pyhealth

Trigger phrases

Phrases that activate this skill when typed to Claude Code:

• EHR machine learning
• MIMIC dataset
• clinical prediction model
• readmission prediction
• drug recommendation model

What it does

pyhealth is a Claude Code skill from K-Dense AI’s scientific-agent-skills repo. It turns Claude into a PyHealth expert for healthcare ML — covering EHR dataset loading and processing (MIMIC-III/IV, eICU, OMOP CDM), medical coding systems (ICD-9/10, NDC, ATC codes), clinical prediction tasks (in-hospital mortality, 30-day readmission, drug recommendation, length of stay), physiological signal processing (EEG, ECG), and deep learning model implementations (RETAIN, SafeDrug, Transformer-based clinical models, GNNs for drug-drug interactions).

A session produces a complete clinical ML pipeline: dataset loading from structured EHR data, task definition, model training, and evaluation with clinically appropriate metrics (AUROC, AUPRC).

When to use it

Reach for it when:

• You’re working with MIMIC-III/IV, eICU, or OMOP CDM data and need a standardized preprocessing pipeline that handles the clinical data complexity
• You want established deep learning models for clinical prediction tasks without implementing them from scratch
• You’re benchmarking a new clinical prediction model against established baselines (RETAIN, Transformer, GRU-based models)

When not to reach for it:

• DICOM medical imaging — use pydicom
• Clinical text NLP (clinical NLP, note processing) — use transformers with a clinical BERT variant

Install

Copy the SKILL.md from K-Dense AI’s pyhealth folder into .claude/skills/pyhealth/ in your project. MIMIC and eICU datasets require PhysioNet data access agreements; PyHealth handles the parsing once you have the raw CSV files.

Trigger phrases: “EHR machine learning”, “MIMIC dataset”, “clinical prediction model”, “readmission prediction”.

What a session looks like

A typical session has three phases:

1. Dataset and task setup. Specify the dataset (MIMIC-III, eICU) and the prediction task. Claude sets up the PyHealth dataset object that handles patient visit aggregation, medical code mapping, and train/validation/test splitting by patient ID (not by visit).
2. Model selection and training. Claude selects an appropriate model architecture from PyHealth’s model zoo — RETAIN for interpretability, Transformer for sequence modeling, GNN for drug-drug interaction tasks — and configures the training loop with appropriate loss function and evaluation metrics.
3. Evaluation. Results are computed with AUROC and AUPRC (both critical for imbalanced clinical outcomes), and Claude flags any data leakage risks in the splitting strategy specific to the clinical task.

Receipts

Where it works well:

• Benchmarking clinical prediction models on MIMIC — PyHealth’s standardized task definitions and data processing make comparison to published results meaningful rather than confounded by data processing differences
• Multi-task clinical prediction where one model needs to predict multiple outcomes — PyHealth’s task definition layer handles this cleanly

Where it backfires:

• Clinical datasets have severe class imbalance (in-hospital mortality rates of 5–15%) that requires careful metric selection; models that optimize accuracy rather than AUROC/AUPRC look good on paper but are clinically useless
• PyHealth’s data loading assumes specific CSV formats from MIMIC; minor version differences in the downloaded files require preprocessing fixes that aren’t always documented

Pattern that works: always split by patient ID, not by visit — splitting by visit allows the same patient to appear in train and test sets, which artificially inflates performance and produces models that don’t generalize.

Source and attribution

Originally authored by K-Dense Inc.. The canonical SKILL.md lives in the pyhealth folder of their public scientific-agent-skills repository.

License: MIT. Install, adapt, and redistribute with attribution preserved.

This page documents the skill from a practitioner’s perspective. For the formal spec and any updates, defer to the source repo.