Pyhealth

If you're building clinical ML models, this gives you the full stack: 10+ healthcare datasets (MIMIC-III/IV, eICU, OMOP), 20+ predefined prediction tasks (mortality, readmission, drug recommendation), and 33+ models including healthcare-specific architectures like RETAIN and SafeDrug. The medical coding translation alone is worth it, handling ICD-9/10, NDC, ATC, and RxNorm conversions that would otherwise take days to implement. It's opinionated about the five-stage pipeline (load, task, model, train, deploy) but that structure actually makes sense for clinical work. Claims 3x faster than pandas for healthcare data processing. The fairness metrics and calibration tools show they're thinking about real deployment, not just academic benchmarks.