gerstung-lab/Delphi

What it solves

Delphi addresses the challenge of modeling the "natural history" of human disease—predicting how diseases progress and the timing of future health events for individual patients based on their medical history.

How it works

The project uses a modified GPT-2 architecture (based on nanoGPT) to treat patient health trajectories as sequences. It was trained on 400,000 patient trajectories from the UK Biobank, learning to predict future disease events and their timing. The system includes tools for evaluating prediction accuracy, analyzing the latent space of disease embeddings using UMAP, and using SHAP values to determine which prior health events most strongly influence specific future predictions.

Who it’s for

Medical researchers and bioinformaticians interested in epidemiological modeling, disease progression, and the use of generative transformers for health data analysis.

Highlights

• Generative Transformer Approach: Applies a GPT-2 style model to predict competing risks and timings of diseases.
• Interpretability Tools: Includes notebooks for SHAP analysis to identify key disease drivers and UMAP for visualizing disease embeddings.
• UK Biobank Integration: Provides specific guides and scripts for processing large-scale real-world health data.
• Synthetic Data Generation: Capable of generating statistically similar synthetic patient trajectories for testing and research.