NYUMedML/DeepEHR

Implements hierarchical CNN-LSTM architectures with multi-task learning to process variable-length clinical notes and structured EHR data for predicting multiple chronic diseases simultaneously. Leverages pre-trained word embeddings (via Starspace) and encounter-level document encoders that capture both local n-gram patterns and sequential dependencies across medical encounters. Built on PyTorch 0.4 with modular training pipelines supporting multiple model variants, lab value extraction, and evaluation across imbalanced disease cohorts.

272 stars. No commits in the last 6 months.