HEALTH INTELLIGENCE SYSTEM · 2024
A graph-based risk stratification system that connects longitudinal health, claims and service-use data to identify patients who may benefit from earlier, better-targeted care.
01 / THE DECISION
Chronic conditions emerge through patterns: diagnoses, medications, service utilisation, demographic context and the pathways patients take through care.
Conventional models usually treat each patient as an isolated row. This project reframed the data as a dynamic network, capturing relationships between patients, conditions, providers and episodes of care.
02 / THE SYSTEM
From fragmented records to an action-ready risk signal.
Securely prepare longitudinal clinical, claims and service-use data.
Represent people, events and care relationships as a heterogeneous graph.
Train temporal graph models to surface emerging risk patterns.
Route calibrated risk signals into prioritisation workflows.
03 / EVALUATION
Published evaluations compare network-informed models against established tabular baselines, with a focus on useful representation of comorbidity and patient similarity.
accuracy for cardiovascular disease prediction
accuracy for chronic pulmonary disease prediction
AUC range in the type 2 diabetes study
Reported results are drawn from the linked peer-reviewed studies. Production performance should be revalidated for each population, care setting and intended decision.
04 / RESEARCH FOUNDATION
Multiple peer-reviewed studies translate graph methods into practical risk-modelling capabilities. The program brings academic rigour and industry context together through collaboration between the University of Sydney and CBHS (Commonwealth Bank Health Society).
SCIENTIFIC REPORTS · 2021
Introduced a weighted patient network and GNN framework for cardiovascular and chronic pulmonary disease prediction using de-identified administrative claims data.
APPLIED INTELLIGENCE · 2022
Combined patient-network attributes with machine learning to improve type 2 diabetes risk prediction and identify the most informative network signals.
APPLIED INTELLIGENCE · 2022
Applied disease networks and recommender systems to rank likely comorbidities and support more targeted care-pathway planning.
HEALTH INFORMATION SCIENCE AND SYSTEMS · 2023
Evaluated graph-embedding and link-prediction approaches to surface potential comorbidity relationships earlier in the care journey.
05 / OUTPUT
Rather than producing a black-box score, the interface surfaces a reviewable group of patients, the associated drivers and a practical route to action.
06 / PROJECT PRINCIPLE
Better prediction is only useful when it makes the next decision clearer.Discuss a health AI project ↗