1A - "Public Health Surveillance, Intelligence, Monitoring and Regulation"
Tracks
Track 1
Case studies on AI use in public health
Educating our workforce for AI
Enabling infrastructure for AI (sharing data, federated learning, etc.)
Projects and programs harnessing AI for public health benefits
| Tuesday, November 10, 2026 |
| 3:30 PM - 5:00 PM |
Speaker
Mr Jacob Madden
Chief Data and Analytics Officer
Australian Centre for Disease Control
Governance first, intelligence second: how CDC is approaching AI-ready public health surveillance
Abstract
Background and Aim
Timely, accurate public health surveillance saves lives — it enables faster outbreak detection, better resource allocation, and more effective policy response. Artificial intelligence offers real capability gains toward that goal — smarter data integration, faster signal detection and more timely decision support. But deploying AI effectively at national scale demands more than technical investment. It demands a governance architecture that earns jurisdictional trust, interoperability standards that make data meaningfully connectable, and design principles that keep human judgement sovereign in an AI-assisted system. As Australia's first dedicated national public health agency, CDC is confronting these challenges in real time, with a mandate to build surveillance infrastructure that is modern, scalable, and accountable from the outset.
Methods and Analysis
CDC is developing its digital and data capability through a value-led framework — one that asks what decision each information source will support, for whom, and under what safeguards — before designing the technical layer. This approach draws on human-centred design methodology, alignment with the Australian Government's Responsible AI framework, and active engagement with state and territory jurisdictions on interoperability standards and data sharing arrangements. The National Public Health Surveillance System (NPHSS), currently in proof-of-concept development, serves as the primary vehicle for testing and refining these principles in practice.
Outcomes
To date, the design process has established foundational governance principles, a value-based approach to data access and use, and a federated architecture oriented around existing jurisdictional realities rather than requiring their replacement. Critically, these foundations were determined before technical build commenced — reflecting the governance-first philosophy that underpins CDC's approach. Work underway is translating these principles into human-centred workflows that augment — not replace — the epidemiological and public health workforce, ensuring AI-assisted capabilities remain connected to operational decision-making rather than sitting outside it.
Conclusion and Future Actions
Building AI-ready public health surveillance infrastructure requires governance and human-centred design to precede — not follow — technical deployment. CDC's experience points to a replicable framework for national agencies navigating this challenge. Future actions include maturing the NPHSS proof-of-concept, enhancing jurisdictional data sharing arrangements, and contributing to national AI governance standards for public health application.
Timely, accurate public health surveillance saves lives — it enables faster outbreak detection, better resource allocation, and more effective policy response. Artificial intelligence offers real capability gains toward that goal — smarter data integration, faster signal detection and more timely decision support. But deploying AI effectively at national scale demands more than technical investment. It demands a governance architecture that earns jurisdictional trust, interoperability standards that make data meaningfully connectable, and design principles that keep human judgement sovereign in an AI-assisted system. As Australia's first dedicated national public health agency, CDC is confronting these challenges in real time, with a mandate to build surveillance infrastructure that is modern, scalable, and accountable from the outset.
Methods and Analysis
CDC is developing its digital and data capability through a value-led framework — one that asks what decision each information source will support, for whom, and under what safeguards — before designing the technical layer. This approach draws on human-centred design methodology, alignment with the Australian Government's Responsible AI framework, and active engagement with state and territory jurisdictions on interoperability standards and data sharing arrangements. The National Public Health Surveillance System (NPHSS), currently in proof-of-concept development, serves as the primary vehicle for testing and refining these principles in practice.
Outcomes
To date, the design process has established foundational governance principles, a value-based approach to data access and use, and a federated architecture oriented around existing jurisdictional realities rather than requiring their replacement. Critically, these foundations were determined before technical build commenced — reflecting the governance-first philosophy that underpins CDC's approach. Work underway is translating these principles into human-centred workflows that augment — not replace — the epidemiological and public health workforce, ensuring AI-assisted capabilities remain connected to operational decision-making rather than sitting outside it.
Conclusion and Future Actions
Building AI-ready public health surveillance infrastructure requires governance and human-centred design to precede — not follow — technical deployment. CDC's experience points to a replicable framework for national agencies navigating this challenge. Future actions include maturing the NPHSS proof-of-concept, enhancing jurisdictional data sharing arrangements, and contributing to national AI governance standards for public health application.
Biography
Jacob Madden is the Chief Data and Analytics Officer at the Australian Centre for Disease Control. Before joining the Australian CDC, Jacob led the Strategy Branch to deliver the policy development and legislation work to establish the Australian CDC. This included extensive stakeholder engagement and complex policy work with state and territory governments to design a data sharing regime to support public health activities.
Previous roles in the Department of Health, Disability and Ageing include leading the response to COVID-19 and other emergencies in aged care settings and supporting aged care prudential reforms. This included building and redesigning reporting of COVID-19 cases and other supports to aged care homes, harnessing data to support delivery of government supports.
Jacob holds a Master of Public Health from the Australian National University and a Bachelor of Laws (Hons)/Bachelor of Arts from the University of Notre Dame Australia.
Dr Charles Alpren
Deputy Director
Western Public Health Unit
AI Skills and Operational Efficiency from Development of an AI-Generated Surveillance Report
Abstract
Background and Aim
The use of large language model (LLM) artificial intelligence (AI) in public health can be limited by lack of availability of enterprise solutions in secure IT environments and the need to use protected data for public health practice. However, AI requires familiarity and practice for safe, effective use. To develop AI skills and find operational efficiencies, Western Public Health Unit (WPHU), developed an AI-generated surveillance report using publicly available interjurisdictional data from Australia and around the world to maintain situational awareness of outbreaks outside our catchment.
Methods and Analysis
Using commercial LLMs (ChatGPT and Claude), we iterated a prompt, examined failures, and designed workarounds for problems with AI-based internet searches. Later versions included quality assurance of each report with operator approval required following data collection prior to report analysis, collation, and distribution. Report generation is automated using python scripts calling the Anthropic API.
Outcomes
We now understand issues with AI internet searches including JavaScript dependencies and Robots.txt errors. Prompt design, quality control, and critical appraisal of AI outputs have all improved through development of this tool.
WPHU produce a weekly report within 10 minutes that previously took three hours of epidemiologist’s time. Major omissions are avoided through the two-step process with publication approval withheld if significant events are absent from search results. Minor errors, contradictions, and miscalculations have been detected in some reports with quality assurance ongoing. The report includes links to sources for ease of verification.
Conclusion and Future Actions
We have used publicly available data and a current operational necessity to develop AI skills and save significant time. With AI technology progressing quickly and limitations on its current use with protected data, development of tools that use public data allow safe development of AI skills in the public health workforce and can bring operational efficiencies.
The use of large language model (LLM) artificial intelligence (AI) in public health can be limited by lack of availability of enterprise solutions in secure IT environments and the need to use protected data for public health practice. However, AI requires familiarity and practice for safe, effective use. To develop AI skills and find operational efficiencies, Western Public Health Unit (WPHU), developed an AI-generated surveillance report using publicly available interjurisdictional data from Australia and around the world to maintain situational awareness of outbreaks outside our catchment.
Methods and Analysis
Using commercial LLMs (ChatGPT and Claude), we iterated a prompt, examined failures, and designed workarounds for problems with AI-based internet searches. Later versions included quality assurance of each report with operator approval required following data collection prior to report analysis, collation, and distribution. Report generation is automated using python scripts calling the Anthropic API.
Outcomes
We now understand issues with AI internet searches including JavaScript dependencies and Robots.txt errors. Prompt design, quality control, and critical appraisal of AI outputs have all improved through development of this tool.
WPHU produce a weekly report within 10 minutes that previously took three hours of epidemiologist’s time. Major omissions are avoided through the two-step process with publication approval withheld if significant events are absent from search results. Minor errors, contradictions, and miscalculations have been detected in some reports with quality assurance ongoing. The report includes links to sources for ease of verification.
Conclusion and Future Actions
We have used publicly available data and a current operational necessity to develop AI skills and save significant time. With AI technology progressing quickly and limitations on its current use with protected data, development of tools that use public data allow safe development of AI skills in the public health workforce and can bring operational efficiencies.
Biography
Deputy Director at Western Public Health Unit, Victoria.
Mr. Amos Okot
Laboratory Technician
Kampala Capital City Authority
Malaria surveillance using Using Mid-Infrared Spectroscopy and AI
Abstract
Background and Aim
Accurate Accurate estimation of mosquito age structure and species composition is central to malaria transmission modelling, insecticide resistance monitoring, and evaluation of vector control interventions. Conventional methods including ovarian dissection and PCR are labor-intensive, reagent-dependent, limiting scalability in endemic regions. We evaluated mid-infrared spectroscopy (MIRS) combined with supervised machine learning as a rapid, non-destructive alternative for entomological surveillance.
Methods and Analysis
We analysed a publicly available dataset containing 41,368 spectra from Anopheles gambiae, Anopheles coluzzii, and Anopheles funestus collected in Burkina Faso, Tanzania, and Scotland. After applying quality-control filters to remove low-intensity or contaminated spectra, selected wavenumbers were used to train and test models for predicting mosquito age classes and species. Several algorithms were evaluated, including Random Forest, logistic regression, and support vector machines. Model performance was compared with traditional methods to assess feasibility for operational surveillance.
Outcomes
Age prediction showed strong performance, with the Random Forest model achieving 94% accuracy across test datasets. Species prediction achieved moderate accuracy (43%), likely influenced by uneven class representation and the limited diversity of spectral features. Despite this, the approach demonstrates clear potential, particularly for high-throughput age grading, which is critical for assessing vector control interventions and identifying shifts in mosquito longevity associated with insecticide resistance.
Conclusion and Future action
These findings indicate that MIRS integrated with machine learning offers a scalable, reagent-free approach for vector phenotyping. With expanded geographic calibration and prospective field validation, this platform could strengthen malaria surveillance by enabling rapid, cost-efficient monitoring of vector population dynamics and earlier detection of transmission risk shifts within endemic health systems.
Accurate Accurate estimation of mosquito age structure and species composition is central to malaria transmission modelling, insecticide resistance monitoring, and evaluation of vector control interventions. Conventional methods including ovarian dissection and PCR are labor-intensive, reagent-dependent, limiting scalability in endemic regions. We evaluated mid-infrared spectroscopy (MIRS) combined with supervised machine learning as a rapid, non-destructive alternative for entomological surveillance.
Methods and Analysis
We analysed a publicly available dataset containing 41,368 spectra from Anopheles gambiae, Anopheles coluzzii, and Anopheles funestus collected in Burkina Faso, Tanzania, and Scotland. After applying quality-control filters to remove low-intensity or contaminated spectra, selected wavenumbers were used to train and test models for predicting mosquito age classes and species. Several algorithms were evaluated, including Random Forest, logistic regression, and support vector machines. Model performance was compared with traditional methods to assess feasibility for operational surveillance.
Outcomes
Age prediction showed strong performance, with the Random Forest model achieving 94% accuracy across test datasets. Species prediction achieved moderate accuracy (43%), likely influenced by uneven class representation and the limited diversity of spectral features. Despite this, the approach demonstrates clear potential, particularly for high-throughput age grading, which is critical for assessing vector control interventions and identifying shifts in mosquito longevity associated with insecticide resistance.
Conclusion and Future action
These findings indicate that MIRS integrated with machine learning offers a scalable, reagent-free approach for vector phenotyping. With expanded geographic calibration and prospective field validation, this platform could strengthen malaria surveillance by enabling rapid, cost-efficient monitoring of vector population dynamics and earlier detection of transmission risk shifts within endemic health systems.
Biography
mos Okot is a laboratory scientist, Acting Deputy Director of Public Health and Environment at the Kampala Capital City Authority uganda. He is developing AI-enabled mid-infrared spectroscopy tools for reagent-free malaria surveillance and antimicrobial testing. His work focuses on scalable diagnostics to strengthen epidemic preparedness and health system resilience in low-resource settings
Dr Gerardo Luis Dimaguila
Informatics Lead
Murdoch Children's Research Institute
When should public health respond? AI-powered vaccine misinformation surveillance around real-world events
Abstract
Background and Aim
In Australia and globally, vaccine misinformation increasingly undermines confidence and uptake. Real-world vaccine policy, safety and programme events can spark discussion in which misinformation may spread, yet evidence is limited on when public health organisations should respond, monitor, or avoid amplifying it. We used VaxPulse, an AI-enabled infodemic surveillance platform (HREC/85026/RCHM-2022), to quantify misinformation dynamics and inform proportionate responses.
Methods and Analysis
We analysed misinformation around three US Advisory Committee on Immunization Practices (ACIP)-linked events that raised concern about misinformation spillover: thimerosal announcement (26 June 2025), vote favouring separate MMR plus varicella vaccines (18 September 2025), and HPV single-dose revision (5 January 2026). Using VaxPulse’s compound AI system of fine-tuned machine learning, deep learning and tuned large language models, we identified event-focused misinformation data. We applied interrupted time series using negative binomial regression to model misinformation rates and compare temporary-change and level-change models.
Outcomes
Misinformation accompanied all three events, although peaks often occurred before the main official announcement, such as discussion preceding the ACIP announcement that HPV vaccines would be reviewed. This shows how AI-powered infodemic surveillance can identify rising misinformation trends that may require earlier education, communication or monitoring. Temporary-change models suggested post-event dips across MMR, thimerosal and HPV events (IRR = 0.49, 0.95 and 0.55, respectively). In contrast, level-change models suggested longer-term increases in estimated baseline misinformation for MMR and thimerosal (IRR = 4.94 and 1.59), while HPV showed an immediate decrease followed by later rebound on visual inspection.
Conclusion and Future Actions
Infodemic risk after vaccine-related events is not uniform. AI-powered surveillance and impact modelling can support proportionate public health action by identifying when responses are needed, when monitoring is sufficient, and when restraint may reduce unnecessary amplification. Future work should translate these insights into practical response thresholds for health services and primary care.
In Australia and globally, vaccine misinformation increasingly undermines confidence and uptake. Real-world vaccine policy, safety and programme events can spark discussion in which misinformation may spread, yet evidence is limited on when public health organisations should respond, monitor, or avoid amplifying it. We used VaxPulse, an AI-enabled infodemic surveillance platform (HREC/85026/RCHM-2022), to quantify misinformation dynamics and inform proportionate responses.
Methods and Analysis
We analysed misinformation around three US Advisory Committee on Immunization Practices (ACIP)-linked events that raised concern about misinformation spillover: thimerosal announcement (26 June 2025), vote favouring separate MMR plus varicella vaccines (18 September 2025), and HPV single-dose revision (5 January 2026). Using VaxPulse’s compound AI system of fine-tuned machine learning, deep learning and tuned large language models, we identified event-focused misinformation data. We applied interrupted time series using negative binomial regression to model misinformation rates and compare temporary-change and level-change models.
Outcomes
Misinformation accompanied all three events, although peaks often occurred before the main official announcement, such as discussion preceding the ACIP announcement that HPV vaccines would be reviewed. This shows how AI-powered infodemic surveillance can identify rising misinformation trends that may require earlier education, communication or monitoring. Temporary-change models suggested post-event dips across MMR, thimerosal and HPV events (IRR = 0.49, 0.95 and 0.55, respectively). In contrast, level-change models suggested longer-term increases in estimated baseline misinformation for MMR and thimerosal (IRR = 4.94 and 1.59), while HPV showed an immediate decrease followed by later rebound on visual inspection.
Conclusion and Future Actions
Infodemic risk after vaccine-related events is not uniform. AI-powered surveillance and impact modelling can support proportionate public health action by identifying when responses are needed, when monitoring is sufficient, and when restraint may reduce unnecessary amplification. Future work should translate these insights into practical response thresholds for health services and primary care.
Biography
Dr Dimaguila is Informatics Lead at the Murdoch Children's Research Institute, and Data Innovation Lead of the Centre for Health Analytics at the Royal Children's Hospital. He also conceived and lead VaxPulse, an infodemic and misinformation platform to help health care workers, immunisers, and vaccine advocates stay ahead of emerging trends of misinformation.
Dr Ana Paula Cardoso Richter
Research Fellow
Deakin University
Development of an AI-Enabled Monitoring System for Breastmilk Substitute Marketing in Australia
Abstract
Background and Aim
Digital marketing has transformed how breastmilk substitute (BMS) companies engage with caregivers through websites and social media. Monitoring compliance with national regulatory frameworks or international Codes (e.g., the WHO Code) is challenging because assessments require interpretation of large volumes of textual and visual content. This study aimed to develop and validate an artificial intelligence (AI) system capable of automatically identifying and classifying potential regulatory contradictions in digital BMS marketing.
Methods and Analysis
Marketing content was collected from Australian BMS manufacturer and retailer websites, Facebook, and Instagram. A multi-stage AI pipeline was developed combining optical character recognition (OCR), object recognition, multimodal large language models (LLMs), and rule-based analytics. The system extracted information from images, videos, captions, product descriptions, and packaging labels, including product names, age recommendations, promotional claims, and key visual elements. OCR was used to capture embedded text from visual media and product descriptions, while object detection techniques identified visual features and product attributes. This information was then used to classify product stages and assess compliance with regulatory frameworks. Rule-based analytics detected price promotions, while LLMs analysed textual and visual content, generating classification outputs with explanatory justifications. System performance was evaluated against manually coded reference data.
Outcomes
The system demonstrated strong agreement with manual coding across product classification and regulatory assessment tasks. By automating the extraction and interpretation of marketing content, the system reduced the manual effort required to assess compliance across large volumes of digital marketing materials.
Conclusion and Future Actions
The integration of OCR, object detection, multimodal AI, and rule-based classification enables large-scale monitoring and automated assessment of digital marketing content against regulatory requirements. By analysing greater volumes of content than would be feasible through manual review, the system provides a scalable approach for supporting compliance monitoring, regulatory oversight, and accountability in digital marketing environments.
Digital marketing has transformed how breastmilk substitute (BMS) companies engage with caregivers through websites and social media. Monitoring compliance with national regulatory frameworks or international Codes (e.g., the WHO Code) is challenging because assessments require interpretation of large volumes of textual and visual content. This study aimed to develop and validate an artificial intelligence (AI) system capable of automatically identifying and classifying potential regulatory contradictions in digital BMS marketing.
Methods and Analysis
Marketing content was collected from Australian BMS manufacturer and retailer websites, Facebook, and Instagram. A multi-stage AI pipeline was developed combining optical character recognition (OCR), object recognition, multimodal large language models (LLMs), and rule-based analytics. The system extracted information from images, videos, captions, product descriptions, and packaging labels, including product names, age recommendations, promotional claims, and key visual elements. OCR was used to capture embedded text from visual media and product descriptions, while object detection techniques identified visual features and product attributes. This information was then used to classify product stages and assess compliance with regulatory frameworks. Rule-based analytics detected price promotions, while LLMs analysed textual and visual content, generating classification outputs with explanatory justifications. System performance was evaluated against manually coded reference data.
Outcomes
The system demonstrated strong agreement with manual coding across product classification and regulatory assessment tasks. By automating the extraction and interpretation of marketing content, the system reduced the manual effort required to assess compliance across large volumes of digital marketing materials.
Conclusion and Future Actions
The integration of OCR, object detection, multimodal AI, and rule-based classification enables large-scale monitoring and automated assessment of digital marketing content against regulatory requirements. By analysing greater volumes of content than would be feasible through manual review, the system provides a scalable approach for supporting compliance monitoring, regulatory oversight, and accountability in digital marketing environments.
Biography
Ana Paula Richter is a Postdoctoral Research Fellow at the Global Centre for Preventive Health and Nutrition, Deakin University. Her research focuses on the digital marketing of breastmilk substitutes and its impact on caregivers. She uses artificial intelligence and automated methods to monitor digital marketing practices, assess compliance with regulatory frameworks, and support large-scale public health surveillance. Her work aims to strengthen the implementation and enforcement of policies that protect breastfeeding and improve infant and young child nutrition.
Mr Matthew Warner-smith
Manager, Business Intelligence & Information Systems
Cancer Institute Nsw
Implementing artificial intelligence for the sustainability of breast cancer screening
Abstract
Problem
Growing demand and radiology workforce shortages threaten the sustainability of breast cancer screening programs. Artificial intelligence (AI) for the reading of mammograms offers the potential to maintain clinical quality while reducing radiologist workload. Implementation of AI-supported reading in BreastScreen NSW considered absolute AI performance alongside program-specific factors including baseline reader performance, variation by imaging technology and population subgroup, acceptability, and regulatory requirements.
What we did
A mixed-methods implementation was undertaken within BreastScreen NSW. Components included retrospective non-inferiority analyses and modelling of human–AI paired reading; prospective validation on contemporary data; legal review; and qualitative assessment of acceptability among clients and clinicians. A staged rollout followed, supported by rigorous monitoring and governance to manage risk and guide optimisation.
Results
In a retrospective validation AI achieved non-inferior sensitivity but lower specificity than human readers. Human–AI paired reading showed non-inferior specificity but lower sensitivity.
On contemporary data, AI sensitivity was lower than human readers, while specificity was comparable to human readers and improved relative to earlier software versions.
AI performance was strongest in younger clients and those without prior screening, where it approximated the reading performance of human readers.
Following evaluation, AI was implemented for initial screens in clients aged 50–59. Early results indicate comparable sensitivity to human readers, with improved specificity and lower recall rates. AI-assisted reading did not increase arbitration rates or alter cancer detection but did reduce false positive recalls.
Lessons
AI-supported reading represents a viable means of sustaining breast cancer screening. The implementation demonstrated a pragmatic approach to introducing AI into public health programs in the face of imperfect data while maintaining safety and trust. Findings support targeted and staged integrations of AI into clinical practice to optimise performance and minimise harms. The implementation illustrates the importance of considering subgroup performance differences, regulatory context, and impacts on clinician behaviour when integrating AI into clinical practice.
Growing demand and radiology workforce shortages threaten the sustainability of breast cancer screening programs. Artificial intelligence (AI) for the reading of mammograms offers the potential to maintain clinical quality while reducing radiologist workload. Implementation of AI-supported reading in BreastScreen NSW considered absolute AI performance alongside program-specific factors including baseline reader performance, variation by imaging technology and population subgroup, acceptability, and regulatory requirements.
What we did
A mixed-methods implementation was undertaken within BreastScreen NSW. Components included retrospective non-inferiority analyses and modelling of human–AI paired reading; prospective validation on contemporary data; legal review; and qualitative assessment of acceptability among clients and clinicians. A staged rollout followed, supported by rigorous monitoring and governance to manage risk and guide optimisation.
Results
In a retrospective validation AI achieved non-inferior sensitivity but lower specificity than human readers. Human–AI paired reading showed non-inferior specificity but lower sensitivity.
On contemporary data, AI sensitivity was lower than human readers, while specificity was comparable to human readers and improved relative to earlier software versions.
AI performance was strongest in younger clients and those without prior screening, where it approximated the reading performance of human readers.
Following evaluation, AI was implemented for initial screens in clients aged 50–59. Early results indicate comparable sensitivity to human readers, with improved specificity and lower recall rates. AI-assisted reading did not increase arbitration rates or alter cancer detection but did reduce false positive recalls.
Lessons
AI-supported reading represents a viable means of sustaining breast cancer screening. The implementation demonstrated a pragmatic approach to introducing AI into public health programs in the face of imperfect data while maintaining safety and trust. Findings support targeted and staged integrations of AI into clinical practice to optimise performance and minimise harms. The implementation illustrates the importance of considering subgroup performance differences, regulatory context, and impacts on clinician behaviour when integrating AI into clinical practice.
Biography
Matthew Warner-Smith is the Manager, Business Intelligence and Information Systems in the Screening and Prevention Division of the Cancer Institute NSW. Matthew is an operational leader at the Cancer Institute NSW with over 25 years’ experience in public health in a diverse range of government and multilateral settings. Prior to joining the Cancer Institute Matthew worked in the United Nations system developing surveillance systems for illicit drug use in Africa and the Middle East, and monitoring and reporting on national responses to HIV and AIDS. Matthew completed his Masters of Public Health in 1997 at the University of Sydney, with a focus on international health and tobacco control.
Dr Sonia El-Zaemey
Epidemiologist
Epidemiology Directorate, Department of Health
AI and Advanced Analytics in Western Australian Population Health Surveillance System
Abstract
Problem
Population health surveys underpin monitoring of risk factors, behaviours and wellbeing, but growing expectations for timeliness, granularity and efficient reporting are placing pressure on surveillance systems. Artificial intelligence (AI) and advanced analytical methods offer opportunities to enhance surveillance, but there is limited applied evidence from operational public health settings and ongoing concerns about bias, transparency and trust.
What We Did
We explored the use of AI and advanced analytical methods at multiple points in the survey pipeline to improve estimates, streamline analysis and support reporting outputs. Within Western Australia’s Health and Wellbeing Surveillance System, we implemented Bayesian hierarchical spatio-temporal models to generate Local Government Area level prevalence estimates from survey data. These addressed limitations associated with unstable direct estimates and suppression due to high relative standard error. In parallel, we undertook early testing of generative AI to support selected reporting tasks and scoped potential applications across the survey lifecycle, including questionnaire development, coding assistance and automated outputs.
Results
Bayesian small area estimation produced more stable Local Government Area estimates with narrower credible intervals compared with direct survey estimates, reducing suppression and improving geographic detail for reporting, including long-term alcohol-related harm indicators. Early use of generative AI demonstrated time savings in drafting routine reporting outputs, with quality dependent on strong human review to ensure accuracy, context and consistency. Additional applications across the survey lifecycle were identified but remain untested.
Lessons
Early experience suggests that AI and advanced analytical methods can add value to population health surveillance through improved estimation, more efficient workflows and clearer reporting. Bayesian modelling delivered immediate gains in estimate quality and local-level reporting, while generative AI showed promise for operational efficiency. Governance, transparency, and human oversight remain essential to maximise benefits while minimising risks associated with bias, error and loss of trust
Population health surveys underpin monitoring of risk factors, behaviours and wellbeing, but growing expectations for timeliness, granularity and efficient reporting are placing pressure on surveillance systems. Artificial intelligence (AI) and advanced analytical methods offer opportunities to enhance surveillance, but there is limited applied evidence from operational public health settings and ongoing concerns about bias, transparency and trust.
What We Did
We explored the use of AI and advanced analytical methods at multiple points in the survey pipeline to improve estimates, streamline analysis and support reporting outputs. Within Western Australia’s Health and Wellbeing Surveillance System, we implemented Bayesian hierarchical spatio-temporal models to generate Local Government Area level prevalence estimates from survey data. These addressed limitations associated with unstable direct estimates and suppression due to high relative standard error. In parallel, we undertook early testing of generative AI to support selected reporting tasks and scoped potential applications across the survey lifecycle, including questionnaire development, coding assistance and automated outputs.
Results
Bayesian small area estimation produced more stable Local Government Area estimates with narrower credible intervals compared with direct survey estimates, reducing suppression and improving geographic detail for reporting, including long-term alcohol-related harm indicators. Early use of generative AI demonstrated time savings in drafting routine reporting outputs, with quality dependent on strong human review to ensure accuracy, context and consistency. Additional applications across the survey lifecycle were identified but remain untested.
Lessons
Early experience suggests that AI and advanced analytical methods can add value to population health surveillance through improved estimation, more efficient workflows and clearer reporting. Bayesian modelling delivered immediate gains in estimate quality and local-level reporting, while generative AI showed promise for operational efficiency. Governance, transparency, and human oversight remain essential to maximise benefits while minimising risks associated with bias, error and loss of trust
Biography
Dr. Sonia El-Zaemey is an Epidemiologist with the Epidemiology Directorate at the Western Australian Department of Health. She leads the Health and Wellbeing Surveillance System and supports the analysis, interpretation and reporting of population health survey data.
Her work focuses on population health surveillance, survey methods, and the use of advanced analytics where appropriate. She is also exploring how artificial intelligence can support efficient reporting, evidence translation, transparency and human oversight.
Mr Damian Honeyman
Phd Candidate
Kirby Institute
Multilingual Event‑Based-Surveillance of Methanol and Toxic Alcohol Poisonings and Detections Worldwide, 2024–2025
Abstract
Background and Aim: Methanol poisoning and toxic alcohol contamination represent significant global public health threats, yet no systematic international surveillance infrastructure exists. Event-based surveillance using open-source intelligence (OSINT) offers a scalable, near real-time complement to traditional indicator-based systems. This study characterised the global epidemiology of methanol and toxic alcohol poisonings and detection events using an artificial intelligence (AI)-assisted multilingual OSINT framework.
Methods and Analysis: A lexicon of 17 English search terms, translated into up to 95 languages, queried the Bing News Application Programming Interface from August 2024 to January 2025. Retrieved articles were scraped, translated, and summarised using GPT-4o. Epidemiological variables were manually extracted and verified, then analysed in STATA/BE 18.0 and mapped using ArcGIS Pro v.3.1. Three multilingual search strategies of increasing linguistic breadth were compared as sensitivity analyses.
Outcomes: Of 9,396 retrieved articles, 777 were relevant, yielding 157 events across 24 countries: nine confirmed methanol poisonings, four suspected methanol poisonings, four toxic alcohol poisonings, and 140 toxic alcohol detections. Events were concentrated in Türkiye (43.3%), India (28.0%), and the Russian Federation (8.3%). Across all poisonings, 745 individuals were affected, including 137 fatalities. Expanding from a two-language baseline to 95 languages increased total event detection from 24 to 157 events and geographic coverage from 11 to 24 countries.
Conclusion and Future actions: AI-assisted multilingual OSINT surveillance can feasibly detect methanol poisoning events and upstream signals of toxic alcohol contamination at a global scale. Multilingual search substantially improves detection yield and geographic coverage, particularly for precursor signals in non-English media. Integration of such frameworks into existing public health early warning systems could support more timely responses to toxic alcohol threats.
Methods and Analysis: A lexicon of 17 English search terms, translated into up to 95 languages, queried the Bing News Application Programming Interface from August 2024 to January 2025. Retrieved articles were scraped, translated, and summarised using GPT-4o. Epidemiological variables were manually extracted and verified, then analysed in STATA/BE 18.0 and mapped using ArcGIS Pro v.3.1. Three multilingual search strategies of increasing linguistic breadth were compared as sensitivity analyses.
Outcomes: Of 9,396 retrieved articles, 777 were relevant, yielding 157 events across 24 countries: nine confirmed methanol poisonings, four suspected methanol poisonings, four toxic alcohol poisonings, and 140 toxic alcohol detections. Events were concentrated in Türkiye (43.3%), India (28.0%), and the Russian Federation (8.3%). Across all poisonings, 745 individuals were affected, including 137 fatalities. Expanding from a two-language baseline to 95 languages increased total event detection from 24 to 157 events and geographic coverage from 11 to 24 countries.
Conclusion and Future actions: AI-assisted multilingual OSINT surveillance can feasibly detect methanol poisoning events and upstream signals of toxic alcohol contamination at a global scale. Multilingual search substantially improves detection yield and geographic coverage, particularly for precursor signals in non-English media. Integration of such frameworks into existing public health early warning systems could support more timely responses to toxic alcohol threats.
Biography
Damian Honeyman is a Registered Nurse, academic, and PhD candidate specialising in AI-driven early warning systems for public health threats. With a background in intensive care, immunisation, and communicable disease control, his work integrates clinical insight with advanced analytics to detect chemical, biological, radiological, nuclear and emerging health risks using open-source intelligence and large language models. Damian has led and contributed to multiple research projects examining global outbreak patterns, including methanol poisoning and botulism, and has published extensively in biosecurity and epidemiology. He currently lectures in nursing and contributes to advancing AI integration in healthcare education and practice. His research focuses on strengthening surveillance systems, improving health system preparedness, and supporting timely, evidence-based responses to complex public health emergencies.
Ms Lyn Buchanan
Research Coordinator
University Of Melbourne
Acceptability of conversational voice AI in bowel cancer screening clinical follow-up
Abstract
Background and Aim
The Australian National Bowel Cancer Screening Program (NBCSP) aims to reduce bowel cancer. Its effectiveness depends on timely diagnostic follow-up of positive screening results. In 2024, 14.6% of individuals with a positive result did not complete follow-up despite the state-funded ‘Participant Follow-Up Function’ (PFUF) phone outreach service. Conversational voice AI (CVAI) has potential to augment PFUF with demonstrated effectiveness at increasing screening uptake internationally. Initial research with two PFUF teams identified opportunities and limitations to incorporating CVAI into the PFUF process. Understanding consumers’ perspectives on including CVAI in PFUF is the next critical step.
The aims of this research are to explore acceptability and requirements for a CVAI agent to increase clinical follow-up in people who test positive to the NBCSP.
Methods and Analysis
Semi-structured interviews were conducted with participants contacted by PFUF from two Australian states (n=30).
Preliminary inductive analysis identified common themes focussing on barriers and facilitators to successful clinical follow-up including the response to, and understanding of, the current follow-up process from receiving the positive result letter to the involvement of PFUF. The acceptability of the use of CVAI was influenced by trust, perceived usefulness, prior experiences, and attitudes toward AI. Participants considered CVAI useful for explaining results, next steps and providing information, while preferring complex queries to remain with PFUF staff. Ethical concerns and preferences regarding information sharing with CVAI were also identified.
Outcomes
This exploratory research will inform the development of a CVAI prototype for future testing.
Conclusions and future actions
CVAI has potential to complement PFUF by answering straightforward queries and providing additional information during the NBCSP follow-up process.
Future research will test the prototype to evaluate the CVAI on:
1. intention to follow-up;
2. effectiveness as a source of information for participants; and
3. usability during the follow-up process.
The Australian National Bowel Cancer Screening Program (NBCSP) aims to reduce bowel cancer. Its effectiveness depends on timely diagnostic follow-up of positive screening results. In 2024, 14.6% of individuals with a positive result did not complete follow-up despite the state-funded ‘Participant Follow-Up Function’ (PFUF) phone outreach service. Conversational voice AI (CVAI) has potential to augment PFUF with demonstrated effectiveness at increasing screening uptake internationally. Initial research with two PFUF teams identified opportunities and limitations to incorporating CVAI into the PFUF process. Understanding consumers’ perspectives on including CVAI in PFUF is the next critical step.
The aims of this research are to explore acceptability and requirements for a CVAI agent to increase clinical follow-up in people who test positive to the NBCSP.
Methods and Analysis
Semi-structured interviews were conducted with participants contacted by PFUF from two Australian states (n=30).
Preliminary inductive analysis identified common themes focussing on barriers and facilitators to successful clinical follow-up including the response to, and understanding of, the current follow-up process from receiving the positive result letter to the involvement of PFUF. The acceptability of the use of CVAI was influenced by trust, perceived usefulness, prior experiences, and attitudes toward AI. Participants considered CVAI useful for explaining results, next steps and providing information, while preferring complex queries to remain with PFUF staff. Ethical concerns and preferences regarding information sharing with CVAI were also identified.
Outcomes
This exploratory research will inform the development of a CVAI prototype for future testing.
Conclusions and future actions
CVAI has potential to complement PFUF by answering straightforward queries and providing additional information during the NBCSP follow-up process.
Future research will test the prototype to evaluate the CVAI on:
1. intention to follow-up;
2. effectiveness as a source of information for participants; and
3. usability during the follow-up process.
Biography
Sandra is currently undertaking a PhD in the School of Population and Global Health at the University of Melbourne. Her research focuses on improving participation in bowel cancer screening, with a particular emphasis on the use of conversational voice AI to support screening pathways and follow-up care.
With a background in pharmaceutical science and clinical trials, she transitioned into public health to pursue research aimed at improving cancer prevention, early detection, and patient engagement in screening programs.
Dr Lisa Sharwood
Injury Epidemiologist & Data Scientist
UNSW School of Population Health
Building National Infrastructure for AI-Enabled Injury Surveillance: The NISAR-ED Program
Abstract
Problem
Emergency departments (ED)s are Australia's frontline for injury, yet ED datasets cannot routinely identify the external causes such as self-harm, assault, family violence, product-related harm, and workplace hazards. Existing surveillance relies heavily on diagnosis coding, missing the contextual detail held within clinical free text. AI could extract these insights, but public health systems often lack the required resources to build a safe, scalable, accurate platform.
What you did
The National Injury Surveillance for Actionable Research-Emergency Department (NISAR-ED) program is creating Australia's first AI-enabled national injury surveillance infrastructure. MRFF funded and UNSW led, with partners including the AIHW, ACCC, ACEM, MUARC, state and territory health departments, clinicians and technical partners, we built a containerised, cloud-agnostic platform that uses AI and NLP through LLMs to classify injuries and derive ICD-11 external-cause information from routinely collected ED records. The program incorporates strict secure protocols within a jurisdiction’s cloud environment, a common data model, an automated ingestion and AI processing pipeline with embedded evaluation and observability, human-in-the-loop validation processes, implementation governance, and reporting tools designed to support public health surveillance while maintaining data sovereignty.
Results
By 2026, NISAR-ED had moved from concept to implementation: platform architecture completed, pilot deployment underway, and multiple jurisdictions engaged. The infrastructure enables standardised capture of prevention-relevant information-intent, mechanism, activity, location and product involvement-creating a foundation for nationally consistent injury surveillance and AI-enabled public health intelligence.
Lessons
NISAR-ED demonstrates how LLMs, paired with a grounded taxonomy and coding rules, can unlock the value of unstructured data already captured during care. The greatest challenge was not technical development but establishing the partnerships, IT collaboration, governance, standards, interoperability, validation and implementation frameworks needed for sustainable deployment. Building trusted infrastructure alongside technical capability is essential for scaling AI in public health and offers a transferable model for other national surveillance systems.
Emergency departments (ED)s are Australia's frontline for injury, yet ED datasets cannot routinely identify the external causes such as self-harm, assault, family violence, product-related harm, and workplace hazards. Existing surveillance relies heavily on diagnosis coding, missing the contextual detail held within clinical free text. AI could extract these insights, but public health systems often lack the required resources to build a safe, scalable, accurate platform.
What you did
The National Injury Surveillance for Actionable Research-Emergency Department (NISAR-ED) program is creating Australia's first AI-enabled national injury surveillance infrastructure. MRFF funded and UNSW led, with partners including the AIHW, ACCC, ACEM, MUARC, state and territory health departments, clinicians and technical partners, we built a containerised, cloud-agnostic platform that uses AI and NLP through LLMs to classify injuries and derive ICD-11 external-cause information from routinely collected ED records. The program incorporates strict secure protocols within a jurisdiction’s cloud environment, a common data model, an automated ingestion and AI processing pipeline with embedded evaluation and observability, human-in-the-loop validation processes, implementation governance, and reporting tools designed to support public health surveillance while maintaining data sovereignty.
Results
By 2026, NISAR-ED had moved from concept to implementation: platform architecture completed, pilot deployment underway, and multiple jurisdictions engaged. The infrastructure enables standardised capture of prevention-relevant information-intent, mechanism, activity, location and product involvement-creating a foundation for nationally consistent injury surveillance and AI-enabled public health intelligence.
Lessons
NISAR-ED demonstrates how LLMs, paired with a grounded taxonomy and coding rules, can unlock the value of unstructured data already captured during care. The greatest challenge was not technical development but establishing the partnerships, IT collaboration, governance, standards, interoperability, validation and implementation frameworks needed for sustainable deployment. Building trusted infrastructure alongside technical capability is essential for scaling AI in public health and offers a transferable model for other national surveillance systems.
Biography
Dr Sharwood is a Senior Research Fellow in the School of Population Health, Faculty of Medicine and Health, UNSW. I am also an Adjunct Professional Fellow in the School of Mechanical Engineering, Faculty of Engineering and IT, University of Technology Sydney.
I am an experienced epidemiologist specialising in injury, public health, health services research and data science, with over 20 years of experience in academic, government and industry settings. I apply my skills in risk assessment and mitigation, utilising comprehensive data and system analyses to provide evidence-based evaluations for improving health and safety outcomes. I am dedicated to advocating for health system optimisation through data-driven strategies, and to translating evidence-based findings to improve practice and policy for cost-effective health care.
Mr Daniel Fry
NA
AIHW
Accelerating Public Health Evidence through AI Agents and Automated Data Pipelines
Abstract
Background and Aim
Timely public health data is critical to effective public health action, enabling governments, researchers and health services to detect emerging issues, monitor population health trends and respond to changing demand. The COVID-19 pandemic highlighted the need for current or near-real-time data for disease surveillance, pandemic preparedness and emergency response.
Existing data pipelines rely on legacy software, manual hand-offs and fragmented codebases that introduce delays from approvals and linkage through to data assembly, review, testing and release. This project envisions an automated, agent-assisted pipeline ready to execute as soon as approvals are in place, enabling data releases in hours rather than weeks or months.
Methods and Analysis
This project uses AI agents to accelerate progress towards this end point, while keeping humans in control of design, review and release decisions. Through experimentation with agent development tools, agents have been found suitable for translating legacy code into flexible, automatable pipelines, summarising governance documents and converting data specifications into machine-readable pipeline inputs. This work has occurred without ever exposing agents to underlying unit-record data, preserving privacy while maintaining oversight and governance controls.
Outcomes
Although early-stage, agents have so far supported the standardisation and translation of legacy code into an open source, faster and more automatable data assembly pipeline. These early wins have reduced manual touchpoints, shortened dataset production timeframes, and decreased dependency on costly proprietary platforms by moving core assembly processes into open, reusable and highly automatable code.
Conclusion and Future actions
The sensitivity of public health data limits direct AI integration into production pipelines, but there is great opportunity to optimise pipeline-adjacent processes. Work is progressing towards a model where AI-supported activities accelerate release of critical health data while preserving privacy, maintaining human oversight and adhering to governance requirements.
Timely public health data is critical to effective public health action, enabling governments, researchers and health services to detect emerging issues, monitor population health trends and respond to changing demand. The COVID-19 pandemic highlighted the need for current or near-real-time data for disease surveillance, pandemic preparedness and emergency response.
Existing data pipelines rely on legacy software, manual hand-offs and fragmented codebases that introduce delays from approvals and linkage through to data assembly, review, testing and release. This project envisions an automated, agent-assisted pipeline ready to execute as soon as approvals are in place, enabling data releases in hours rather than weeks or months.
Methods and Analysis
This project uses AI agents to accelerate progress towards this end point, while keeping humans in control of design, review and release decisions. Through experimentation with agent development tools, agents have been found suitable for translating legacy code into flexible, automatable pipelines, summarising governance documents and converting data specifications into machine-readable pipeline inputs. This work has occurred without ever exposing agents to underlying unit-record data, preserving privacy while maintaining oversight and governance controls.
Outcomes
Although early-stage, agents have so far supported the standardisation and translation of legacy code into an open source, faster and more automatable data assembly pipeline. These early wins have reduced manual touchpoints, shortened dataset production timeframes, and decreased dependency on costly proprietary platforms by moving core assembly processes into open, reusable and highly automatable code.
Conclusion and Future actions
The sensitivity of public health data limits direct AI integration into production pipelines, but there is great opportunity to optimise pipeline-adjacent processes. Work is progressing towards a model where AI-supported activities accelerate release of critical health data while preserving privacy, maintaining human oversight and adhering to governance requirements.
Biography
Daniel Fry is a data integration specialist in the Data Governance and Integration Group at the Australian Institute of Health and Welfare. His work focuses on data integration, transformation and validation for national health data assets, with experience across SAS, Python, Polars and PySpark. He is currently involved in work to modernise legacy data processes, improve reproducibility and explore the practical use of AI-assisted tools in data pipeline development, automation and documentation.
Daniel holds a Graduate Certificate in Data Analytics from the University of Sydney, a Master of Teaching (Secondary) from UNSW and a Bachelor of Science with Honours from the University of Adelaide.
