how_to_get_started.rst

How to Get Started with PyHealth

Welcome to PyHealth! This guide will help you build machine learning pipelines for healthcare tasks using PyHealth.

Installing PyHealth

You can install PyHealth using pip:

pip install pyhealth

To install the latest development version from GitHub:

git clone https://github.com/sunlabuiuc/PyHealth.git
cd PyHealth
pip install -e .

Ensure you have the required dependencies installed before proceeding.

Overview of ML Pipelines

All healthcare tasks in PyHealth follow a five-stage pipeline:

Each stage is modular, allowing customization based on your needs.

Stage 1: Loading Datasets

pyhealth.datasets provides structured datasets independent of tasks. PyHealth supports MIMIC-III, MIMIC-IV, eICU, and more.

Example:

from pyhealth.datasets import MIMIC3Dataset

mimic3base = MIMIC3Dataset(
    root="https://storage.googleapis.com/pyhealth/Synthetic_MIMIC-III/",
    tables=["DIAGNOSES_ICD", "PROCEDURES_ICD", "PRESCRIPTIONS"],
    code_mapping={"NDC": "CCSCM"},
)

Stage 2: Defining Tasks

pyhealth.tasks processes patient data into task-specific samples.

Example:

from pyhealth.tasks import readmission_prediction_mimic3_fn

mimic3sample = mimic3base.set_task(task_fn=readmission_prediction_mimic3_fn)

To split data and create DataLoaders:

from pyhealth.datasets import split_by_patient, get_dataloader

train_ds, val_ds, test_ds = split_by_patient(mimic3sample, [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)

Stage 3: Building ML Models

pyhealth.models provides various machine learning models.

Example:

from pyhealth.models import Transformer

model = Transformer(
    dataset=mimic3sample,
    feature_keys=["conditions", "procedures", "drugs"],
    label_key="label",
    mode="binary",
)

Stage 4: Training the Model

pyhealth.trainer allows specifying training parameters such as optimizer, epochs, and learning rate.

Example:

from pyhealth.trainer import Trainer

trainer = Trainer(model=model)
trainer.train(
    train_dataloader=train_loader,
    val_dataloader=val_loader,
    epochs=50,
    monitor="pr_auc_samples",
)

Stage 5: Evaluating Model Performance

pyhealth.metrics provides evaluation metrics.

Example:

trainer.evaluate(test_loader)

from pyhealth.metrics.binary import binary_metrics_fn

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