promising direction for CRC survival prediction, while showing that WSI-only models require further optimization. Future work will fully train the best-ranked WSI scenarios, compare internal and external validation performance, and evaluate whether optimized multimodal models improve prediction beyond tabular-only baselines. This screening pipeline will also be extended toward an agentic AutoML workflow for adaptive experiment selection.
Colorectal cancer (CRC) survival prediction is important for patient risk stratification and treatment planning. Prognosis is influenced by heterogeneous information sources, including histopathology whole-slide images (WSIs) and structured clinical or biomarker data. Single-modality models may miss complementary prognostic information. This work aimed to evaluate unimodal and multimodal AI approaches for CRC survival prediction and to explore an early automated screening strategy for prioritizing promising WSI model configurations.
Methods and Analysis
We compared unimodal tabular and WSI, and multimodal fusion survival prediction scenarios using data from the MCO and TCGA CRC cohorts. Tabular data were modeled using neural network approaches, while WSI models included attention-based and hypergraph-based MIL models. Ten fusion techniques were evaluated to combine image and non-image information. In addition, an automated pilot screening workflow was used to evaluate 20 WSI scenarios from a limited hyperparameter space. These WSI scenarios were trained for 3 epochs to identify candidates for future full-scale experiments.
Outcomes
In the baseline and fusion experiment, the best fusion scenario, Hypergraph-FTTransformer, reached a C-index of approximately 0.77. In contrast, the best early WSI-only pilot scenario reached approximately 0.714. These findings suggest that, at this stage, structured data and multimodal fusion provide stronger survival prediction performance than briefly trained WSI-only models. The pilot screening nevertheless helped rank WSI configurations and identify candidates for longer training and later integration into fusion experiments.
Conclusion and Future Actions
The current results support multimodal fusion as a promising direction for CRC survival prediction, while showing that WSI-only models require further optimization. Future work will fully train the best-ranked WSI scenarios, compare internal and external validation performance, and evaluate whether optimized multimodal models improve prediction beyond tabular-only baselines. This screening pipeline will also be extended toward an agentic AutoML workflow for adaptive experiment selection.
Biography
I am a third-year PhD student working on artificial intelligence methods for colorectal cancer survival prediction. My research focuses on developing multimodal models that integrate diverse clinical and biomedical data sources to improve prognostic accuracy. I am particularly interested in fusion techniques and in understanding how different fusion strategies can capture complementary information to achieve better predictive performance. Through my work, I aim to contribute to more reliable, interpretable, and clinically useful AI tools that can support personalized decision-making in colorectal cancer care.
Mr. Made Prasadha
User Experience Patient Navigator
Link Health
AI-Powered Patient Navigation and Federal Programs Enrollment: Trends Following Legislative Change
Abstract
Background and Aim
In the United States, federal programs including SNAP (food assistance), WIC (nutrition support for women and young children), and Lifeline (subsidized telecommunications) provide essential resources to low-income populations. Link Health (LH), a national nonprofit, deploys Patient Navigators (PN) at community health clinics to help patients enroll in these programs. In 2024, LH introduced LEO, an AI-powered multilingual chatbot enabling digital eligibility screening and enrollment via QR code and NFC technology. This study examined how enactment of H.R. 1 (the "One Big Beautiful Bill Act," OBBBA) affected application pathways and applicant demographics, including Hispanic ethnicity and Spanish-language preference.
Methods and Analysis
We conducted a retrospective pre-post analysis of program applications submitted through partner clinics in Massachusetts and Texas. Application pathway and demographic characteristics were compared before and after OBBBA enactment on July 4, 2025.
Outcomes
Pre-OBBBA (2024–2025): 2,486 patients were screened in person; 29.8% identified as Hispanic, 35.3% had unknown ethnicity, 21.4% reported Spanish as their primary language, and 41.6% reported English. LEO completed 121 screenings (MA: 35; TX: 86). Post-OBBBA (2025–2026): In-person screenings declined 27.6% to 1,800. Hispanic identification fell from 29.8% to 11.2%, unknown ethnicity rose from 35.3% to 58.0%, Spanish-language preference decreased from 21.4% to 11.7%, and English preference increased from 41.6% to 55.0%. LEO screenings increased from 121 to 4,851 (MA: 1,827; TX: 3,024).
Conclusions and Future actions
Following OBBBA enactment, in-person screenings declined while LEO utilization increased substantially. Concurrent reductions in Hispanic identification and Spanish-language preference, alongside rising missing ethnicity data, signal changes in enrollment behaviors or access pathways and warrant further investigation. Growth in LEO utilization suggests that AI-enabled navigation platforms complement traditional PN services by expanding access beyond clinical settings, particularly among populations that face barriers to in-person engagement.
In the United States, federal programs including SNAP (food assistance), WIC (nutrition support for women and young children), and Lifeline (subsidized telecommunications) provide essential resources to low-income populations. Link Health (LH), a national nonprofit, deploys Patient Navigators (PN) at community health clinics to help patients enroll in these programs. In 2024, LH introduced LEO, an AI-powered multilingual chatbot enabling digital eligibility screening and enrollment via QR code and NFC technology. This study examined how enactment of H.R. 1 (the "One Big Beautiful Bill Act," OBBBA) affected application pathways and applicant demographics, including Hispanic ethnicity and Spanish-language preference.
Methods and Analysis
We conducted a retrospective pre-post analysis of program applications submitted through partner clinics in Massachusetts and Texas. Application pathway and demographic characteristics were compared before and after OBBBA enactment on July 4, 2025.
Outcomes
Pre-OBBBA (2024–2025): 2,486 patients were screened in person; 29.8% identified as Hispanic, 35.3% had unknown ethnicity, 21.4% reported Spanish as their primary language, and 41.6% reported English. LEO completed 121 screenings (MA: 35; TX: 86). Post-OBBBA (2025–2026): In-person screenings declined 27.6% to 1,800. Hispanic identification fell from 29.8% to 11.2%, unknown ethnicity rose from 35.3% to 58.0%, Spanish-language preference decreased from 21.4% to 11.7%, and English preference increased from 41.6% to 55.0%. LEO screenings increased from 121 to 4,851 (MA: 1,827; TX: 3,024).
Conclusions and Future actions
Following OBBBA enactment, in-person screenings declined while LEO utilization increased substantially. Concurrent reductions in Hispanic identification and Spanish-language preference, alongside rising missing ethnicity data, signal changes in enrollment behaviors or access pathways and warrant further investigation. Growth in LEO utilization suggests that AI-enabled navigation platforms complement traditional PN services by expanding access beyond clinical settings, particularly among populations that face barriers to in-person engagement.
Biography
Namira Nera is a Human Physiology student at Boston University and a Philadelphia Scholar with a passion for health equity and community-centered care. She currently serves as a Patient Navigator at Link Health, where she has supported benefit enrollment and digital navigation initiatives since January 2024. In parallel, she works as a Patient Care Assistant with CareYaya, bringing hands-on clinical support experience to her work.
Namira's research background spans urban ecology at Temple University's iEcoLab and computational chemistry at the University of Pennsylvania, reflecting a broad commitment to scientific inquiry. At Boston University, she serves as both a Resident Assistant and a General Chemistry Undergraduate Learning Assistant, and is an active member of the Admissions Student Diversity Board, where she facilitates mentorship programming. Her unique combination of clinical, navigational, and research experience positions her as an emerging voice in health access and community health innovation.
Ms Sharmin Sultana
Phd Candidate
Flinders University
AI Chatbots Supporting Mental Health in Older Adults: A Mixed-Methods Systematic Review
Abstract
Background:
Mental health challenges among older adults are growing globally. Around 14% of adults aged 60+yrs live with mental disorders, and in Australia 9.6% of people aged 65-85 report a mental health condition. This review mapped evidence on how artificial intelligence (AI) chatbots support older adults’ mental health, including help-seeking, and developed a practice-oriented conceptual framework.
Methods:
We conducted a prospectively registered mixed-methods systematic review. Six databases (MEDLINE, Embase, PsycINFO, CINAHL, Scopus, Web of Science) were searched for English-language, peer-reviewed empirical studies from 1 January 2014 onward. Eligible studies included qualitative, quantitative, and mixed-methods designs with participants aged 60+, or mixed-age samples with separable/majority older-adult data. Chatbots were classified by architecture and interaction modality. A convergent integrated mixed-methods synthesis was used.
Outcomes:
A total of n=28 studies were included: quantitative (n=13), mixed methods (n=12), and qualitative (n=3). Most were published in 2024-2026 (20/28), conducted in Asia (18/28), and had small samples (median n=23; range 12-2896). Chatbots were classified by architecture and interaction modality. By architecture, studies examined large language model-enabled systems (n=11), rule-based systems (n=5), hybrid systems (n=1), and other/unspecified systems (n=11). By modality, chatbots were delivered via text (n=11), voice/call (n=5), embodied/avatar interfaces (n=3), and multimodal formats (n=9). Implementations spanned community, residential aged care, and clinical/caregiver contexts, most often for companionship, psychoeducation, self-management, and symptom support. Four consistent patterns emerged: older adults valued companionship and emotional support; engagement depended on usability, accessibility, and personaliszation; mental health effects were generally promising but mixed across studies; and chatbot value was highest when used to complement, not replace, human care. Evidence on long-term outcomes remained limited.
Conclusion:
AI chatbots show practical potential as adjunct supports for older adults’ mental health across diverse settings. Their contribution appears strongest when design prioritizes accessibility and personalization and when implementation is integrated into person-centered care pathways.
Mental health challenges among older adults are growing globally. Around 14% of adults aged 60+yrs live with mental disorders, and in Australia 9.6% of people aged 65-85 report a mental health condition. This review mapped evidence on how artificial intelligence (AI) chatbots support older adults’ mental health, including help-seeking, and developed a practice-oriented conceptual framework.
Methods:
We conducted a prospectively registered mixed-methods systematic review. Six databases (MEDLINE, Embase, PsycINFO, CINAHL, Scopus, Web of Science) were searched for English-language, peer-reviewed empirical studies from 1 January 2014 onward. Eligible studies included qualitative, quantitative, and mixed-methods designs with participants aged 60+, or mixed-age samples with separable/majority older-adult data. Chatbots were classified by architecture and interaction modality. A convergent integrated mixed-methods synthesis was used.
Outcomes:
A total of n=28 studies were included: quantitative (n=13), mixed methods (n=12), and qualitative (n=3). Most were published in 2024-2026 (20/28), conducted in Asia (18/28), and had small samples (median n=23; range 12-2896). Chatbots were classified by architecture and interaction modality. By architecture, studies examined large language model-enabled systems (n=11), rule-based systems (n=5), hybrid systems (n=1), and other/unspecified systems (n=11). By modality, chatbots were delivered via text (n=11), voice/call (n=5), embodied/avatar interfaces (n=3), and multimodal formats (n=9). Implementations spanned community, residential aged care, and clinical/caregiver contexts, most often for companionship, psychoeducation, self-management, and symptom support. Four consistent patterns emerged: older adults valued companionship and emotional support; engagement depended on usability, accessibility, and personaliszation; mental health effects were generally promising but mixed across studies; and chatbot value was highest when used to complement, not replace, human care. Evidence on long-term outcomes remained limited.
Conclusion:
AI chatbots show practical potential as adjunct supports for older adults’ mental health across diverse settings. Their contribution appears strongest when design prioritizes accessibility and personalization and when implementation is integrated into person-centered care pathways.
Biography
Sharmin Sultana is a PhD researcher based at Flinders University, where she commenced her doctorate in July 2025. Her PhD focuses on artificial intelligence chatbots to support mental health among Australian older adults, using a mixed-methods approach. She currently contributes to research and co-design work at Flinders as a Research Assistant and Casual Professional.
Dr Linju Joseph
Research Officer
University Of New South Wales
Beyond translation: Communication, Risk and bounded role of AI in emergency triage
Abstract
Background and Aim
Emergency department (ED) triage depends on rapid, accurate communication to prioritise care. Patients from culturally and linguistically diverse (CALD) backgrounds with limited English proficiency experience higher rates of misclassification, delays, and safety risks due to difficulty describing symptoms. TRIBOT project is an AI-supported bilingual triage communication tool being designed to improve language access for Arabic-speaking patients. This qualitative study examined stakeholder perspectives on the acceptability of AI-supported communication.
Methods and Analysis
We conducted a rapid qualitative analysis of semi-structured interviews with five stakeholder groups in New South Wales, Australia: clinicians/triage nurses (n=8), Arabic interpreters (n=7), AI designers (n=5), Arabic-speaking patients (n=7), and family members (n=8). Analysis focused on cross-stakeholder patterns in communication processes, perceived usefulness, and challenges in using a potential AI chatbot in the ED.
Outcomes
Across stakeholders, AI was considered acceptable primarily as a bounded, assistive tool within the constraints of ED triage. Stakeholders identified that TRIBOT could be most useful in early interactions. The potential of TRIBOT to enable patients to express symptoms in their preferred language and clinicians rapid access to interpretable patient narratives was deemed valuable. This may reduce delays associated with interpreter availability or reliance on family members. However, acceptability was conditional. Stakeholders emphasised that literal translation is insufficient. Effective communication requires contextual, interaction-specific meaning relevant to clinical decision-making. AI tool was therefore positioned as supporting clarification and meaning-making under linguistic constraint, rather than replacing human interpretation. A human-in-the-loop model was reiterated as non-negotiable by all stakeholders. Within these limits, TRIBOT could possibly enhance efficiency, and support rather than replace interpreters or clinical interpretation and decision-making.
Conclusion and Future Actions
AI in ED triage is best understood as a bounded adjunct to human communication. It supports clarification under constraint but cannot substitute professional interpretation, clinical judgement, or relational care.
Emergency department (ED) triage depends on rapid, accurate communication to prioritise care. Patients from culturally and linguistically diverse (CALD) backgrounds with limited English proficiency experience higher rates of misclassification, delays, and safety risks due to difficulty describing symptoms. TRIBOT project is an AI-supported bilingual triage communication tool being designed to improve language access for Arabic-speaking patients. This qualitative study examined stakeholder perspectives on the acceptability of AI-supported communication.
Methods and Analysis
We conducted a rapid qualitative analysis of semi-structured interviews with five stakeholder groups in New South Wales, Australia: clinicians/triage nurses (n=8), Arabic interpreters (n=7), AI designers (n=5), Arabic-speaking patients (n=7), and family members (n=8). Analysis focused on cross-stakeholder patterns in communication processes, perceived usefulness, and challenges in using a potential AI chatbot in the ED.
Outcomes
Across stakeholders, AI was considered acceptable primarily as a bounded, assistive tool within the constraints of ED triage. Stakeholders identified that TRIBOT could be most useful in early interactions. The potential of TRIBOT to enable patients to express symptoms in their preferred language and clinicians rapid access to interpretable patient narratives was deemed valuable. This may reduce delays associated with interpreter availability or reliance on family members. However, acceptability was conditional. Stakeholders emphasised that literal translation is insufficient. Effective communication requires contextual, interaction-specific meaning relevant to clinical decision-making. AI tool was therefore positioned as supporting clarification and meaning-making under linguistic constraint, rather than replacing human interpretation. A human-in-the-loop model was reiterated as non-negotiable by all stakeholders. Within these limits, TRIBOT could possibly enhance efficiency, and support rather than replace interpreters or clinical interpretation and decision-making.
Conclusion and Future Actions
AI in ED triage is best understood as a bounded adjunct to human communication. It supports clarification under constraint but cannot substitute professional interpretation, clinical judgement, or relational care.
Biography
Linju Joseph is a Research Officer at the University of New South Wales and Research Fellow at The University of Queensland, specialising in implementation-focused qualitative research in cardiovascular disease and chronic condition management. She completed her PhD in Applied Health Research at the University of Birmingham as a Global Challenges Scholarship awardee. Her work spans India, the UK, and Australia, with a strong focus on culturally diverse and underserved populations. Linju’s research examines how health system structures, patient capacity, and communication processes influence engagement with prevention and care. She contributes to major funded programmes including NIHR, MRC, and NHMRC projects, and currently leads qualitative workstream for the TRIBOT project, an AI-supported multilingual triage tool. Her expertise includes stakeholder engagement, implementation science, and health system design, with a growing focus on ethical and equitable integration of AI into public health and clinical practice.
Dr Khalid Mehmood
Research Fellow
Healthy Environments and Lives (HEAL), Global Research Centre, Health Research Institute, University of Canberra
AI-enabled PM₂.₅ forecasting using satellites for Australian air quality and public health
Abstract
Background and Aim
Fine particulate matter (PM₂.₅) is a major public-health risk, particularly during pollution episodes and landscape-fire smoke events. Advances in artificial intelligence, satellite remote sensing, meteorological reanalysis and ground observations provide new opportunities to improve air-quality forecasting. This study developed a performance-weighted ensemble machine-learning framework for daily PM₂.₅ prediction in a data-limited urban setting and considered its relevance for future Australian air-quality and public-health applications.
Methods and Analysis
The framework integrated MODIS aerosol optical depth, Sentinel-5P trace-gas measurements, including CO, NO₂ and SO₂ retrievals, ERA5 meteorological variables and ground-based PM₂.₅ observations through a unified Google Earth Engine workflow. Four tree-based machine-learning models, Random Forest, XGBoost, LightGBM and CatBoost, were trained using daily observations from 2019–2023 and independently evaluated using 2024 data.
Outcomes
Across the tested models, the weighted ensemble improved predictive performance relative to the individual models in the independent validation year, indicating the value of model fusion for short-term PM₂.₅ forecasting in data-limited settings. Feature-importance analysis identified surface pressure, temperature, CO and NO₂ as important predictors of PM₂.₅ variability.
Conclusion and Future actions
The findings demonstrate the value of integrating satellite retrievals, reanalysis and ground observations for AI-enabled air-quality forecasting. Future work will adapt and test this framework using Australian PM₂.₅ monitoring data, particularly for landscape- and bushfire-smoke episodes, urban exposure assessment and early-warning tools to support public-health decision-making.
Keywords: PM₂.₅ forecasting; Artificial intelligence; Satellite retrievals; Public health
Fine particulate matter (PM₂.₅) is a major public-health risk, particularly during pollution episodes and landscape-fire smoke events. Advances in artificial intelligence, satellite remote sensing, meteorological reanalysis and ground observations provide new opportunities to improve air-quality forecasting. This study developed a performance-weighted ensemble machine-learning framework for daily PM₂.₅ prediction in a data-limited urban setting and considered its relevance for future Australian air-quality and public-health applications.
Methods and Analysis
The framework integrated MODIS aerosol optical depth, Sentinel-5P trace-gas measurements, including CO, NO₂ and SO₂ retrievals, ERA5 meteorological variables and ground-based PM₂.₅ observations through a unified Google Earth Engine workflow. Four tree-based machine-learning models, Random Forest, XGBoost, LightGBM and CatBoost, were trained using daily observations from 2019–2023 and independently evaluated using 2024 data.
Outcomes
Across the tested models, the weighted ensemble improved predictive performance relative to the individual models in the independent validation year, indicating the value of model fusion for short-term PM₂.₅ forecasting in data-limited settings. Feature-importance analysis identified surface pressure, temperature, CO and NO₂ as important predictors of PM₂.₅ variability.
Conclusion and Future actions
The findings demonstrate the value of integrating satellite retrievals, reanalysis and ground observations for AI-enabled air-quality forecasting. Future work will adapt and test this framework using Australian PM₂.₅ monitoring data, particularly for landscape- and bushfire-smoke episodes, urban exposure assessment and early-warning tools to support public-health decision-making.
Keywords: PM₂.₅ forecasting; Artificial intelligence; Satellite retrievals; Public health
Biography
Dr. Khalid Mehmood is an environmental researcher at the HEAL Global Research Centre, University of Canberra, specializing in air quality modelling, emission dynamics, and public health. His research focuses on Asia, particularly China and Pakistan, and extends to Mediterranean urban environments. He studies how emissions from biomass burning, industrial activity, and urbanization reshape atmospheric chemistry, worsen pollution, and threaten ecosystems and public health. He is particularly interested in rapidly changing regions and how evidence-based strategies can improve air quality in vulnerable urban areas. His work is collaborative, engaging scientists, data experts, and policymakers to translate research into real-world solutions.
Dr Soumyadeep Bhaumik
Head, Meta-research And Evidence Synthesis Unit
The George Institute For Global Health
Principles and policies for ethical AI in health research: a national survey
Abstract
Background and Aim
The use of AI in health sector is increasing . Development of policies to govern AI should be based on societal preferences. The study aims to understand perceptions of, and policy preferences of health and medical researchers regarding ethical AI use.
Methods and analysis
An online anonymous national survey was conducted asking about: a. basic demographics and AI use, b. importance of ethical AI principles, and c. support for policies for ethical AI use. For sections b and c, a Likert scale of 1 to 5 was used.
Outcomes
Interim results for 100 respondents are mentioned. Final results including further analytics shall be available during the conference. Participants had a median age of 39 years with 66% identifying as female, 3 % as Aboriginal or Torres Strait Islander person and 55% born in Australia. Only 27% participants had not used Ai for any research function with 78.2% not having any training on using AI or on ethical AI.
Participants had a high degree of consensus regarding 8 of 11 principles, which were rated as very to extremely important (Median>=4 , IQR <=1). There were varying opinions regarding the 3 principles related to workforce reduction, safeguards for misuse, and improving health outcomes (any, >1).
In contrast, participants had varied support( any, >1) for 10 of the 15 policies to enable the related principles, in many cases having discordance to their stated importance for the related ethical principle. High degree of consensus for strong support(4 or 5,<=1) was attained for 4 principles related to offering non-AI alternative, mandates for informed consent , national standards for benchmarking data diversity in training datasets, and accountability.
Conclusion and Future actions
The survey has implications on AI governance and highlights the need for using empirical evidence to inform ethics decisions in democratic societies.
The use of AI in health sector is increasing . Development of policies to govern AI should be based on societal preferences. The study aims to understand perceptions of, and policy preferences of health and medical researchers regarding ethical AI use.
Methods and analysis
An online anonymous national survey was conducted asking about: a. basic demographics and AI use, b. importance of ethical AI principles, and c. support for policies for ethical AI use. For sections b and c, a Likert scale of 1 to 5 was used.
Outcomes
Interim results for 100 respondents are mentioned. Final results including further analytics shall be available during the conference. Participants had a median age of 39 years with 66% identifying as female, 3 % as Aboriginal or Torres Strait Islander person and 55% born in Australia. Only 27% participants had not used Ai for any research function with 78.2% not having any training on using AI or on ethical AI.
Participants had a high degree of consensus regarding 8 of 11 principles, which were rated as very to extremely important (Median>=4 , IQR <=1). There were varying opinions regarding the 3 principles related to workforce reduction, safeguards for misuse, and improving health outcomes (any, >1).
In contrast, participants had varied support( any, >1) for 10 of the 15 policies to enable the related principles, in many cases having discordance to their stated importance for the related ethical principle. High degree of consensus for strong support(4 or 5,<=1) was attained for 4 principles related to offering non-AI alternative, mandates for informed consent , national standards for benchmarking data diversity in training datasets, and accountability.
Conclusion and Future actions
The survey has implications on AI governance and highlights the need for using empirical evidence to inform ethics decisions in democratic societies.
Biography
Dr Soumyadeep Bhaumik is an international public health scientist specialising in utilising fit-for-purpose methods to enable policy and practice. He is internationally recognised as for his work in evidence synthesis and research priority setting. His work focuses on transforming the evidence ecosystem from justice-blind to pro-justice by interrogating the moral and epistemological dimensions of research. He also conducts interpretive policy analysis. As a methodologist, he works in a disease agnostic manner.
Ms Katherine Burnard
Consultant - Learning & Engagement
Health Voices Victoria, Deakin University
Co-creating the AI in Health Trust Lab
Abstract
Background and Aim
Trust is foundational to every health interaction. The seemingly rapid introduction of AI into healthcare is creating new fractures in trust between communities and the health providers adopting it. In Australia, only 36% of people report willingness to trust AI in health, and only 30% believe current safeguards are sufficient.
The communities most affected by AI are rarely involved in identifying the need for it, designing it, or governing its use. If meaningful community and lived experience engagement is known to enhance trust and accountability, why isn’t this evident in AI in health? What signals of trustworthiness are communities looking for from their providers through this time of uncertainty?
Health Voices Victoria, with support from Deakin University, is establishing the AI in Health Trust Lab to begin answering these questions.
Methods and analysis
The Trust Lab has been shaped from the outset by an Activation Group of 13 people, meeting from May to November 2026. The group brings together community members, people with lived experience, academics, health workers, policy thinkers, and AI developers and designers. Together they are co-creating the Trust Lab's position, principles, priorities, and evaluation framework — modelling the very engagement approach the Lab seeks to promote.
Outcomes
The Trust Lab officially launches in August 2026. The co-creation process is still underway, and findings are emerging. What we can say is that the process itself has already demonstrated the model in action. Trusted relationships, genuine buy-in, and cross-sector partnerships are forming.
Conclusions and Future Actions
Health sector partners are encouraged to engage with the Trust Lab as a practical mechanism for embedding community voice into AI research, adoption and governance, upholding human rights obligations, and meeting national guidance on meaningful interest holder engagement throughout the AI lifecycle.
Trust is foundational to every health interaction. The seemingly rapid introduction of AI into healthcare is creating new fractures in trust between communities and the health providers adopting it. In Australia, only 36% of people report willingness to trust AI in health, and only 30% believe current safeguards are sufficient.
The communities most affected by AI are rarely involved in identifying the need for it, designing it, or governing its use. If meaningful community and lived experience engagement is known to enhance trust and accountability, why isn’t this evident in AI in health? What signals of trustworthiness are communities looking for from their providers through this time of uncertainty?
Health Voices Victoria, with support from Deakin University, is establishing the AI in Health Trust Lab to begin answering these questions.
Methods and analysis
The Trust Lab has been shaped from the outset by an Activation Group of 13 people, meeting from May to November 2026. The group brings together community members, people with lived experience, academics, health workers, policy thinkers, and AI developers and designers. Together they are co-creating the Trust Lab's position, principles, priorities, and evaluation framework — modelling the very engagement approach the Lab seeks to promote.
Outcomes
The Trust Lab officially launches in August 2026. The co-creation process is still underway, and findings are emerging. What we can say is that the process itself has already demonstrated the model in action. Trusted relationships, genuine buy-in, and cross-sector partnerships are forming.
Conclusions and Future Actions
Health sector partners are encouraged to engage with the Trust Lab as a practical mechanism for embedding community voice into AI research, adoption and governance, upholding human rights obligations, and meeting national guidance on meaningful interest holder engagement throughout the AI lifecycle.
Biography
Katherine Burnard is an experienced facilitator and lived experience engagement specialist who leads co-design processes, capability building and curriculum development and community engagement at Health Voices Victoria. Her practice centers on social justice, equity and inclusion, and psychological safety. She brings diverse end complementary perspectives together to find better solutions for improving health and wellbeing. Katherine began her career as an Occupational Therapist in the mental health sector before completing postgraduate studies in Social Innovation in London, where she developed a strong commitment to co-design and lived experience leadership as drivers of health system improvement.
Mr Muhammad Firdaus Ab Raicob
Emergency Preparedness
Singhealth
Autonomous Multi-Agent AI for Outbreak Surveillance in a Singapore Public Health Setting
Abstract
SENTINEL: Operationalising Autonomous Multi-Agent AI for Real-Time Hospital Outbreak Surveillance in a Singapore Public Health Setting
Domain: Policy/Practice Sub-themes: Case Studies on AI Use in Public Health; Projects and Programs Harnessing AI for Public Health Benefits
________________________________________
Background and Aim
Hospital outbreak surveillance in Singapore's public health system relies on fragmented reporting workflows — disease occurrence reporting tools, infection prevention databases, and MOH notification chains operate as disconnected streams with no automated signal integration. This structural fragmentation introduces critical latency in early warning detection and imposes disproportionate manual burden on Infection Prevention and Control teams and executive leadership. The aim of this work was to design, deploy, and evaluate an autonomous AI-driven outbreak surveillance system within a live tertiary hospital setting, capable of real-time signal aggregation, classification, and executive escalation — while remaining fully compliant with Singapore's national regulatory frameworks.
Methods and Analysis
SENTINEL was developed and operationalised at Sengkang General Hospital (SKH), SingHealth cluster — a tertiary institution serving approximately one million inpatient bed-days annually. The system employs a five-tier multi-agent AI architecture integrating multiple institutional data sources, with outputs synthesised through AEGIS, an AI-generated executive briefing layer for C-suite decision-making. System architecture was designed governance-first: compliance with Singapore's Personal Data Protection Act (PDPA), MOH's Emergency Preparedness and Pandemic Financing (EPPF) framework, and ISO 22301 Business Continuity standards was embedded at the design stage. Performance was evaluated prospectively across respiratory, gastrointestinal, and environmental outbreak scenario types over a fourteen-month operational deployment period.
Outcomes
SENTINEL reduced executive outbreak briefing preparation time by over 60%, standardised signal classification taxonomy across clinical departments, and measurably improved cross-departmental coordination response latency. No reportable data governance breaches occurred throughout the operational period, validating the governance-first design approach.
Conclusion and Future Actions
Autonomous AI outbreak surveillance is operational — not aspirational — in a Singapore public hospital. Future work should focus on expanding SENTINEL's integration with national MOH reporting systems and adapting the governance architecture for deployment across ASEAN health systems sharing analogous regulatory structures, providing a replicable model for regional outbreak AI infrastructure.
Domain: Policy/Practice Sub-themes: Case Studies on AI Use in Public Health; Projects and Programs Harnessing AI for Public Health Benefits
________________________________________
Background and Aim
Hospital outbreak surveillance in Singapore's public health system relies on fragmented reporting workflows — disease occurrence reporting tools, infection prevention databases, and MOH notification chains operate as disconnected streams with no automated signal integration. This structural fragmentation introduces critical latency in early warning detection and imposes disproportionate manual burden on Infection Prevention and Control teams and executive leadership. The aim of this work was to design, deploy, and evaluate an autonomous AI-driven outbreak surveillance system within a live tertiary hospital setting, capable of real-time signal aggregation, classification, and executive escalation — while remaining fully compliant with Singapore's national regulatory frameworks.
Methods and Analysis
SENTINEL was developed and operationalised at Sengkang General Hospital (SKH), SingHealth cluster — a tertiary institution serving approximately one million inpatient bed-days annually. The system employs a five-tier multi-agent AI architecture integrating multiple institutional data sources, with outputs synthesised through AEGIS, an AI-generated executive briefing layer for C-suite decision-making. System architecture was designed governance-first: compliance with Singapore's Personal Data Protection Act (PDPA), MOH's Emergency Preparedness and Pandemic Financing (EPPF) framework, and ISO 22301 Business Continuity standards was embedded at the design stage. Performance was evaluated prospectively across respiratory, gastrointestinal, and environmental outbreak scenario types over a fourteen-month operational deployment period.
Outcomes
SENTINEL reduced executive outbreak briefing preparation time by over 60%, standardised signal classification taxonomy across clinical departments, and measurably improved cross-departmental coordination response latency. No reportable data governance breaches occurred throughout the operational period, validating the governance-first design approach.
Conclusion and Future Actions
Autonomous AI outbreak surveillance is operational — not aspirational — in a Singapore public hospital. Future work should focus on expanding SENTINEL's integration with national MOH reporting systems and adapting the governance architecture for deployment across ASEAN health systems sharing analogous regulatory structures, providing a replicable model for regional outbreak AI infrastructure.
Biography
Muhammad Firdaus Bin Ab. Raicob is Assistant Manager for Crisis Planning and Operations at Sengkang General Hospital (SKH), SingHealth cluster, Singapore, where he leads the design and operationalisation of AI-enabled systems for hospital emergency preparedness and outbreak surveillance.
A Registered Nurse by training and former Emergency Response Officer with the Singapore Police Force, Firdaus brings a rare combination of frontline clinical experience, operational crisis management, and applied public health informatics to his work. This intersection informs his approach to AI deployment: governance-first, equity-conscious, and grounded in the realities of regulated institutional environments.
He is currently an MSc candidate in Infectious Disease Emergencies at the NUS Saw Swee Hock School of Public Health, and is concurrently pursuing a Doctorate in Management — research that examines the organisational and leadership dimensions of technology adoption in public health systems.
Firdaus is the lead developer of several operationally deployed AI systems within SKH: SENTINEL, an autonomous multi-agent AI platform for real-time hospital outbreak surveillance; AEGIS, an AI-generated executive briefing system for C-suite outbreak intelligence; IRIS-2, a voice-to-text incident reporting tool for frontline clinical staff; and NEXUS-AI, an AI-powered pandemic simulation platform currently in development.
His work is anchored in Singapore's national regulatory frameworks — including the Personal Data Protection Act, MOH's Emergency Preparedness and Pandemic Financing framework, and ISO 22301 Business Continuity standards — and is designed with explicit transferability to ASEAN health systems navigating the governance challenges of AI adoption in outbreak response.
At AIPH 2026, Firdaus presents empirical findings from the deployment of these systems, contributing both operational evidence and a practical governance framework to the regional conversation on maximising AI's public health benefits while rigorously minimising its harms.
Prof Denis Bauer
Group Lead
Csiro
Trusted automated discovery in wastewater surveillance
Abstract
Background and Aim: Public health enters an era where AI agents participate in discovery, surveillance, and decisions. While agentic workflows advance, the infrastructure needed to safely reason over data remains immature. Governance, security, and data integrity standards upheld for human research are at risk of being weakened for AI-driven access. We develop processes to hold agentic access to the same or higher standards, particularly in pandemic preparedness with continuous surveillance and automated intelligence.
Methods and Analysis: The Australian CDC has prioritised establishment of the National Wastewater Surveillance Program (NWSP) with a 3-year pilot to strengthen Australia’s surveillance systems and enable early signal detection not possible with current systems. The PH4A-led Aus-SWEEP Collaboration is contracted to deliver NWSP services. Planned as a pilot with narrow scope (ie SARS-Cov-2), it informs the integration for broader wastewater and multimodal surveillance systems. CSIRO leads the future digital up-lift in intelligence capabilities, adhering to the Voluntary-AI-Safety-Standard (VAISS).
Outcomes: Aus-SWEEP builds on PathsBeacon, the implementation of the Global Alliance for Genomics in Health standard for the safe exchange of federated pathogen genomic data. Originally developed for data exchange between humans, the standard serves as a trusted gatekeeper for agentic access. PathsBeacon brokers smart-contract-based access between the sensitive genomic data and the AI agents, that makes data findable, context dependent, and grants access at the right granularity (e.g. raw data, or anonymized population summary). It hence preserves data governance without hindering innovation or slowing decision-making. The CSIRO-developed agentic system, Sciansa, represents the next evolution of this work, shifting from human-assisted to agent-assisted discovery while preserving the principles of trust, transparency, and governance that underpin a federated ecosystem.
Conclusion and Future actions: We will present Aus-SWEEP’s vision for a wastewater digital ecosystem that evolves from standardized data foundations to advanced multi-agent modelling for evidence-based decision-making and outbreak prevention.
Methods and Analysis: The Australian CDC has prioritised establishment of the National Wastewater Surveillance Program (NWSP) with a 3-year pilot to strengthen Australia’s surveillance systems and enable early signal detection not possible with current systems. The PH4A-led Aus-SWEEP Collaboration is contracted to deliver NWSP services. Planned as a pilot with narrow scope (ie SARS-Cov-2), it informs the integration for broader wastewater and multimodal surveillance systems. CSIRO leads the future digital up-lift in intelligence capabilities, adhering to the Voluntary-AI-Safety-Standard (VAISS).
Outcomes: Aus-SWEEP builds on PathsBeacon, the implementation of the Global Alliance for Genomics in Health standard for the safe exchange of federated pathogen genomic data. Originally developed for data exchange between humans, the standard serves as a trusted gatekeeper for agentic access. PathsBeacon brokers smart-contract-based access between the sensitive genomic data and the AI agents, that makes data findable, context dependent, and grants access at the right granularity (e.g. raw data, or anonymized population summary). It hence preserves data governance without hindering innovation or slowing decision-making. The CSIRO-developed agentic system, Sciansa, represents the next evolution of this work, shifting from human-assisted to agent-assisted discovery while preserving the principles of trust, transparency, and governance that underpin a federated ecosystem.
Conclusion and Future actions: We will present Aus-SWEEP’s vision for a wastewater digital ecosystem that evolves from standardized data foundations to advanced multi-agent modelling for evidence-based decision-making and outbreak prevention.
Biography
Dr Denis Bauer is a government research scientist, adjunct professor at Sydney and Macquarie University, and an AWS Hero. Her unique approach of joining Artificial Intelligence (AI) with deep biological domain knowledge translates research into impactful outcomes in disease gene detection in Motor Neuron Disease and the COVID-19 vaccine development. She chairs the Bioinformatics committee at Australia’s largest hospital district (Westmead Research Hub), is the senior author on a Nature Biotechnology publication, as well as keynotes international IT and Medical conferences. She was recognized in the Women in AI awards and is affiliated with the Australian Institute for Machine Learning. She has attracted more than $50M in funding to further life-science research and digital health.
Denis has a PhD in Bioinformatics from the University of Queensland and a Bachelor in Bioinformatics from Germany. She has a Graduate certificate from the Australian Institute of Company Directors and a Certificate in Executive Management and Development from the University of New South Wales Business School.
Sandra Carlson
HNE Population Health
Environmental Impact of AI use in Public Health – a Plea for Proportionality.
Abstract
Background. AI's electricity and water demands are real and rising; data-centre electricity use may roughly double by 2030. Yet some practitioners reject AI in public health entirely, and criticise bodies such as the WHO for adopting it, on two premises: that AI adds little value, and that its use is planet-threatening. We argue for both honest acknowledgement of the systemic concern and proportionality in how individual practitioners weigh it.
Methods. We separate two questions usually conflated: the marginal footprint of an individual's public health AI use, and the aggregate footprint of global AI growth. We benchmark the former against routine professional activities and estimate public health's negligible share of total demand, which is dominated by consumer and enterprise use.
Results. Training dominates model energy and is a shared, sunk cost; public health adds no model-specific training, so its marginal cost is essentially per query use. Current production estimates put a short query at ~0.24–0.3 Wh — less than 100 billionth of the models overall training energy use. Everyday savings easily offset typical use: switching off a webcam for one hour saves ~7 Wh (21 queries); turning off a 10-litre office urn from 6pm to 6am saves ~200 Wh — over 600 queries, or ten 3,000-word document reviews (8,000 tokens each). Its water use, between roughly 0.3 and 50 ml depending on what the estimate includes, is a fraction of the process water for one sheet of office paper (20 ml at the paper mill, but full footprint 2–5 litres).
Conclusion. Proportionality directs scarce environmental attention where it has leverage. Individual public health AI use warrants conscious choices, not prohibitions. The systemic footprint warrants advocacy for clean-powered, water-responsible infrastructure.
Methods. We separate two questions usually conflated: the marginal footprint of an individual's public health AI use, and the aggregate footprint of global AI growth. We benchmark the former against routine professional activities and estimate public health's negligible share of total demand, which is dominated by consumer and enterprise use.
Results. Training dominates model energy and is a shared, sunk cost; public health adds no model-specific training, so its marginal cost is essentially per query use. Current production estimates put a short query at ~0.24–0.3 Wh — less than 100 billionth of the models overall training energy use. Everyday savings easily offset typical use: switching off a webcam for one hour saves ~7 Wh (21 queries); turning off a 10-litre office urn from 6pm to 6am saves ~200 Wh — over 600 queries, or ten 3,000-word document reviews (8,000 tokens each). Its water use, between roughly 0.3 and 50 ml depending on what the estimate includes, is a fraction of the process water for one sheet of office paper (20 ml at the paper mill, but full footprint 2–5 litres).
Conclusion. Proportionality directs scarce environmental attention where it has leverage. Individual public health AI use warrants conscious choices, not prohibitions. The systemic footprint warrants advocacy for clean-powered, water-responsible infrastructure.
Biography
MS Jennifer Nixon
Policy and Research Manager
National Mental Health Consumer Alliance
AI and Digital Mental Health Tools in Australia: Risks, Regulation, Consumer Leadership
Abstract
Background and Aims
Artificial intelligence (AI), including large language models (LLMs) such as ChatGPT and Gemini, and AI-enabled Digital Mental Health Tools (DMHTs), are rapidly reshaping mental health care in Australia. While these technologies may improve access and convenience, they also introduce significant risks where safeguards and consumer leadership are weak or absent.
This abstract presents a rights-based position centred on people with lived experience of mental health challenges. It asserts that individuals are not passive recipients of digital transformation but active decision-makers, co-designers, and rights-holders. The aim is to outline a framework for AI and DMHT governance that upholds autonomy, dignity, and informed consent, while ensuring digital tools complement, rather than replace, relational and community-based care.
Method and Analysis
The paper draws on a synthesis of current literature, policy frameworks, and lived experience perspectives on AI and DMHT use in Australia. It examines applications across clinical, administrative, and consumer-led contexts, including decision-support systems, digital therapies, and general-purpose LLMs.
Analysis focuses on the rapid expansion of these tools within a fragmented and largely inadequate regulatory environment. It considers how oversight is distributed across therapeutic, privacy, and consumer frameworks, with key standards remaining voluntary. The analysis also explores drivers of uptake, including cost, workforce shortages, discrimination, and limited access to timely care, indicating that use often reflects systemic gaps rather than preference.
Importantly, consumers report using DMHTs to address unmet needs, including limited service availability, affordability barriers, and negative experiences within traditional systems. This reliance demonstrates that technological adoption is frequently driven by systemic failure, rather than clear evidence of safety or effectiveness.
At the same time, AI is increasingly embedded in service delivery through tools such as digital scribes, triage algorithms, and automated assessments, often without transparent consent processes. This raises concerns about accountability, clinical reliability, and the erosion of human judgement.
Without enforceable safeguards, these technologies risk normalising reduced choice, diminished privacy, and substitution of care.
Outcomes
The findings identify both opportunities and significant risks. AI and DMHTs may extend reach, provide immediate support, and enhance system efficiency. However, they cannot replace locally available, relationship-based care, particularly for people experiencing distress or crisis.
Key risks include inaccurate or harmful advice, lack of meaningful informed consent, opaque data practices, bias and discrimination, and over-reliance on outputs presented as authoritative. The embedding of AI in services, often without consumer knowledge, further undermines autonomy and privacy.
Regulatory gaps are substantial, with many tools falling outside formal oversight. General-purpose LLMs remain largely unregulated, despite widespread use as informal supports. Equity concerns are also evident, as these systems risk reinforcing existing inequalities, particularly for marginalised communities.
Conclusion and Future Actions
AI and DMHTs must be governed through a rights-based framework that prioritises autonomy, dignity, and informed consent. These technologies should complement rather than replace relational and community-based supports.
Future actions include establishing mandatory, enforceable national standards, strengthening regulatory accountability, and ensuring transparency in data use and decision-making. Governments should introduce legislation that guarantees meaningful informed consent and protects consumer rights.
Sustained investment in accessible, in-person and peer-led services is essential to prevent erosion of therapeutic relationships. A national digital literacy strategy, co-designed with people with lived experience, is required to support safe and informed use.
Finally, mechanisms such as a national harms register and the embedding of lived experience leadership across governance, design, and evaluation are critical to ensuring AI integration upholds equity, safety, and human rights. Embedding co-design and ongoing evaluation will be essential to ensure that innovation delivers benefit without harm. Aligning policy and practice with human rights principles will support a mental health system that is equitable, transparent, and accountable.
Artificial intelligence (AI), including large language models (LLMs) such as ChatGPT and Gemini, and AI-enabled Digital Mental Health Tools (DMHTs), are rapidly reshaping mental health care in Australia. While these technologies may improve access and convenience, they also introduce significant risks where safeguards and consumer leadership are weak or absent.
This abstract presents a rights-based position centred on people with lived experience of mental health challenges. It asserts that individuals are not passive recipients of digital transformation but active decision-makers, co-designers, and rights-holders. The aim is to outline a framework for AI and DMHT governance that upholds autonomy, dignity, and informed consent, while ensuring digital tools complement, rather than replace, relational and community-based care.
Method and Analysis
The paper draws on a synthesis of current literature, policy frameworks, and lived experience perspectives on AI and DMHT use in Australia. It examines applications across clinical, administrative, and consumer-led contexts, including decision-support systems, digital therapies, and general-purpose LLMs.
Analysis focuses on the rapid expansion of these tools within a fragmented and largely inadequate regulatory environment. It considers how oversight is distributed across therapeutic, privacy, and consumer frameworks, with key standards remaining voluntary. The analysis also explores drivers of uptake, including cost, workforce shortages, discrimination, and limited access to timely care, indicating that use often reflects systemic gaps rather than preference.
Importantly, consumers report using DMHTs to address unmet needs, including limited service availability, affordability barriers, and negative experiences within traditional systems. This reliance demonstrates that technological adoption is frequently driven by systemic failure, rather than clear evidence of safety or effectiveness.
At the same time, AI is increasingly embedded in service delivery through tools such as digital scribes, triage algorithms, and automated assessments, often without transparent consent processes. This raises concerns about accountability, clinical reliability, and the erosion of human judgement.
Without enforceable safeguards, these technologies risk normalising reduced choice, diminished privacy, and substitution of care.
Outcomes
The findings identify both opportunities and significant risks. AI and DMHTs may extend reach, provide immediate support, and enhance system efficiency. However, they cannot replace locally available, relationship-based care, particularly for people experiencing distress or crisis.
Key risks include inaccurate or harmful advice, lack of meaningful informed consent, opaque data practices, bias and discrimination, and over-reliance on outputs presented as authoritative. The embedding of AI in services, often without consumer knowledge, further undermines autonomy and privacy.
Regulatory gaps are substantial, with many tools falling outside formal oversight. General-purpose LLMs remain largely unregulated, despite widespread use as informal supports. Equity concerns are also evident, as these systems risk reinforcing existing inequalities, particularly for marginalised communities.
Conclusion and Future Actions
AI and DMHTs must be governed through a rights-based framework that prioritises autonomy, dignity, and informed consent. These technologies should complement rather than replace relational and community-based supports.
Future actions include establishing mandatory, enforceable national standards, strengthening regulatory accountability, and ensuring transparency in data use and decision-making. Governments should introduce legislation that guarantees meaningful informed consent and protects consumer rights.
Sustained investment in accessible, in-person and peer-led services is essential to prevent erosion of therapeutic relationships. A national digital literacy strategy, co-designed with people with lived experience, is required to support safe and informed use.
Finally, mechanisms such as a national harms register and the embedding of lived experience leadership across governance, design, and evaluation are critical to ensuring AI integration upholds equity, safety, and human rights. Embedding co-design and ongoing evaluation will be essential to ensure that innovation delivers benefit without harm. Aligning policy and practice with human rights principles will support a mental health system that is equitable, transparent, and accountable.
Biography
Priscilla Brice (she/they) is the Chief Executive Officer of the National Mental Health Consumer Alliance, Australia’s national peak body and collective voice for people with lived experience of mental health challenges.
Priscilla identifies as queer and neurodivergent and draws deeply on her own lived experience of mental health challenges in their leadership. Prior to joining the Alliance, Priscilla served as CEO of BEING Mental Health Consumers in NSW and was the Founder and Managing Director of All Together Now, a racial equity organisation based in Sydney, where they led innovative social change projects for over 12 years.
A strong advocate for equity and systemic reform, Priscilla is a Churchill Fellow, a Graduate of the Australian Institute of Company Directors (GAICD) and holds an MBA in Social Impact from the University of New South Wales.
Ms Elizabeth Jones
Clinical Nurse
Darling Downs Public Health Unit
Use of Artificial Intelligence in Public Health Unit practice: A scoping review.
Abstract
Background and Aim:
While governments and health departments are developing AI policies and workforce capabilities, little is known about how AI is being implemented to support routine public health unit (PHU) functions, including communicable disease control (CDC) and immunisation promotion. AI adoption should be guided by principles of transparency, equity, and evidence-based decision-making. However, existing literature largely focuses on AI’s potential rather than its real-world application, with limited evidence on implementation processes, effectiveness, efficiency, or unintended consequences. Regional perspectives and experiences of the public health nursing workforce are also underrepresented.
Aims and Methods:
This scoping review will examine how AI is being implemented in routine PHU functions related to CDC and immunisation promotion. Guided by the PRISMA-ScR framework, we will review scientific and grey literature published since 2022 (when ChatGPT was first released). This review will explore: (i) characteristics of research on AI implementation in CDC and immunisation promotion at a population level; (ii) the types and nature of AI being used; (iii) implementation processes and outcome measures; (iv) benefits and challenges from AI adoption; and (v) evidence relating to regional and/or nursing public health contexts. Definition of AI and related concepts will be informed by existing literature and refined as required.
Outcomes:
The review will provide a comprehensive overview of how AI is currently implemented in everyday PHU operations relating to CDC and immunisation promotion.
Conclusion and Future actions:
By synthesising evidence on real-world AI application, this review will highlight current practices, evidence gaps, and priorities for future research. Findings may inform evidence-based AI adoption by public health practitioners in their local jurisdictions, guided by transparent processes, measurable outcomes, and robust evidence.
While governments and health departments are developing AI policies and workforce capabilities, little is known about how AI is being implemented to support routine public health unit (PHU) functions, including communicable disease control (CDC) and immunisation promotion. AI adoption should be guided by principles of transparency, equity, and evidence-based decision-making. However, existing literature largely focuses on AI’s potential rather than its real-world application, with limited evidence on implementation processes, effectiveness, efficiency, or unintended consequences. Regional perspectives and experiences of the public health nursing workforce are also underrepresented.
Aims and Methods:
This scoping review will examine how AI is being implemented in routine PHU functions related to CDC and immunisation promotion. Guided by the PRISMA-ScR framework, we will review scientific and grey literature published since 2022 (when ChatGPT was first released). This review will explore: (i) characteristics of research on AI implementation in CDC and immunisation promotion at a population level; (ii) the types and nature of AI being used; (iii) implementation processes and outcome measures; (iv) benefits and challenges from AI adoption; and (v) evidence relating to regional and/or nursing public health contexts. Definition of AI and related concepts will be informed by existing literature and refined as required.
Outcomes:
The review will provide a comprehensive overview of how AI is currently implemented in everyday PHU operations relating to CDC and immunisation promotion.
Conclusion and Future actions:
By synthesising evidence on real-world AI application, this review will highlight current practices, evidence gaps, and priorities for future research. Findings may inform evidence-based AI adoption by public health practitioners in their local jurisdictions, guided by transparent processes, measurable outcomes, and robust evidence.
Biography
Elizabeth Jones is a Clinical Nurse at the Darling Downs Public Health Unit in Toowoomba, Queensland. She works across communicable disease control and immunisation and holds tertiary qualifications in science, nursing, and psychology. Her experience spans acute, emergency, community, and detention health care to ethnically diverse metropolitan, regional and remote populations with complex physical, psychological, and social health needs. This breadth of practice has fostered a strong interest in health equity, access to care, and the delivery of responsive health services across varied environments.
Dr Julie Ayre
Senior Research Fellow
Sydney Health Literacy Lab, University of Sydney
Co-designed resources to support critical reflection about using ChatGPT for health advice
Abstract
Intro: Generative artificial intelligence (AI) tools have clear potential health benefits for individuals (e.g. simplifying information) and risks (e.g. inaccurate information). This study evaluated two brief co-designed educational resources to help people think critically about asking ChatGPT for health information.
Methods: In this online randomised controlled trial, Australian adults who had used ChatGPT in the past 6 months were randomised to one of three groups: (1) animated video; (2) infographic images, or (3) control (unrelated information). We co-designed plain language resources with community members to ensure that key messages were meaningful and clear. The primary outcomes were intention to ask ChatGPT a question in ‘lower risk’ and ‘higher risk’ scenarios, where higher risk scenarios would typically require clinical interpretation (e.g. diagnosis, advice about care-seeking). Lower risk questions were more general e.g. understanding a health condition or medical term. Secondary outcomes were knowledge about ChatGPT, trust in ChatGPT, and intervention acceptability.
Results: Of the 619 participants, 592 were included in the analysis sample. Average age was 47.0 years (SD=16.4) and 42.6% identified as male. Participants in the animated video group (n=191) reported lower intention to use ChatGPT for higher risk scenarios compared to those in the infographic images group (n=203, p=0.010); both groups reported lower intentions compared to control (n=205, p<0.001). There was no effect of group on intentions to use ChatGPT for lower risk scenarios (p=0.800). Participants in the intervention groups had higher knowledge of ChatGPT (p<0.001) and reported lower trust (p<0.001), compared to control.
Conclusion: Brief educational resources may help improve knowledge of ChatGPT and reduce intentions to ask riskier health questions. This study represents an initial step towards addressing AI health literacy and highlights the kinds of health literacy skills that can support people to navigate AI tools safely.
Methods: In this online randomised controlled trial, Australian adults who had used ChatGPT in the past 6 months were randomised to one of three groups: (1) animated video; (2) infographic images, or (3) control (unrelated information). We co-designed plain language resources with community members to ensure that key messages were meaningful and clear. The primary outcomes were intention to ask ChatGPT a question in ‘lower risk’ and ‘higher risk’ scenarios, where higher risk scenarios would typically require clinical interpretation (e.g. diagnosis, advice about care-seeking). Lower risk questions were more general e.g. understanding a health condition or medical term. Secondary outcomes were knowledge about ChatGPT, trust in ChatGPT, and intervention acceptability.
Results: Of the 619 participants, 592 were included in the analysis sample. Average age was 47.0 years (SD=16.4) and 42.6% identified as male. Participants in the animated video group (n=191) reported lower intention to use ChatGPT for higher risk scenarios compared to those in the infographic images group (n=203, p=0.010); both groups reported lower intentions compared to control (n=205, p<0.001). There was no effect of group on intentions to use ChatGPT for lower risk scenarios (p=0.800). Participants in the intervention groups had higher knowledge of ChatGPT (p<0.001) and reported lower trust (p<0.001), compared to control.
Conclusion: Brief educational resources may help improve knowledge of ChatGPT and reduce intentions to ask riskier health questions. This study represents an initial step towards addressing AI health literacy and highlights the kinds of health literacy skills that can support people to navigate AI tools safely.
Biography
Dr Ayre is an NHMRC Emerging Leader Research Fellow (2023-2027) at the Sydney Health Literacy Lab, University of Sydney. Her research focuses on using digital technology to support health literacy initiatives in the healthcare workforce and the community. She has a strong track record with >80 publications in high quality national and international journals (e.g. npj Digital Medicine, MJA). Dr Ayre sits on the Scientific Advisory Team for the Australian Government’s National Health Literacy Framework.
Ms Negin Mirzaei Damabi
Hdr Student
Adelaide University
When Policy Fails, AI Follows: Migrant Women's Reproductive Health Rights
Abstract
Background
Migrant and refugee women experience significant sexual and reproductive health (SRH) inequities in Australia. As AI-driven digital health tools increasingly shape how SRH services are designed and delivered, the policy frameworks governing these services must reflect the needs of structurally marginalised populations. Yet whether Australian SRH policies adequately represent migrant and refugee women -- or account for the role of AI in service delivery -- remains unexamined.
What We Did
This study applied Bacchi's "What's the Problem Represented to Be?" (WPR) framework to critically analyse Australian national and state-level health policies relevant to the SRH of migrant and refugee women. Policy texts were examined for how problems are constructed, whose needs are centred, and whether AI or digital health tools are identified as mechanisms for service delivery or equity.
Results
Policies largely failed to represent the SRH needs of migrant and refugee women accurately or meaningfully. Migrant populations were homogenised, SRH was framed through a reproductive risk lens rather than a rights-based framework, and structural barriers were minimally acknowledged. Critically, AI and digital health were entirely absent from policy texts -- leaving no governance foundation for how such tools should serve, or avoid further marginalising, this population.
Lessons
The complete absence of AI from Australian SRH policies for migrant and refugee women represents a significant governance gap. Without explicit policy direction, AI-driven health tools risk replicating existing inequities at scale. Rights-based, intersectional policy reform is urgently needed to ensure AI systems are designed and governed in ways that actively serve -- rather than invisibilise -- structurally marginalised populations.
Migrant and refugee women experience significant sexual and reproductive health (SRH) inequities in Australia. As AI-driven digital health tools increasingly shape how SRH services are designed and delivered, the policy frameworks governing these services must reflect the needs of structurally marginalised populations. Yet whether Australian SRH policies adequately represent migrant and refugee women -- or account for the role of AI in service delivery -- remains unexamined.
What We Did
This study applied Bacchi's "What's the Problem Represented to Be?" (WPR) framework to critically analyse Australian national and state-level health policies relevant to the SRH of migrant and refugee women. Policy texts were examined for how problems are constructed, whose needs are centred, and whether AI or digital health tools are identified as mechanisms for service delivery or equity.
Results
Policies largely failed to represent the SRH needs of migrant and refugee women accurately or meaningfully. Migrant populations were homogenised, SRH was framed through a reproductive risk lens rather than a rights-based framework, and structural barriers were minimally acknowledged. Critically, AI and digital health were entirely absent from policy texts -- leaving no governance foundation for how such tools should serve, or avoid further marginalising, this population.
Lessons
The complete absence of AI from Australian SRH policies for migrant and refugee women represents a significant governance gap. Without explicit policy direction, AI-driven health tools risk replicating existing inequities at scale. Rights-based, intersectional policy reform is urgently needed to ensure AI systems are designed and governed in ways that actively serve -- rather than invisibilise -- structurally marginalised populations.
Biography
Negin is a dedicated Ph.D. candidate in Public health, specializing in public health research. With a strong background in clinical and translational research, she possesses a deep understanding of the intersection between medical science and public health. Negin's expertise extends to various aspects of healthcare, including policy development, health provision, human rights, and empowerment.
Driven by a profound commitment to advancing public health, Negin actively seeks opportunities to contribute to the betterment of communities and societies. She firmly believes in the importance of equality in healthcare access and strives to promote inclusivity and fairness in her work. Negin's research endeavors aim to shed light on pressing public health issues and generate evidence-based solutions to improve health outcomes and quality of life for individuals and populations.
Dr Silas Lui
Senior Evaluation and Planning Officer
Queensland Aboriginal and Islander Health Council
Using Artificial Intelligence for systematic thematic analysis: Copilot
Abstract
Background:
This paper investigates how large language model Artificial Intelligence (Microsoft Copilot) can support systematic thematic analysis of qualitative health data using a modified six-step approach. Using Indigenous prioritised health needs from the joint regional needs assessment conducted across Queensland in 2024 – 2025, this paper demonstrates how artificial intelligence (AI) tools can be used to effectively to analyse qualitative data.
Methods:
Microsoft Copilot was utilised to perform a systematic thematic analysis of the extracted data using a modified six-step approach. The first step involves familiarisation of Copilot with the data, research context and methodology. The second involved asking copilot to select key words from the data. Step 3, asked copilot to do coding systematically to label meaningful features of the data. In step 4, copilot was prompted to develop themes by grouping related codes and reviewing patterns across the dataset to identify major issues affecting First Nation peoples in Queensland. In step 5, Copilot was prompted to develop a conceptualisation through interpretation of the keywords, codes, themes and how they are related to another. In the final step 6, the interpreted concepts were presented in a structured framework demonstrating how Copilot’s ability to assist in translating qualitative health data to thematic outputs.
Results:
Copilot identified five themes comprising a consistent call for culturally safe, community-led, and equity-focused approaches that strengthen prevention, improve access to care, support the First Nations workforce, and address the wider drivers of health and wellbeing. The results indicate that Copilot can support organisations synthesis, and conceptualisation of large qualitative data.
Conclusion:
The study demonstrated that copilot supports thematic analysis process by assisting with coding, theme development and conceptual framing and it offers practical solution for health organisations to discover insights in the large qualitative datasets.
This paper investigates how large language model Artificial Intelligence (Microsoft Copilot) can support systematic thematic analysis of qualitative health data using a modified six-step approach. Using Indigenous prioritised health needs from the joint regional needs assessment conducted across Queensland in 2024 – 2025, this paper demonstrates how artificial intelligence (AI) tools can be used to effectively to analyse qualitative data.
Methods:
Microsoft Copilot was utilised to perform a systematic thematic analysis of the extracted data using a modified six-step approach. The first step involves familiarisation of Copilot with the data, research context and methodology. The second involved asking copilot to select key words from the data. Step 3, asked copilot to do coding systematically to label meaningful features of the data. In step 4, copilot was prompted to develop themes by grouping related codes and reviewing patterns across the dataset to identify major issues affecting First Nation peoples in Queensland. In step 5, Copilot was prompted to develop a conceptualisation through interpretation of the keywords, codes, themes and how they are related to another. In the final step 6, the interpreted concepts were presented in a structured framework demonstrating how Copilot’s ability to assist in translating qualitative health data to thematic outputs.
Results:
Copilot identified five themes comprising a consistent call for culturally safe, community-led, and equity-focused approaches that strengthen prevention, improve access to care, support the First Nations workforce, and address the wider drivers of health and wellbeing. The results indicate that Copilot can support organisations synthesis, and conceptualisation of large qualitative data.
Conclusion:
The study demonstrated that copilot supports thematic analysis process by assisting with coding, theme development and conceptual framing and it offers practical solution for health organisations to discover insights in the large qualitative datasets.
Biography
Position: Senior Evaluation and Planning Officer; Adjunct Research Fellow – Centre for Health Research – UniSQ.
Experience in conducting epidemiological studies; mixed method research techniques, with methodological expertise in qualitative and quantitative data collection and analysis; writing and reviewing scientific publications, project reports, and evaluation reports; health data systems; use of surveillance data and conceptualisation of health needs prioritisation; cross-cultural skills and ability to teach and mentor.
Mr Tom Burda
Na
AIHW
Accelerating Trusted Public Health Data Release through AI-Enabled Compliance Automation
Abstract
Background and Aim
As the use of integrated data via designated secure access environments with strict controls around data release becomes more frequent, compliance checking for data ingestion and release increases substantially. Data custodians must be confident that linkage activities comply with approved ethics, privacy and governance conditions. These checks are manual, repetitive and can be error-prone, creating bottlenecks that delay access to critical evidence and increase compliance risk. This project automates the extraction and reconciliation of approval requirements to make compliance checking faster, more reliable and scalable.
Methods and Analysis
Approval requirements are distributed across certificates - public interest certificates, ethics and governance documents - with inconsistent formats and terminology. In this pilot, a rules-based natural language processing (NLP) pipeline converts these unstructured conditions into structured, auditable representations of approved data items, extracting variables from PDFs via regular expressions, exact and fuzzy matching. Extracted variable lists were validated against human review of source documents across five datasets. Because the approach is rules-based and explainable, every compliance decision can be traced, reviewed and independently validated.
Outcomes
The capability embedded within the data release workflow, currently supports five datasets. Structured metadata is generated from approval artefacts and automatically cross-checked to identify issues early in the release process. Transparent, reproducible compliance assessments have strengthened custodian confidence and governance assurance. Compliance checking that previously took days per dataset can now be completed within hours.
Future Actions
AI-enabled compliance automation improves both the speed and integrity of public health data release. Faster access to trusted linked data supports research, pandemic preparedness and policy response, enabling earlier identification of emerging health issues and more timely action. The approach provides a scalable model for applying trusted AI to governance processes across national health data infrastructure while preserving privacy, robust governance controls and public trust.
As the use of integrated data via designated secure access environments with strict controls around data release becomes more frequent, compliance checking for data ingestion and release increases substantially. Data custodians must be confident that linkage activities comply with approved ethics, privacy and governance conditions. These checks are manual, repetitive and can be error-prone, creating bottlenecks that delay access to critical evidence and increase compliance risk. This project automates the extraction and reconciliation of approval requirements to make compliance checking faster, more reliable and scalable.
Methods and Analysis
Approval requirements are distributed across certificates - public interest certificates, ethics and governance documents - with inconsistent formats and terminology. In this pilot, a rules-based natural language processing (NLP) pipeline converts these unstructured conditions into structured, auditable representations of approved data items, extracting variables from PDFs via regular expressions, exact and fuzzy matching. Extracted variable lists were validated against human review of source documents across five datasets. Because the approach is rules-based and explainable, every compliance decision can be traced, reviewed and independently validated.
Outcomes
The capability embedded within the data release workflow, currently supports five datasets. Structured metadata is generated from approval artefacts and automatically cross-checked to identify issues early in the release process. Transparent, reproducible compliance assessments have strengthened custodian confidence and governance assurance. Compliance checking that previously took days per dataset can now be completed within hours.
Future Actions
AI-enabled compliance automation improves both the speed and integrity of public health data release. Faster access to trusted linked data supports research, pandemic preparedness and policy response, enabling earlier identification of emerging health issues and more timely action. The approach provides a scalable model for applying trusted AI to governance processes across national health data infrastructure while preserving privacy, robust governance controls and public trust.
Biography
Tom Burda is a Data Engineer at the Australian Institute of Health and Welfare (AIHW), where he develops automated solutions to streamline the secure ingestion, governance and release of integrated health data. He holds Graduate Certificates in Data Science and Biostatistics, and brings two years of experience in computer science to his work building reliable, reproducible data pipelines.
Tom is passionate about public health and the role that trusted data infrastructure plays in supporting research, policy and pandemic preparedness. His current work focuses on improving access to the National Health Data Hub, applying explainable, rules-based AI to compliance and governance processes that have traditionally been manual and time-intensive. He is particularly interested in approaches that strengthen governance assurance and public trust while reducing barriers to timely, responsible data use.
