Background and Aim
Artificial intelligence can support public health decision-making by improving prediction of long-term clinical outcomes. However, many high-value health datasets are distributed across institutions and cannot be centrally pooled because of governance, privacy, and confidentiality constraints. This work aimed to develop and evaluate a federated survival modelling framework that enables collaborative health AI while minimising risks associated with sharing patient-level data.

Methods and Analysis
We developed a federated proportional hazard metaparametric neural network (PH-MNN) for privacy-preserving survival analysis. The method combines neural network feature learning with time-basis functions to model nonlinear and time-dependent hazard relationships while retaining a Cox-type survival structure. A federated Breslow-type baseline hazard estimator was introduced to aggregate only client-level event counts and risk-set summaries. The framework was evaluated using synthetic survival data and National Joint Registry primary hip replacement data. Outcomes included all-cause mortality and revision surgery. PH-MNN was compared with federated Cox, DeepSurv, and DeepHit baseline models using IPCW concordance, IPCW Brier score, and marginal cumulative hazard ratio diagnostics.

Outcomes
In the clinical registry analysis, PH-MNN achieved the highest concordance for mortality prediction at 2, 4, and 8 years: 0.7575, 0.7488, and 0.7647, respectively. For revision prediction, PH-MNN achieved the highest concordance at 4 and 8 years: 0.5988 and 0.6001. PH-MNN also achieved the lowest Brier score for mortality at 2 and 4 years: 0.0239 and 0.0594, and the lowest Brier score for revision at 8 years: 0.0229. Hazard-ratio diagnostics showed that PH-MNN recovered clinically interpretable age- and BMI-dependent risk patterns more accurately than comparator models.

Conclusion and Future actions
Federated survival modelling provides practical infrastructure for responsible public health AI by enabling collaborative model development without centralising sensitive health data. Future work should evaluate this approach using real multi-institutional client settings and implementation pathways.

This Australian Research Council-funded project explores how artificial intelligence can be used to support culturally responsive workforce development while maintaining Indigenous governance, cultural integrity, and data sovereignty. Led by an Aboriginal researcher in partnership with Elders and community members, the project is developing immersive simulation and e-learning resources to prepare social work students for practice with rural and remote First Peoples communities.

While artificial intelligence offers significant opportunities for scalable educational resource development, it also presents substantial risks for Indigenous communities, including cultural misrepresentation, stereotyping, unauthorised use of cultural knowledge, and the reproduction of colonial narratives. To address these concerns, the project embedded Aboriginal Participatory Action Research processes and Indigenous governance structures throughout the design and development lifecycle.

Community members guided decisions regarding cultural representation, acceptable imagery, narrative development, and the use of AI-generated content. The project applied CARE Principles for Indigenous Data Governance alongside Indigenous Cultural and Intellectual Property (ICIP) protections to ensure community authority over cultural knowledge and representation. Iterative feedback processes enabled the identification and correction of culturally unsafe outputs, highlighting the limitations of current generative AI systems when operating without Indigenous oversight.

Preliminary findings suggest that First Peoples governance must be embedded at every stage of AI development rather than applied retrospectively as an ethical review process. Key lessons include the importance of community authority, relational accountability, transparent decision-making, and ongoing human oversight to mitigate bias and cultural harm.

This presentation provides a practical case study of Indigenous-led AI governance and offers a transferable framework for public health, education, and community sectors seeking to harness AI while protecting First Peoples rights, knowledges, and sovereignty.

Aboriginal Governance Statement

The project is Aboriginal-led and governed through ongoing partnerships with Aboriginal Elders, community members, researchers, and service providers. Decision-making authority regarding cultural content, imagery, narratives, and representations remains with Aboriginal participants and communities. The project applies CARE Principles for Indigenous Data Governance and Indigenous Cultural and Intellectual Property principles to ensure collective benefit, authority to control, responsibility, and ethics are maintained throughout project development. AI-generated content is reviewed through iterative community feedback processes before inclusion, ensuring cultural integrity, community accountability, and protection against inappropriate use or representation of Indigenous knowledges.

Background and Aim

Artificial intelligence (AI) is increasingly being adopted across healthcare, public health, disability, and government services.

While AI offers significant opportunities to improve service delivery, decision-making, and operational efficiency, failures in healthcare settings have demonstrated the potential for serious harm, including missed diagnoses, inappropriate treatments, biased clinical outcomes, and patient injury.

Examples of AI-related failures include surgical instrument navigation errors, over-reliance on AI-assisted endoscopy systems, unsafe cancer treatment recommendations, inaccurate sepsis prediction models, and diagnostic bias in melanoma detection for people with darker skin tones.

Despite growing adoption, many organisations lack a structured framework to govern AI use. This presentation introduces the Australian Government's AI Governance Standard (DTA AI Technical Standard), released in July 2025 and adopted by the Department of Health, Disability and Ageing. The Department has encouraged healthcare providers to begin implementing the Standard in preparation for emerging regulatory expectations and increased scrutiny of AI-enabled decision-making.

Methods and Analysis

The DTA AI Technical Standard provides a practical framework for governing AI systems throughout their lifecycle. Drawing on AI governance reviews, organisational readiness assessments, and lessons from recent AI failure case studies across public and private sector organisations, this presentation examines how healthcare organisations can operationalise AI governance in practice.

Key areas of focus include:
• Establishing clear ownership and accountability across the AI lifecycle
• Developing inventories of AI systems, data assets, and use cases
• Implementing appropriate human oversight for higher-risk applications
• Educating staff on responsible AI use, limitations, risks, and secure practices

Outcomes

Organisations that adopt structured AI governance practices report improved visibility of AI use, stronger executive oversight, clearer accountability, reduced compliance risk, and increased confidence in deploying AI-enabled solutions.

The Australian AI Governance Standard provides a practical pathway for identifying high-risk AI use cases, implementing proportionate controls, and demonstrating responsible AI practices to regulators, stakeholders, and the community.

Future Actions

Public health and healthcare organisations should assess their current AI use and establish governance mechanisms that support safe, transparent, and accountable adoption. Early implementation of Australia's AI Governance Standard can help organisations realise the benefits of AI while protecting public trust and ensuring equitable outcomes for the communities they serve.

Background and Aim: Generative AI (GenAI) has been promoted as a means of significantly enhancing public health practice. However, the ethical implications of using GenAI in response to outbreak investigations have not been explored in depth. To better understand these issues, we conducted a narrative review of ethical issues relating to the use of GenAI in outbreak investigations.

Methods and Analysis: We reviewed the peer-reviewed and grey literature to identify studies related to the ethics and use of GenAI in outbreak investigations. We identified common themes relating to ethical concerns over GenAI use across three phases of outbreak investigation – detection, investigation and response.

Outcomes: Of the 3557 studies reviewed, 37 met criteria for inclusion. 18 of the studies examined GenAI use for outbreak detection, 14 for investigation, and 23 for response. Most ethical issues related to data privacy and security, data ownership and consent, algorithmic transparency and accountability, lack of appropriate governance, data quality, and biases and limited accessibility relating to underrepresented groups. No study looked at the use of GenAI in outbreak investigation preparation, including personnel or resource planning. In addition, these studies largely focused on GenAI’s applicability and related ethical considerations with respect to major outbreaks like COVID-19, while none evaluated GenAI’s role for more routine outbreak investigations.

Conclusion and Future Actions: While there are substantial ethical concerns regarding the use of GenAI during the major phases of an outbreak investigation, there exists no definitive assessment of the ethics of using GenAI across the entire span of an investigation. Where discussed, ethics are relegated to secondary consideration. These findings underscore the need for future outbreak investigation reports to not only document GenAI use but also ethical concerns regarding its use and appropriate solutions.

Background and Aim:

There is increasing interest in artificial intelligence (AI) and machine learning (ML) to enhance data-driven decision-making and planning. However, biases in design and implementation risks producing BADDR (Blaming, Aggregated, Decontextualised, Deficit-based, and Restricted access) approaches, perpetuating inequities for Aboriginal and Torres Strait Islander peoples.

Indigenous Data Sovereignty (IDSov) and Indigenous Data Governance (IDGov) principles redress these colonising data practices and are fundamental to the health and wellbeing of Aboriginal and Torres Strait Islander people. Embedding IDSov and IDGov (collectively IDSov/Gov) principles was, therefore, central to SMART-PH from the outset.

SMART-PH (DigitiSing InforMAtion for PRacTice in Public Health) involves building a state-wide Public Health Data Lake integrating public health data assets in South Australia’s Digital Analytics Platform. It will feature a suite of AI/ML tools including, but not limited to, data visualisation, automated reporting, scenario and predictive modelling capabilities. Consistent with Carlson and Worrell’s assertion that “AI not just a technical issue, but a relational, cultural, and political one”, this project adopts multilayered approaches to embedding IDSov/Gov.

Methods and Analysis:

Approaches include having a lead Aboriginal investigator and leadership on the Steering Committee and an Indigenous Data Council (IDC) to provide critical oversight, governance, and strategic advice to ensure cultural safety and ethical engagement with Indigenous Data. Further, the project includes Aboriginal HREC-approved empirical data collection to identify public health priorities and data needs, wants, and gaps for organisations working with Indigenous Data and multi-sectoral collaborative efforts to understand Aboriginal communities’ perspectives about data sovereignty in AI-enabled contexts.

Outcomes:

Early outcomes demonstrate how IDSov/Gov principles can be operationalised from the outset in AI-enabled public health infrastructure.

Conclusion and Future actions:

Embedding IDSov/Gov is critical to ensuring equity focused AI advances. By sharing our work, we provide best practice approaches surrounding IDSov/Gov for others working with AI in public health.

Background:
Survival determinants and racial disparities in early-onset small intestine cancer (ages 20-49) remain poorly described. We developed and compared Cox and Random Forest Survival models using SEER data and assessed racial survival differences.
Methods:
We identified 5,217 patients aged 20-49 diagnosed with malignant small intestine cancer (ICD-O-3 C17.0-C17.9) between 2004 and 2020 from the SEER 17-registry database. Missing data were handled with MICE (m=5) and data split 70/30 into training (n=3,651) and test (n=1,566) sets. Random Forest hyperparameters were tuned via 5-fold cross-validation grid search. Both models were evaluated on the same held-out test set using the C-index, calibration plots, Brier scores, and Decision Curve Analysis at 36 months.
Results:
Over 96.3 months median follow-up, 1,556 deaths occurred (29.8%). Non-Hispanic Black patients had shorter mean survival than Non-Hispanic White patients (85.6 vs 100.1 months, p<0.0001) and 32% more deaths than expected (O/E ratio 1.32). Distant stage (HR 3.73), regional stage (HR 1.89), and male sex (HR 1.33) were the strongest predictors of mortality (all p<0.0001). Race was not independently associated with survival after adjusting for stage and treatment, indicating the survival gap operates through later stage at presentation. The optimal Random Forest used mtry=2, node size=10. On the held-out test set, Cox achieved C-index 0.690 and Random Forest 0.693. The Cox model showed good calibration (AUC 73.7, Brier 12.4) and net clinical benefit across threshold probabilities of 5-45%.
Conclusions:
Both models performed comparably with acceptable discrimination. The racial survival gap is driven by stage at diagnosis, pointing to earlier detection as the primary lever for reducing disparities. These findings underscore the importance of evaluating AI survival models for racial equity to ensure fair and equitable clinical application.

Background and Aim
Classifying a consumer health tool as a medical device settles whether it is regulated; it leaves open the degree. Australia's risk-based system already scales the evidence it requires to a device's risk class. The unsettled question is whether a diagnostic role must keep a clinician reviewing its output. Treating supervision as a proxy for safety has a concrete cost: where clinicians are scarce, a supervised-only path means delayed triage and foregone screening for rural and underserved patients. The aim is to identify when supervision should be mandatory and when demonstrated performance is satisfactory.

Methods and Analysis
This analysis is a legal-policy comparison. Abroad the trend is toward mandated oversight: the EU AI Act's Article 14 will require human oversight of high-risk medical AI from 2026. Australia's scheme is so far silent on supervision, and autonomous tools have already been cleared overseas without a human reviewer. The governance question is whether a tool performs to the standard its risk demands, a matter of evidence about the tool, separate from whether a human stands beside it.

Outcomes
The principle already operates in practice. In the United States, LumineticsCore was authorised in 2018 to return a diabetic-retinopathy screening result with no clinician interpreting the image, after a 900-patient primary-care trial (87.4% sensitivity, 89.5% specificity). It runs in primary care, removing a specialist bottleneck. Autonomy can be granted on evidence; the open question is how far up the risk scale that licence should extend.

Conclusion and Future actions
Diagnostic classification should set a performance bar, met with a clinician or without, rather than importing supervision as a reflex. Regulators should publish a demanding evidence threshold at which autonomous operation is permitted for each risk class, so validated tools reach rural and underserved patients who have the least access to care.

Background and Aim
Public health discussions on artificial intelligence (AI) have largely focused on clinical care and service delivery, with little attention to how AI shapes sexual wellbeing. This gap is significant for people with physical disability, who experience barriers to sexual wellbeing, intimacy, and relationships in care contexts where sexuality is often overlooked or constrained. AI systems used to support sexuality and intimacy, such as romantic chatbots, virtual companions, and emerging autonomous sex technologies, may create opportunities to support sexual health, autonomy, pleasure, identity, education, and social connection. However, their implications for public health, equity, safety, and governance remain underexplored. This research examines how AI systems can support sexuality for people with physical disability in care settings while identifying and addressing associated risks.

Methods and Analysis
This human-computer interaction research employs an interdisciplinary approach drawing on digital health and disability studies. Guided by value sensitive design, queer theory, and care ethics, this work combines a systematic review, lived experience consultation, surveys, and co-design. It examines stakeholder values, opportunities, risks, implementation challenges, and governance requirements associated with AI systems used to support sexuality and intimacy in physical disability support contexts.

Outcomes
This research will generate evidence and a practical framework to inform the design, implementation, and governance of these AI systems. Expected outputs include recommendations addressing accessibility, autonomy, privacy, bias, stigma, safety, and sexual wellbeing, with implications for disability policy, care practice, and public health governance.

Conclusion and Future actions
AI systems supporting sexuality and intimacy carry potential to advance equity for people with physical disability, but their public health implications remain unknown. Future action should prioritise inclusive co-design, ethical governance, workforce capability, and policy frameworks that ensure these technologies are safe and empowering. This research contributes to an emerging public health agenda on AI, sexuality, and disability.

Background and Aim
People ask AI about their health millions of times each day. No appointment. No referral. No clinical relationship. The governance debate has focused on purpose-built health apps regulated as medical devices, a frame that assumes clinical intervention rather than population contact. Most AI health contact does not sit at that end.
In early 2026, ChatGPT, Copilot, Perplexity, and others launched health functions integrating medical records, wearables, and lab results. Apple's Health app already embeds clinically-adjacent screening tools on billions of devices. None carry mandatory adverse event reporting obligations or medical device classification, a status many platforms deliberately maintain by positioning as wellness rather than health applications. And none of this is the sharpest problem. Open source AI models create a category of AI health contact where no actor exists to mandate reporting from. The question the field is not asking is who actually can act.

Methods and Analysis
Regulatory frameworks were examined against three governance requirements: mandatory adverse event reporting, population-level monitoring, and cross-platform harm detection. Frameworks examined include the TGA, AHPRA, the EU AI Act, and FDA medical device pathways. None meets them for general AI health contact.
Voluntary incident reporting infrastructure was examined, including the AI Incident Database documenting over 1,300 incidents including mental health harms, and the MIT AI Incident Tracker. Both confirm proof-of-concept surveillance exists but neither mandates participation. A 2026 Danish clinical surveillance study of 53,000 patients identified AI chatbot harms at an increasing rate.

Outcomes
The governance debate is aimed at the wrong target. Here is what the analysis reveals.
In January 2026, a Guardian investigation found Google's AI Overviews, reaching approximately 2 billion people monthly, were providing dangerous health misinformation. Google removed some summaries in response to media pressure, with no mandatory reporting, no regulatory notification, and no adverse event pathway triggered. This is not a failure of a purpose-built health app. It is what the accountability gap predicts will keep happening.
Domestic device regulators cannot reach general AI systems operating across borders. AHPRA reaches registered practitioners, not the general AI assistants people use without a practitioner in the loop. The TGA's regulatory trigger is intended therapeutic purpose, which general assistants do not claim. New federal legislation, not extension of existing mechanisms, would be required for any mandatory AI health adverse event reporting framework in Australia. Populations least served by existing health infrastructure are most exposed to this unmonitored contact, and neither risks nor benefits can be measured without mandatory reporting infrastructure. It is a predictable consequence of asking the wrong actors to solve a problem outside their authority.

Conclusion and Future Actions
Pharmacovigilance offers the model. In 1963, a World Health Assembly resolution led to a global adverse drug reaction reporting system now with 160 member countries. Professional bodies did not build it. Governments cooperating through an international institution did.
Global consensus is unlikely. The Trump administration rebranded the US AI Safety Institute in 2026, reorienting it away from safety governance, and US participation cannot be assumed. A coalition of willing jurisdictions could act without waiting, establishing what reporting looks like and making non-participation a visible political choice rather than an invisible governance gap.
Three actions are required. New domestic legislation creating AI health adverse event reporting obligations. Bilateral agreements with coalition jurisdictions to pool incident data. Active Australian participation in AI health adverse event reporting governance, where action has not yet followed recommendation. Australia has the domestic precedent, institutional relationships through the International Network of AI Safety Institutes, and a government roundtable recommendation already made. The pathway exists. What is missing is the political will to follow it.

Background and Aim: GenAI is rapidly transforming higher education by offering new learning opportunities while raising concerns about academic integrity, equity, and responsible use. International students, who often navigate linguistic and cultural transitions, may experience unique benefits and risks when using GenAI. Understanding these experiences is important for supporting equitable and responsible AI use, informing institutional policies, and promoting inclusive learning environments in higher education. This systematic review aimed to critically examine the benefits, risks, and institutional responses to GenAI among international students in higher education.
Methods and Analysis: Following PRISMA-P guidelines, seven electronic databases (Web of Science, Scopus, ERIC, IEEE Xplore, Educational Research Complete, Education Database, and MEDLINE) were searched for studies published between 2022 and 2026. Three reviewers independently screened studies, with data synthesised thematically and methodological quality assessed using the Mixed Methods Appraisal Tool.
Outcomes: Twenty-three studies met the inclusion criteria. ChatGPT was the most frequently examined GenAI tool. Reported benefits included academic writing and language support, enhanced engagement, confidence and motivation, personalised learning, and improved research efficiency. Key risks included academic integrity concerns, over-reliance on GenAI, bias and cultural or linguistic limitations, ethical concerns, accuracy and trust issues, and potential negative effects on learning and skill development. No comprehensive GenAI policies or guidelines specifically tailored to the needs of international students were identified.
Conclusion and Future Actions: GenAI offers substantial opportunities to support international students’ academic writing, language development, and engagement in higher education. However, concerns regarding academic integrity, overreliance, and algorithmic bias remain significant challenges. The absence of GenAI policies or guidelines specifically designed for international students represents an important gap, underscoring the need for context-sensitive institutional frameworks that address linguistic, cultural, and educational equity considerations.

Background and Aim: Artificial intelligence (AI) is increasingly being used to identify risk, support clinical decision-making, and address inequities in healthcare. However, AI systems are developed and evaluated using definitions of need, harm, and quality that are shaped by human interpretation and institutional priorities. Critical Race Theory and Critical Disability Studies suggest that experiences of racism, ableism, and other forms of oppression affecting equity-seeking communities are often rendered invisible within dominant healthcare systems. This study examines how lived and professional positionality shape understandings of medical bias, unmet health needs, and healthcare harms, and considers the implications for equitable AI development.

Methods and Analysis: Semi-structured qualitative interviews were conducted with 20 participants whose lived and professional experiences provided insight into how healthcare inequities are experienced, recognized, documented, and addressed. Participants included individuals from equity-seeking communities and others engaged in addressing healthcare inequities. Reflexive thematic analysis was used to explore how participants understood healthcare harms and the factors that shaped their interpretations.

Outcomes: Participants identified different but overlapping forms of healthcare harm. Across interviews, participants emphasized fragmented care, barriers to access, self-advocacy burdens, continuity of care, and experiences of discrimination, harmful assumptions, and dismissal within healthcare encounters. Participants also highlighted the importance of trust-building and attention to patients’ social realities. Experiences of marginalization, community connection, and exposure to health equity perspectives influenced participants’ capacity to recognize structural determinants of health and healthcare inequities. These findings suggest that positionality shapes not only experiences of healthcare but also what individuals recognize as evidence of harm.

Conclusion and Future Actions: As healthcare organizations increasingly adopt AI tools to identify, predict, and address inequities, questions remain about whose understandings of harm are embedded within these systems. Findings suggest that community-generated knowledge and diverse forms of lived and professional expertise may be essential for identifying healthcare harms that are often absent from clinical documentation and institutional indicators. Future AI governance, auditing, and development efforts should consider how positionality shapes the recognition of bias, unmet health needs, and inequity, and how communities affected by healthcare inequities can meaningfully inform the design and evaluation of AI systems.