PyHealth Clinical Deep Learning Pipeline

What it does

PyHealth is a Python toolkit for clinical deep learning. This skill teaches Claude to write idiomatic PyHealth code that follows the library's 5-stage pipeline (Dataset → Task → Model → Trainer → Metrics).

• Dataset loading: MIMIC-III/IV, eICU, OMOP-CDM, EHRShot, SleepEDF, SHHS, ISRUC, ChestX-ray14, COVID19-CXR, TUEV/TUAB
• Task definition: mortality, readmission, length of stay, drug recommendation, sleep staging, ICD coding, EEG event detection, de-identification
• Model selection: Transformer, RETAIN, GAMENet, SafeDrug, MICRON, StageNet, AdaCare, CNN/RNN/MLP
• Training & evaluation: PyHealth Trainer handles checkpointing, logging and best-model selection; clinical metrics like PR-AUC, ROC-AUC, Jaccard
• Medical code utilities: lookup and cross-mapping for ICD-9/10-CM, ATC, NDC, RxNorm, CCS

The most valuable part is the "critical things to get right" list — pass a SampleDataset (not BaseDataset) to models, always split by patient to avoid leakage, MIMIC-IV uses ehr_root=, and pick a monitor metric matching the task type.

Who it's for

• Healthcare AI researchers and grad students modeling EHR data from MIMIC/eICU
• Data scientists prototyping hospital risk-prediction models (mortality, readmission)
• Health informatics practitioners doing frequent ICD/ATC/NDC code mapping
• Anyone building signal or imaging classification pipelines (polysomnography, EEG, chest X-ray)

Example uses

1. "Build a mortality prediction model on synthetic MIMIC-III in under 20 lines" → generates a script using the public synthetic bucket with MortalityPredictionMIMIC3 + Transformer + Trainer, and sets monitor="pr_auc" correctly.
2. "Create a drug recommendation pipeline on MIMIC-IV and compare SafeDrug vs GAMENet" → uses multilabel-appropriate metrics (jaccard_samples, pr_auc_samples) and avoids the ehr_root= constructor gotcha.
3. "Map NDC codes in my prescription table to ATC level 3" → produces medcode-based NDC→RxNorm→ATC cross-mapping code.

Install the skill I found on Claude Skill Mart.
Copy the skills/pyhealth folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/pyhealth/.
When it's done, tell me in one line what this skill can do.