EHR Risk Supported
Logistic regression, random forest, XGBoost, optional LightGBM. Windowed tabular features. Optional isotonic calibration. SHAP explanations.
Limitations & model card
What this framework does not claim. Use alongside the capability pages when writing theses or lab reports.
Clinical / legal / safety
- Not for diagnosis, treatment, or triage without institutional governance
- No FDA/CE claim — outputs are research probabilities only
- You own PHI, IRB, DUA, and PhysioNet credentialing
- Local API is unsafe on public networks without
API_KEY
Modeling boundaries
Not supported
LSTM / Transformers / foundation models as a production path. Full FHIR/OMOP servers. Multi-tenant SaaS. Full AutoML HPO.
Data & science honesty
- Demo CSVs are for smoke tests — not clinical performance claims
- MIMIC-IV requires your PhysioNet credentials; restricted data is never shipped
- Leakage audit catches common mistakes — not every contamination pattern
- SHAP explains the fitted model — not causal effects
- Fairness helpers need group columns; otherwise they skip with a reason
- Calibration improves probability quality; it does not guarantee clinical utility
Platform
- Single-host research workbench — one heavy job at a time
- Filesystem experiment runs — not a full MLflow/W&B replacement
- Upload size default ~50MB; SQL import is read-only
SELECT
Model card
Repo companion
Full model card
Intended use, training data notes, metrics, and ethical considerations live in the repository.
How this pairs with the strengths
Documenting limits keeps the capability claims accurate: temporal integrity, calibration, SHAP in the workbench, and Docker researcher loop.