-
Notifications
You must be signed in to change notification settings - Fork 0
Home
DelphiTrajectories is a fully client-side, privacy-preserving web application that runs the Delphi 2M generative transformer model entirely in the browser to predict personalized disease trajectories from patient medical histories. No data ever leaves the user's device, no installation required, no server inference
Live demo: https://epiverse.github.io/delphiTrajectories/
GitHub repository: https://github.com/epiverse/delphiTrajectories
Associated paper (2025): "Engineering FAIR privacy-preserving applications that learn from natural histories to predict clinical outcomes" – Duarte I, Bhawsar P, Mason L, Balasubramanian J, Russ D, Oliveira A, Almeida J.
DelphiTrajectories is a proof-of-concept web application that ports the Delphi 2M model (Shmatko, Ghaffari Laleh, Gerstung & Kather et al., Nature 2025) from its original PyTorch/server environment to a pure client-side JavaScript/WebAssembly execution stack using ONNX runtime web.
The entire inference runs locally in the user's browser, which makes the application inherently privacy-preserving: sensitive health data never travels over the network for prediction, satisfying the data governance requirements.
The project was specifically designed as a real-world test of the Reusability (R) and Interoperability (I) components of the FAIR principles, using only the public documentation and code from the original Delphi 2M authors. Current model: trained on ~10k synthetic trajectories (smaller than the full 400k model in the original paper, therefore reduced predictive performance but sufficient for demonstration).
Traditional deployment of large generative models for health data requires users to upload sensitive records to cloud servers which leads to regulatory, ethical, and trust barriers.
This project eliminates that barrier completely by moving all computation to the client:
- Absolute privacy: health data never leaves the device
- No installation needed: works in any browser
- FAIR compliance: transformer logic is truly reusable without hand-holding
- Democratizes access: anyone with a browser can experiment the model