Skip to content

Latest commit

 

History

History
64 lines (47 loc) · 2.56 KB

File metadata and controls

64 lines (47 loc) · 2.56 KB

Limitations

This framework is an open research / clinical decision-support prototyping toolkit. It is not a regulated medical device.

Questions or feedback: support@larucare.com · docs/HOW_IT_HELPS.md

A. Clinical / legal / safety

  • Not for diagnosis, treatment, or triage without institutional governance
  • No FDA/CE claim; outputs are research probabilities only
  • User is responsible for PHI, IRB, DUA, and PhysioNet rules
  • Local API is unsafe on public networks without API_KEY
  • Clinical-research mode is a prototype, not production CDS

B. Data coverage

  • Canonical path: tabular/longitudinal events → window aggregates
  • Not full free-text NLP, imaging, waveforms, or genomics
  • FHIR/OMOP adapters are subset importers, not full CDM/FHIR servers
  • MIMIC requires user credentials; restricted data is never shipped
  • Schema mapping cannot fix wrong labels or upstream leakage
  • Default upload size limit ~50MB
  • SQL import is read-only SELECT / WITH … SELECT

C. Modeling / ML

  • Supported: logistic regression, random forest, XGBoost, optional LightGBM
  • LSTM / Transformers / foundation models: not supported
  • Multi-model compare ranks by hold-out ROC-AUC — not full AutoML/HPO
  • Tiny cohorts → unstable or undefined AUC (framework warns)
  • Calibration improves probability quality; does not guarantee clinical utility
  • SHAP explains the fitted model — not causal effects

D. Evaluation / science

  • Default splits are research hold-outs; multi-site validation is user-provided
  • Fairness requires group columns; otherwise skipped with an explicit reason
  • Leakage audit catches common classes — not every contamination pattern
  • Synthetic results are for method/CI verification — not clinical performance claims

E. Platform / ops

  • Single-host research workbench; not multi-tenant SaaS
  • One heavy job at a time (not a cluster scheduler)
  • Filesystem experiment runs — not a full MLflow/W&B replacement
  • No Model Hub / community marketplace in this release
  • Windows/macOS ease depends on Docker Desktop
  • User backs up reports/runs/ and config

F. Customization boundaries

  • Configurable: tasks, windows, horizon, splits, models, calibrate, persona
  • Formal plugin marketplace not included
  • Advanced feature engineering beyond YAML may require Python changes

G. Forbidden implications

We do not claim to:

  • Solve all healthcare AI
  • Replace hospital EHR systems
  • Be state-of-the-art on every MIMIC benchmark
  • Be “bias-free” or to have “solved fairness”