Ehr Analysis Agent Skill

EHR Analysis

Version Compatibility

Reference examples assume:

• pyhealth 1.1.6+ (stable) or 2.0+ (latest, requires Python ≥ 3.12)
• torch 2.0+
• pandas 2.0+

Verify the runtime first:

• Python: python -c "import pyhealth; print(pyhealth.__version__)"
• CLI: pip show pyhealth
• If APIs differ between 1.x and 2.x, consult the installed docs and adapt the pattern accordingly.

Overview

Use this skill when the user needs to:

• load and structure EHR data from MIMIC-III, MIMIC-IV, eICU, OMOP-CDM, or a custom dataset
• define clinical prediction tasks such as mortality, readmission, length of stay, or drug recommendation
• train and evaluate deep learning or ML models on patient visit sequences
• apply medical code normalization, ontology mapping, or tokenization
• calibrate model outputs or generate post-hoc interpretability artifacts

When To Use This Skill

• the data source is a structured EHR database or tabular patient record export
• the task requires longitudinal patient visit sequences as model input
• the user needs benchmark-compatible evaluation metrics for clinical AI
• medical code systems (ICD, ATC, NDC, CPT, CCS) must be normalized or mapped before modeling

Quick Route

• Known EHR source (MIMIC, eICU, OMOP): load with a pyhealth.datasets class, then set_task.
• Custom tabular data: wrap in SampleBaseDataset, then follow the same pipeline.
• Code mapping only: use pyhealth.medcode independently without the full pipeline.
• Evaluation only: use pyhealth.metrics independently on existing predictions.

Progressive Disclosure

• Read references/technical_reference.md for dataset branching rules, task schema details, model selection guidance, and calibration caveats.
• Read references/commands_and_thresholds.md for concrete PyHealth code patterns, recommended thresholds, and output file conventions.
• Keep SKILL.md as the main execution path and load reference files only when the task or failure mode needs the extra detail.

Default Rules

• Always use raw integer counts or visit-level code lists as model inputs; do not pre-aggregate before loading.
• Keep dataset loading, task definition, model training, and evaluation as separate, inspectable stages.
• Record dataset name, table selections, code mapping applied, task function, model class, and PyHealth version in every experiment.
• Favor reproducible train/val/test splits via split_by_patient over random splits that can leak patient data.

Expected Inputs

• EHR source files or database dumps (MIMIC-III, MIMIC-IV, eICU, OMOP-CDM, or custom CSV/Parquet)
• table names to load (e.g., diagnoses_icd, procedures_icd, prescriptions)
• optional medical code mapping specifications
• task definition (predefined or custom)
• model architecture choice and training hyperparameters

Expected Outputs

• results/model_checkpoint.pth
• results/test_metrics.json
• results/predictions.tsv
• qc/dataset_stats.tsv
• figures/calibration_curve.pdf
• figures/feature_importance.pdf

Preferred Tools

• pyhealth.datasets — MIMIC3Dataset, MIMIC4Dataset, eICUDataset, OMOPDataset, SampleBaseDataset
• pyhealth.tasks — MortalityPrediction, ReadmissionPrediction, DrugRecommendation, LengthOfStayPrediction, and custom task functions
• pyhealth.models — RNN, LSTM, GRU, Transformer, RETAIN, SafeDrug, GAMENet, AdaCare, ConCare, StageNet
• pyhealth.trainer — Trainer
• pyhealth.metrics — binary, multiclass, multilabel, fairness, interpretability metrics
• pyhealth.medcode — InnerMap, CrossMap for ontology lookup and code mapping
• pyhealth.tokenizer — Tokenizer for code-to-index transformations
• pyhealth.calib — calibration and prediction set construction

Starter Pattern

from pyhealth.datasets import MIMIC3Dataset
from pyhealth.tasks import drug_recommendation_mimic3_fn
from pyhealth.datasets import split_by_patient, get_dataloader
from pyhealth.models import Transformer
from pyhealth.trainer import Trainer

# Step 1: load dataset
dataset = MIMIC3Dataset(
    root="/path/to/mimic3/",
    tables=["DIAGNOSES_ICD", "PROCEDURES_ICD", "PRESCRIPTIONS"],
    code_mapping={"NDC": ("ATC", {"target_kwargs": {"level": 3}})},
)

# Step 2: define task
samples = dataset.set_task(task_fn=drug_recommendation_mimic3_fn)

# Step 3: split and build dataloaders
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)

# Step 4: initialize model
model = Transformer(
    dataset=samples,
    feature_keys=["conditions", "procedures"],
    label_key="drugs",
    mode="multilabel",
)

# Step 5: train
trainer = Trainer(model=model)
trainer.train(
    train_dataloader=train_loader,
    val_dataloader=val_loader,
    epochs=50,
    monitor="pr_auc_samples",
)

# Step 6: evaluate
metrics = trainer.evaluate(test_loader)
print(metrics)

Workflow

1. Validate dataset access and structure

Confirm the EHR source is accessible and the required tables are present. Inspect patient counts, visit counts, and code distributions before building tasks.

2. Load data with pyhealth.datasets

Use the appropriate dataset class. Apply code mappings at load time so downstream tasks and models operate on a consistent vocabulary. Record which tables and mappings were used.

3. Define the clinical task

Choose a predefined task function or write a custom one. Confirm the input schema (feature keys and types) and output schema (label key and mode: binary, multiclass, or multilabel) match the clinical question.

4. Split patients and build dataloaders

Always split by patient, not by visit or row, to prevent patient-level data leakage. Use split_by_patient and get_dataloader from pyhealth.datasets.

5. Initialize and train the model

Select an architecture matched to the task and data size. Configure feature_keys, label_key, and mode explicitly. Use Trainer for training, validation monitoring, and checkpoint saving.

6. Evaluate and report

Compute task-appropriate metrics using pyhealth.metrics. For binary tasks report AUROC and AUPRC. For multilabel tasks report Jaccard and PR-AUC. Export a metrics JSON and a predictions TSV alongside the model checkpoint.

7. Calibrate and interpret where needed

Apply calibration from pyhealth.calib when predicted probabilities will be used for clinical decision support. Use pyhealth.interpret for feature-level or visit-level attribution when the use case requires explainability.

Output Artifacts

results/
├── model_checkpoint.pth
├── test_metrics.json
└── predictions.tsv
qc/
├── dataset_stats.tsv
└── split_summary.tsv
figures/
├── calibration_curve.pdf
└── feature_importance.pdf

Key Parameters

| Parameter | Typical value | Notes |

|---|---|---|

| batch_size | 32 | reduce for large visit sequences or memory-constrained environments |

| epochs | 50 | use early stopping via monitor metric |

| monitor | "pr_auc_samples" for multilabel; "roc_auc" for binary | task-dependent |

| split_by_patient ratios | [0.8, 0.1, 0.1] | always split by patient not by visit |

| code mapping level | ATC level 3 for drug codes | balance specificity and sparsity |

| n_top_genes (HVG analog) | N/A for EHR; use vocabulary filtering | filter rare codes with min_count |

Quality Review

• Confirm patient count, visit count, and code vocabulary size before training.
• Verify that train, validation, and test sets have no overlapping patient IDs.
• Check class balance for binary and multiclass tasks before reporting metrics.
• Review calibration curves if predicted probabilities feed downstream clinical logic.
• Do not report only AUROC; include AUPRC especially for imbalanced clinical outcomes.
• Record dataset version, table selection, code mapping, task function, and model class for every experiment.

Anti-Patterns

• splitting by visit or row instead of by patient, causing data leakage
• using AUROC as the sole metric for rare clinical events
• applying code mappings inconsistently across train and test sets
• mixing raw codes and mapped codes in the same feature without documenting it
• reporting model results without checking data coverage for the target cohort
• treating predicted probabilities as calibrated without verification

Related Skills

• Machine Learning For Omics
• Multi-Omics Integration
• Pathway Analysis
• Reporting And Figure Export

Optional Supplements

• pyhealth.medcode for standalone code mapping
• pyhealth.calib for post-hoc calibration
• pyhealth.interpret for interpretability