HeartRisk Assist - Cardiac Risk Triage (Medi-Hack 2025)

HeartRisk Assist provides clinicians with a calibrated probability of cardiac risk, fairness slice metrics, and point-of-care explanations in a single, intuitive Streamlit app.

Comment

Inspiration

Crowded triage lines, limited cardiology slots, and inconsistent risk estimates. I wanted a tiny, transparent tool that provides clinicians with a calibrated probability (not just a score), explains why the prediction occurred, and highlights fairness gaps - so they can route patients more efficiently and safely.

What it does

  • Predicts a calibrated probability that heart disease is present from 13 routine inputs.
  • Maps probability into Low / Medium / High bands using a documented demo policy (Low < 7%, High ≥ 35%).
  • Explains each decision (SHAP or robust fallbacks) with signed bars.
  • Monitors fairness via simple slice AUCs (sex, age buckets, chest‐pain type).
  • Shows model quality (ROC, PR, calibration curve, AUC 95% CI).
  • Scores uploaded CSVs in batch and lets you push any dataset row directly into Triage.

How we built it

  • Data: Kaggle Heart Disease dataset; I deduplicated to 302 rows to avoid leakage/duplicates.
  • Modeling: ColumnTransformer (scale numeric + one-hot categoricals) → Random Forest (final) and Logistic Regression (baseline) → isotonic calibration with CalibratedClassifierCV on a held-out fold → 20% test evaluation.
  • Explainability: shap.TreeExplainer for RF, shap.LinearExplainer for LR, with coefficient/importance or directional ablation fallbacks.
  • App: Streamlit multipage UI (Triage, Explanations, Fairness, Model Quality, Data Explorer, Batch Scoring).
  • Thresholds: I documented a principled selection method (capacity/PPV for High, miss-rate/NPV for Low) and used 7% / 35% as sensible demo defaults.

Challenges we ran into

  • Tiny dataset → risk of overfitting; required restrained RF tuning and held-out calibration.
  • Explaining calibrated models → I had to unwrap CalibratedClassifierCV to explain the underlying estimator cleanly.
  • UI clarity → making calibration/bands and “per-100 patients” interpretation obvious to non-ML users.
  • Data hygiene → removing duplicates so train/validation/test splits were honest.

Accomplishments that we're proud of

  • End-to-end, reproducible pipeline with artifacts and a clean demo app.
  • Calibrated probabilities that make thresholds meaningful, not arbitrary.
  • Clear, signed local explanations and quick fairness slices that surface operational watchpoints.
  • Lightweight app that runs anywhere without PHI.

What we learned

  • Calibration changes the conversation: 30% means ~30/100, which clinicians understand.
  • Thresholds must be tied to capacity, PPV/NPV, and miss tolerance, not gut feel.
  • SHAP with calibrated models requires care; transparency beats raw accuracy in triage tools.
  • Small UX choices (vertical bars, per-100 wording) dramatically improve comprehension.

What's next for HeartRisk Assist - Cardiac Risk Triage (Medi-Hack 2025)

  • Train on-site EHR cohorts (e.g., MIMIC-IV/PhysioNet) with IRB/governance.
  • Add an auto-threshold finder (capacity/PPV and miss-rate/NPV rules, decision curves) and cohort-specific bands.
  • Prospective A/B tests, drift/bias monitoring, and clinician-in-the-loop feedback.
  • Expand features (labs/meds), improve accessibility, and package for hospital deployment.

Team

Solo: Sweety Seelam

Built With

Share this project:

Updates