Summary Delphi is an open-source project that models human health trajectories using generative transformers, specifically a modified GPT-2 architecture. It aims to learn the natural history of diseases by analyzing longitudinal patient data, enabling researchers to understand disease progression and predict future health events.

How It Works The project leverages Andrej Karpathy's nanoGPT implementation of GPT-2 to process sequential patient health records. It treats disease progression as a generative task, learning patterns from data to predict future states. A key feature is the generation of statistically similar synthetic patient trajectories, facilitating research while preserving privacy.

Quick Start & Requirements Installation requires cloning the repo, setting up a Python 3.11 Conda environment, and running pip install -r requirements.txt. Training uses UK Biobank data (400K trajectories, requires application) or a provided synthetic dataset. Demo training takes ~10 minutes on GPU, original model ~1 GPU-hour (V100), and M1 MPS ~10 hours. A Dockerfile is available. Data preparation notebooks are included.

Highlighted Details

• Notebooks for evaluating prediction accuracy, analyzing attention mechanisms, and UMAP-based latent space exploration.
• SHAP analysis for identifying key disease drivers and predicting future events.
• Tools for comparing generated synthetic data against real-world distributions.
• Example notebooks for transforming raw UK Biobank data into a compatible format.

Maintenance & Community The README does not specify community channels, active contributors, sponsorships, or a roadmap