PyHealth

PyHealth (https://pyhealth.dev/) is a Python toolkit for clinical deep learning. It provides a unified, modular pipeline across electronic health records (EHR), physiological signals, and medical imaging.

The library is built around a 5-stage pipeline — Dataset → Task → Model → Trainer → Metrics — where each stage is replaceable and the interfaces between stages are stable. Code that follows this pipeline shape composes well; code that bypasses it usually fights the library.

When to use this skill

Use this skill whenever the user is doing clinical/healthcare ML and any of the following are true:

• They mention PyHealth, MIMIC-III/IV, eICU, OMOP-CDM, EHRShot, SleepEDF, SHHS, ISRUC, COVID19-CXR, ChestX-ray14, TUEV/TUAB.
• They want to predict mortality, readmission, length of stay, drug recommendations, sleep stages, ICD codes, EEG events, or de-identification.
• They need to look up or cross-map medical codes (ICD-9-CM, ICD-10-CM, ATC, NDC, RxNorm, CCS).
• They have EHR-shaped data and want to train a clinical model without writing the plumbing themselves.

PyHealth is the right tool when the workflow fits its 5 stages. If the user just wants generic PyTorch on tabular data, this skill is not necessary.

Installation (uv)

PyHealth 2.0 requires Python ≥ 3.12, < 3.14. Use uv for environment management — it's faster and reproducible.

# Create a project with the right Python
uv init my-pyhealth-project
cd my-pyhealth-project
uv python pin 3.12

# Add PyHealth (this also pulls in PyTorch and friends)
uv add pyhealth

# Run scripts inside the env
uv run python train.py

For a one-off script without a project, use uv run --with pyhealth python script.py. For the legacy 1.x line (Python 3.9+), uv add pyhealth==1.16. Detailed install notes, MIMIC access, and GPU/CPU device tips are in references/installation.md.

The 5-stage pipeline

A complete pipeline is typically <20 lines. This is the canonical shape — start here and modify pieces:

from pyhealth.datasets import MIMIC3Dataset, split_by_patient, get_dataloader
from pyhealth.tasks import MortalityPredictionMIMIC3
from pyhealth.models import Transformer
from pyhealth.trainer import Trainer
from pyhealth.metrics.binary import binary_metrics_fn

# 1. Dataset — raw patient registry
base = MIMIC3Dataset(
    root="https://storage.googleapis.com/pyhealth/Synthetic_MIMIC-III/",
    tables=["DIAGNOSES_ICD", "PROCEDURES_ICD", "PRESCRIPTIONS"],
)

# 2. Task — converts patients into supervised samples
samples = base.set_task(MortalityPredictionMIMIC3())

# 3. Split + DataLoaders (split by patient to avoid leakage)
train_ds, val_ds, test_ds = split_by_patient(samples, [0.8, 0.1, 0.1])
train_loader = get_dataloader(train_ds, batch_size=32, shuffle=True)
val_loader   = get_dataloader(val_ds,   batch_size=32, shuffle=False)
test_loader  = get_dataloader(test_ds,  batch_size=32, shuffle=False)

# 4. Model — must be passed the SampleDataset, not the BaseDataset
model = Transformer(dataset=samples)

# 5. Train + evaluate
trainer = Trainer(model=model)
trainer.train(
    train_dataloader=train_loader,
    val_dataloader=val_loader,
    epochs=50,
    monitor="pr_auc",
)

y_true, y_prob, _ = trainer.inference(test_loader)
print(binary_metrics_fn(y_true, y_prob, metrics=["pr_auc", "roc_auc"]))

A copy-pasteable starter is in assets/starter_pipeline.py.

Critical things to get right

These are the mistakes that PyHealth code most commonly trips on. Internalize them before writing pipelines:

1. Models take a SampleDataset, not a BaseDataset. MIMIC3Dataset(...) returns a BaseDataset (a queryable patient registry). Only after .set_task(task) do you get a SampleDataset, which is what models, splitters, and DataLoaders expect. If you pass base to a model, it will fail or behave wrong.

2. Always split by patient (or visit), not by sample. Random sample-level splits leak information across train/test because the same patient can appear in both. Use split_by_patient for patient-level prediction, split_by_visit only when visits are independent.

3. Match the task to the dataset. Tasks are dataset-specific: MortalityPredictionMIMIC3 won't work on MIMIC-IV — use MortalityPredictionMIMIC4 or InHospitalMortalityMIMIC4. The full mapping is in references/tasks.md.

4. Pick monitor to match the task type. For binary classification use "pr_auc" or "roc_auc". For multilabel (drug rec) use "pr_auc_samples" or "jaccard_samples". For multiclass use "accuracy" or "f1_macro". Wrong monitor → checkpoint selection saves the wrong epoch.

5. MIMIC-IV uses ehr_root=, not root=. This is the one inconsistency in the dataset constructors.

6. For reproducible work, point cache_dir= somewhere persistent. PyHealth caches the parsed dataset; without cache_dir, you re-parse every run.

How to use this skill

PyHealth has a large API surface — there's no point loading it all at once. Read the reference file that matches the user's task:

If the user is asking about…	Read
Installing, env setup, MIMIC access, GPU	references/installation.md
Which dataset class to use, loading patterns, splitting	references/datasets.md
What prediction task to choose (mortality, readmission, drug rec, sleep…)	references/tasks.md
Picking a model architecture, model-specific arguments	references/models.md
Looking up or cross-mapping ICD/ATC/NDC/RxNorm/CCS codes, tokenizers	references/medcode.md
End-to-end recipes for common scenarios	references/examples.md

For multi-step tasks (e.g., "build a drug recommendation pipeline on MIMIC-IV"), read tasks.md + models.md + examples.md together — they cross-reference each other.

A note on style

Write minimal, idiomatic PyHealth. The library is opinionated; lean into its abstractions instead of reimplementing them in raw PyTorch. If you find yourself writing a custom training loop, ask whether Trainer would do the job — it almost always will, and it handles checkpointing, logging, and best-model selection for free.

When the user has private MIMIC access, point them at the local CSV root; for demos and learning, the synthetic MIMIC-III bucket (https://storage.googleapis.com/pyhealth/Synthetic_MIMIC-III/) is fine and works without credentialing.