Background and aim: Evidence regarding the safety of language models used in personalised patient communications remains limited. Retrieval-augmented generation (RAG) frameworks are designed to focus responses on trusted information sources and are frequently assumed to improve safety. We evaluated whether augmenting a personalised question-answering (QA) system with context-specific knowledge reduced potential safety risks when responding to patient questions following hospital discharge.

Methods and analysis: We examined eight configurations of a QA system designed to respond to patient questions following hospital discharge. Clinical experts evaluated safety risks for responses to 111 questions compared with answers written by other clinical experts.

Outcomes: The overall proportion of potentially unsafe responses was similar between clinical experts (6.3%) and QA systems with augmented GPT-4o (9.9%, p=0.50) and Qwen-2.5 (3.6%, p=0.57) configurations. Retrieval augmentation did not consistently improve safety. There was no evidence that augmenting Qwen-2.5 with context-specific knowledge improved safety (9.9% baseline, 4.5% discharge summaries, 4.0% synthetic QA examples, 3.6% both knowledge bases; p=0.158 by Kruskal-Wallis), and augmented GPT-4o configurations were found to be less safe than the baseline (0.9%, 9.5%, 9.9%, 9.9%; p=0.026). Compared with clinician-written answers (0.9% of responses in higher categories of risk), augmented GPT-4o configurations produced more responses in higher categories of safety risk (2.7%).

Conclusion and future actions: While RAG frameworks augment language models with additional information, they do not guarantee safety or reduce the severity of safety risks. Safety varied across model configurations and clinical contexts, highlighting the need for systematic safety evaluation and ongoing monitoring before deployment. When developing patient-facing QA systems, developers may need to balance the provision of additional information in responses against potential increases in the frequency and severity of safety risks.

Background and Aim
The integration of artificial intelligence (AI) into mammogram reading offers significant opportunities to enhance breast cancer screening while introducing new operational and clinical considerations. BreastScreen Victoria (BSV) is currently progressing three strategically aligned research trials designed to build organisational capability and generate robust evidence to support future AI-enabled mammography reading within the screening program. This presentation will focus on one of these initiatives: a Randomised Controlled Trial (RCT) evaluating whether the implementation of an AI mammogram reader can improve breast cancer screening outcomes. This RCT is a first for Australia and examines clinical performance and operational impacts, as well as ethical and social impacts.

Methods and Analysis
The current standard of care in breast cancer screening is two radiologists independently reading mammogram images. The RCT compares this to a model where an AI reader replaces one radiologist.

Use of an AI reader can improve accuracy, address workforce challenges, and improve the client experience, enabling the faster provision of results and less waiting time if a client requires further tests.

Outcomes
The AI reader used in the RCT has been developed in Australia and a retrospective study demonstrates its strong performance. Early results from the RCT have reinforced the AI reader’s performance, along with the importance of retaining the involvement of radiologists.

There is a need for strong governance when using AI in healthcare. There was significant work done to prepare for the implementation of the RCT, with project streams including legal, contractual and operational modelling; change management, communications and training; IT integration; and clinical governance.

Over 200,000 BreastScreen clients will participate in the RCT, with over 35,000 participating to date.

Conclusion and Future actions
Early indications are that an AI reader can bring significant enhancements to the BreastScreen program, and is well accepted by most clients and staff.

Background and Aim
Childhood dementia (CD) encompasses more than 170 rare genetic neurodegenerative disorders affecting 1,500 Australian children. Onset occurs from infancy to adolescence with symptoms beginning at an age of 2.5 years. Early signs may include deterioration in vision, hearing, motor function, cognition, communication or behaviour; loss of acquired abilities; and sometimes seizures or abnormal head growth. Parents often notice unusual development, but may not know to report these changes to GPs and/or other health professionals, often leaving behavioural patterns unrecognised. Many studies reported diagnostic delays of at least two years; three reported six years or more. We present an approach that could substantially shorten the time to diagnosis.

Methods and Analysis
An expert round-table examined symptoms of developmental regression in childhood dementia, family diagnostic experiences, rare-disease pathways as well as the role of health information, artificial intelligence (AI) and clinical decision support (CDS). Findings included the identification of the "Red Flags" concept, plausible signs of CD from teacher observations, GPs, other professionals and parents' stories as first steps to clinical diagnoses by neurologists.

Outcomes
Modern electronic health records (EHRs), extended with non-clinical data and processed with AI agents trained to spot "Red Flags", can plausibly raise CD alerts. Human-in-the-loop AI could support neurologists with case summaries and background information to assist with their exercise of clinical judgement.

Conclusion and Future Actions
The proposed approach needs further analysis and evaluation including approaches to extend EHRs to accept non-clinical observations and the development of specialised AI agents to infer "Red Flag" information and make recommendations. Benefits include earlier recognition of unexpected regression, a shortened time-to-diagnosis, earlier commencement of available treatments, minimized harm, as well as earlier access to genetic counselling, clinical trials and emerging treatments for currently untreatable variants. Finally, an early diagnosis increases parents' precious time with their child.

References:
• Elvidge KL, Christodoulou J, Farrar MA, Tilden D, Maack M, Valeri M, et al.; Childhood Dementia Working Group. The collective burden of childhood dementia: a scoping review. Brain. 2023;146(11):4446–4455. doi:10.1093/brain/awad242.
• Djafar JV, Johnson AM, Elvidge KL, Farrar MA. Childhood dementia: a collective clinical approach to advance therapeutic development and care. Pediatric Neurology. 2023;139:76–85. doi:10.1016/j.pediatrneurol.2022.11.015.
• Nevin SM, McGill BC, Kelada L, Hilton G, Maack M, Elvidge KL, et al. The psychosocial impact of childhood dementia on children and their parents: a systematic review. Orphanet Journal of Rare Diseases. 2023;18(1):277. doi:10.1186/s13023-023-02859-3.
• Furley K, Mehra C, Goin-Kochel RP, Fahey MC, Hunter MF, Williams K, Absoud M. Developmental regression in children: current and future directions. Cortex. 2023;169:5–17. doi:10.1016/j.cortex.2023.09.001.
• Furley K, Hunter MF, Fahey M, Williams K. Diagnostic findings and yield of investigations for children with developmental regression. American Journal of Medical Genetics Part A. 2024;194(8):e63607. doi:10.1002/ajmg.a.63607.
• Furley K, Teo A, Williams K, Alshawsh M, Brignell A. The diagnostic yield of investigating developmental regression in children: a systematic review and meta-analysis. Journal of Autism and Developmental Disorders. Published online February 20, 2025. doi:10.1007/s10803-025-06749-4.
• Furley K, Hunter M, Gawade G, Absoud M, Mehra C, Kochel R, Fahey MC, Williams K. Towards an agreed approach to investigate children with developmental regression. BMJ Paediatrics Open. 2025;9:e003594. doi:10.1136/bmjpo-2025-003594.
• Özçetin E, Baş SE, Özpay F. Architectural and translational perspectives on clinical decision support systems for rare disease diagnosis: a scoping review. International Journal of Medical Informatics. 2026;215:106442. doi:10.1016/j.ijmedinf.2026.106442.
• Hersh WR, Cohen AM, Nguyen MM, Bensching KL, Deloughery TG. Clinical study applying machine learning to detect a rare disease: results and lessons learned. JAMIA Open. 2022;5(2):ooac053. doi:10.1093/jamiaopen/ooac053.
• Dementia Australia. Dementia facts and figures. Updated 2026. Accessed June 25, 2026.

Background and Aim
Keratinocyte cancers (KCs) are the most prevalent cancers in white-skinned individuals, yet remain underrepresented in cancer registries because reporting requirements differ greatly across jurisdictions. Manual extraction of KC subtypes from medical reports is labor-intensive and time-consuming, particularly as reports often document multiple co-excised skin lesions. Artificial intelligence offers automated solutions for disease phenotyping from unstructured clinical text.

Methods and Analysis
We fine-tuned the open-source large language model Meta-AI (LLaMA) 3.1-8B-Instruct on 26,179 manually reviewed pathology reports from 10,326 Australian QSkin Sun and health study participants who had histologically confirmed KCs. Independently validation was performed on 217 pathology reports from the Skin Tumors in Allograft Recipients cohort.

Outcomes
The model achieved F1-scores above 0.90 for the four KC subtypes of interest: squamous cell carcinoma, basal cell carcinoma, keratoacanthoma, and intraepidermal carcinoma. Frequency-weighted mean F1-scores reached 0.84 (95% confidence interval: 0.843-0.846) for diagnosis classification and 0.83 (95% confidence interval: 0.826-0.828) for anatomical site. External validation demonstrated robust performance with F1-scores between 0.73-0.86 for the four KCs of interest.

Conclusion and Future actions
Our fine-tuned model QSkin-llama-3.1-8b works locally. It accurately classifies lesion counts, diagnoses, and anatomical sites. It processes 24 pathology reports per minute with minimal preprocessing, enabling scalable automated disease phenotyping for large health datasets. Our approach makes it possible to analyse large collections of medical data that would otherwise remain unused. This could improve how keratinocyte cancers are monitored and studied, and may help researchers and health systems better understand patterns of disease and plan care more effectively.

Introduction: Automated International Classification of Diseases (ICD) coding in general practice is susceptible to variation. ICD-11 adds complexity through post-coordinated code clusters, while general-practice data are often stored as longitudinal tabular event streams rather than narrative summaries. This study developed and evaluated a local hybrid multi-agent framework for automated dual ICD-10 and ICD-11 coding from these records.

Methods: A two stage pipeline was implemented using Australian clinical summary principles, with all language models (Llama 3, GLM 5.2, and Gemma 4) deployed locally via Ollama to ensure that no patient data were transmitted externally. In Stage 1, these models were combined with deterministic rule based extraction to transform raw tabular event streams into structured patient summaries. In Stage 2, the same models were integrated into a retrieval augmented multi agent coding framework: SapBERT retrieved candidate ICD concepts from indexed corpora, three agents proposed codes independently, and a chairman agent resolved disagreements to select the final code. ICD 11 post coordination was then generated by a hybrid module that combines rules and language models and operates across 13 extension axes.

Results: The pipeline was evaluated on 2,232 general-practice patient records and generated approximately 110,000 ICD-10 and ICD-11 coded terms, with a near-zero unmapped rate. Post-coordination was required for 14.2% of ICD-11 assignments. Expert review of 500 records showed almost perfect agreement for ICD-10 (kappa = 0.87) and substantial agreement for ICD-11 (kappa = 0.68), with 90% acceptance of principal diagnoses. The most frequent conditions were upper respiratory tract infections and hypertension.

Conclusion: Automated dual ICD-10 and ICD-11 coding from longitudinal general-practice records was feasible using a locally deployed retrieval, summarisation and adjudication framework. The approach may support ICD-11 transition planning while preserving ICD-10 compatibility for longitudinal surveillance and population health research.

Background: Cervical cancer causes an estimated 3,200 to 5,000 preventable deaths annually in Pakistan. Despite effective HPV vaccines, uptake is hindered by limited awareness, misinformation, and safety and fertility concerns. This study examined how a Generative AI voice helpline compared with the existing human-staffed helpline in supporting caregivers, including on fertility, sexuality, and religious concerns.

Methods: A two-stage adaptive design was used. Stage 1 (July to August 2025) optimized the GenAI helpline through 25 caregiver interviews; development of a RAG-grounded Q&A library validated by HPV experts and cultural advisors; a multi-arm bandit voice experiment (approximately 1,800 GenAI calls); and a 100-caregiver system pilot (SUS at least 70; accuracy at least 95%). Stage 2 (September 2025 to February 2026) employed a quasi-experimental, treatment-on-the-treated comparison of the GenAI and human helplines across three IVR/CATI survey waves: pre-call baseline (n=6,357), post-interaction (n=7,871; GenAI=2,973, Human=4,898), and 7 to 14-day vaccination follow-up (n=5,014). Primary analysis used inverse probability of treatment weighting (IPTW), adjusting for demographics, language, baseline HPV knowledge, vaccination intention, and anxiety. Safety was monitored through weekly clinician-coded transcript audits of GenAI calls, NLP evaluation, and monthly mystery calls.

Results: Before any call, 49% of caregivers lacked sufficient information to decide on vaccination, indicating an information vacuum rather than a refinement opportunity. The GenAI line drew a younger, more rural, less educated, and lower-income caller profile, groups with lower baseline vaccination intent (41.8% vs. 47.6%) and greater access constraints (53% needing household permission vs. 24%), and it drew 2.7 times more Pashto and Balochi speakers. Across the campaign the GenAI line handled 233,808 HPV calls and the human line 92,202. Despite serving a structurally harder population, vaccination rates were similar (78.9% vs. 77.8%; adjusted difference +1.1 pp, within the ±5 pp equivalence margin). Satisfaction was lower for GenAI in crude analysis (−8.4 pp), yet on the adjusted model the two lines were at parity, and within the dominant CATI mode (80%) GenAI was marginally ahead (+1.5 pp), with the reversal due to survey mode rather than helpline performance. Medical accuracy was 97.2% across 1,241 Q&A pairs over 11 audit rounds, with high cultural sensitivity and no hate speech. The 86% grounding rate reflected source content gaps rather than inaccuracy.

Conclusion: Evidence supports a complementary model, GenAI for scalable triage and after-hours access, with human agents handling complex cases, rather than replacement.

Background and Aim Endometriosis is a significant public health challenge, affecting 1 in 7 women. Currently, a debilitating 6.4-year average delay between symptom onset and diagnosis, results in chronic pelvic pain, infertility, and immense economic burdens. Pain induced school and work absences reduce productivity and economic loss has been estimated at ~$21,000-31,000 per year per woman. While laparoscopy has been the traditional diagnostic, recent guidelines encourage imaging. We have developed IMAGENDO®, an AI-driven tool that combines endometriosis ultrasound (eTVUS) and MRI (eMRI) to provide a faster, non-surgical diagnostic pathway.

Methods and Analysis We developed deep learning models using temporal residual networks and 3D Vision Transformers trained on specialist imaging datasets. To address the limitations of single-modality imaging, we employed a knowledge distillation approach. This involves a "teacher" model trained on high-specificity eTVUS data to enhance a "student" model interpreting eMRI scans. To ensure data integrity and mitigate "information bias", models underwent rigorous student auditing and expert radiology review to compensate for imaging artifacts and mislabeling.
Outcomes The integration of multimodal AI significantly improved diagnostic performance. For Pouch of Douglas (POD) obliteration, accuracy on eMRI datasets increased from 65% to a 90.6% AUC. Furthermore, automated models for detecting rectal nodules achieved an 85.5% AUROC for eTVUS. These results demonstrate AI’s ability to maximize public health benefits by bridging the gap in specialist expertise and providing a reliable, non-invasive diagnostic alternative.
Conclusion and Future actions Future efforts will expand the algorithm to include additional markers and larger datasets to ensure global generalisability. To safeguard communities, we must develop shared directions for AI governance and workforce education, ensuring clinicians can ethically integrate these tools. By strengthening cross-sector collaboration, IMAGENDO® aims to promote health equity, providing accessible, high-quality diagnostics that reduce the long-term physical and financial harms of undiagnosed endometriosis.

Background and Aim
Patient-reported outcomes (PROs) are a cornerstone of person-centred cancer care, providing direct insight into symptoms, functioning and quality of life. When collected and acted upon routinely, PROs have demonstrated meaningful clinical benefits, including earlier detection of deterioration and improved survival. However, implementation barriers persist for patients in completing PROs, and for clinicians in interpreting and integrating PRO data into care plans. The process of collecting, analysing, and acting on PROs creates opportunities to harness Artificial Intelligence (AI) and ease the manual burden.
Methods and Analysis
We conducted a scoping review to map how Artificial Intelligence (AI) is being applied to support cancer care informed by PRO data. Data were extracted and structured around the stage of AI model development, the role of PROs, intended clinical use, implementation outcomes (Proctor et al. 2011), and the extent to which ethical and governance considerations were reported in accordance with TRIPOD-AI and CONSORT-AI.
Outcomes
Of 1119 records screened, 67 studies met inclusion. The most common objectives were predicting PROs (n=18), adverse clinical events (n=15) and risk stratification (n=9). Fewer studies addressed survival prediction (n=8), automated data extraction (n=7), clinical decision support (n=4), patient self-management (n=3), or symptom pattern analysis (n=3). Most applications were intended for clinician use (n=64). Critically, most studies reported early model development or validation stages (90%), of which 33% referenced a reporting guideline.
Conclusion and Future Actions
AI is being applied across a range of PRO-related clinical objectives, but evidence remains concentrated in model development. This creates a timely opportunity to examine implementation considerations, clinician and patient attitudes to uptake and workflow readiness before tools are scaled. Public health systems and cancer services should collaborate in the early stages of model development to establish appropriate governance frameworks that enable responsible and safe integration of AI into PRO-informed cancer care.

Background and Aim
Methanol poisoning and toxic alcohol contamination remain significant global public health threats, often resulting in mass casualties, permanent blindness, and death, particularly in settings with limited surveillance capacity. This study aimed to evaluate the performance of a large language model (GPT-4o) for automated detection and structured extraction of methanol and toxic alcohol events from open-source intelligence (OSINT) data.
Methods and Analysis
News articles published between August 2024 and January 2025 were collected using the Bing News Application Programming Interface (API) with a curated set of methanol- and toxic-alcohol-related search terms. A hierarchical zero-shot prompting framework was applied to classify events and extract epidemiologically relevant entities. Model performance was evaluated against a manually annotated ground-truth dataset of 100 articles using precision, recall, and F₁-score.
Outcomes
GPT-4o demonstrated high performance for event identification (F₁ = 0.98), temporal expressions (F₁ = 0.97), and location extraction (F₁ = 0.95). Numerical entities, including case counts and fatalities, showed high recall but lower precision, reflecting over-identification in narrative contexts. Exact date extraction yielded the lowest performance (F₁ = 0.86), primarily due to false positives from contextual temporal references.
Conclusion and Future Actions
These findings indicate that GPT-4o can support early detection and structured characterisation of methanol poisoning and toxic alcohol events from open-source intelligence. While not a replacement for formal surveillance systems, LLM-assisted OSINT pipelines may enhance situational awareness and rapid risk assessment in regions where traditional reporting is delayed or incomplete.

Background and Aims:
Chronic diseases including diabetes mellitus, hypertension, cardiovascular disease, and obesity remain a major global healthcare burden. Traditional disease monitoring often relies on periodic clinical assessments, which may not capture dynamic physiological changes in everyday life. This study aims to review current evidence regarding the role of wearable technologies and digital biomarkers in advancing personalized chronic disease prevention and management.

Methods:
A literature review was conducted using databases including PubMed, Scopus, and Web of Science to identify relevant studies investigating wearable devices, digital biomarkers, and artificial intelligence-based health monitoring in chronic disease care. Studies were selected based on clinical relevance and recent evidence in the field.

Results:
Current evidence demonstrates increasing applications of wearable technologies in chronic disease management. Devices capable of monitoring parameters such as heart rate, heart rate variability, sleep, physical activity, and glucose trends have shown potential in improving patient engagement, risk assessment, and individualized interventions. However, limitations including data accuracy, privacy concerns, accessibility, and the need for further clinical validation remain.

Conclusions:
Wearable technologies represent a promising advancement in modern healthcare, enabling a shift from reactive disease treatment toward proactive, continuous, and personalized care. Further research is required to establish standardized integration into routine clinical practice.

Title: Evaluating an AI assistant to scale equitable, gold-standard access to genetic testing.

Background and Aim: Demand for clinical genetics and precision medicine is growing far faster than the genetic counselling workforce can meet. Long waiting lists and inequitable geographic and cultural access leave many people — particularly in rural, regional and culturally diverse communities — without timely, high-quality pre-test genetic education and counselling. Scaling these with AI technology risks trading access for quality; any digital solution must match, not dilute, the gold standard.
Methods and Analysis: We developed an AI-powered conversational assistant ("Tara") that provides plain-language pre-test genetic education and guides the testing pathway — consent, family-history collection and scheduling — with mandatory hand-off to human genetic counsellors when clinical judgement is required. Built-in guardrails prevent it from interpreting personal results or offering medical, legal or reproductive advice.
Using structured rubrics, independent accredited-genetic counsellor review, blinded AI-versus-human comparison and a pre-specified non-inferiority margin, we designed a multi-phase evaluation process. The AI assistant was benchmarked against gold-standard genetic counselling across pre-specified domains to determine:
• Clinical accuracy
• Appropriate escalation
• Safety
• Comprehension and
• User experience.
Outcomes: The assistant is being assessed against pre-defined thresholds for accuracy, mandatory escalation and safety, and compared head-to-head with human counsellors under a non-inferiority framework. Findings from the completed internal and external review phases and early deployment data will be presented.
Conclusion and Future actions: Rigorous, counsellor-led evaluation against the gold standard is essential before deploying AI in genetics. Designed conservatively and held to clinical benchmarks, AI can widen equitable access and ease workforce pressure without compromising quality — maximising benefit while minimising harm.