ehr-analysis

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.