Vertical Federated Learning (vFL) Pilot for Type 2 Diabetes Risk Prediction Using Linked HELIOS and TRUST EHR Data
Main Applicant – Prof John Chambers, Professor of Cardiovascular Epidemiology and Director of Health Screening Centre, Lee Kong Chian School of Medicine, Nanyang Technological University (NTU)

This pilot aims to demonstrate a privacy-preserving workflow for predicting Type 2 Diabetes (T2D) onset by integrating health data held in different secure environments without pooling underlying participant-level data.

The HELIOS study (a subset of SG100K) team at NTU LKC Medicine holds rich research phenotypes (demographics, lifestyle questionnaires, genomics), while TRUST holds Electronic Health Record (EHR) outcomes (diagnosis codes, medications, laboratory tests).

We will implement a Vertical Federated Learning (vFL) workflow using the FLWR (Flower) framework. Each node computes model updates locally and exchanges only encrypted partial predictions or gradients, coordinated by the TRUST/enTRUST environment.

We will train a federated prediction models to identify factors associated with T2D using the distributed features. This project validates a secure infrastructure for precision medicine, enabling the development of robust risk models for chronic diseases like diabetes while strictly adhering to data residency and privacy governance.