Reproduce “Doctor AI: Predicting Clinical Events via Recurrent Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Edward Choi, Mohammad Taha Bahadori, Andy Schuetz, Walter F. Stewart, Jimeng Sun
Doctor AI, developed by researchers at Georgia Tech and Sutter Health, employs Gated Recurrent Units (GRUs) to predict future diagnoses, medication orders, and time to next patient visit using longitudinal electronic health record data. The model achieved up to 79.58% recall@30 for diagnosis code prediction and an R^2 of 0.25 for next visit time prediction, while also demonstrating successful knowledge transfer between different hospital datasets.
Key insights
- Recurrent Neural Networks, particularly GRUs, are effective in learning robust patient representations and temporal patterns from longitudinal EHR data, enabling accurate multi-label prediction of future clinical events.
- Initializing medical code embeddings with Skip-gram pre-training can enhance model performance and potentially accelerate learning in the context of EHR sequence modeling.
- Knowledge transfer is feasible; models pre-trained on large EHR datasets can be fine-tuned on smaller, different institutional datasets, leading to substantial performance gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.