CVD Digital Twin

a CVD that predicts an individual's risk, explains the camyes of that risk mying SHAP, and allows for simulation of "what if" lifestyle and clinical interventions.

Comment

🫀 CVD Digital Twin — Interpretable Cardiovascular Risk with “What‑If” Simulation

Inspiration

Most risk calculators give a single opaque number. Clinicians and patients want more:
1) What’s my current risk?
2) Why is it high or low (for me specifically)?
3) What would change my risk if I lower BP, quit smoking, or get more active?

That’s the gap this project addresses with an interpretable, interactive CVD Digital Twin.

What it does

  • Predicts CVD risk from routine clinical & lifestyle features.
  • Explains every prediction with SHAP (global & per‑patient drivers).
  • Simulates “what‑if” scenarios (e.g., lower systolic BP to 120/80, stop smoking, increase activity) and quantifies the risk delta in real time.
  • Ships with a Streamlit demo so judges/users can try scenarios live.

Research prototype — not a medical device.

How we built it

Data (Modal A): cardio_train.csv (or cardio_base.csv)
We derive: ( \text{Age (years)}, \text{BMI} = \frac{\text{weight}}{(\text{height}/100)^2} ), pulse pressure, and flags for hypertension (ap_hi ≥ 140 OR ap_lo ≥ 90) and obesity (BMI ≥ 30). Physiological sanity filters remove implausible BP/BMI entries.

Modeling:

  • XGBoost classifier with a small GridSearchCV over depth/trees/lr using 5‑fold Stratified CV.
  • Probability calibration (isotonic) for trustworthy risk probabilities (calibration curve + Brier score).
  • Metrics: ROC‑AUC, PR‑AUC, F1, confusion matrix, Brier; plots exported for the report & gallery.

Explainability & Simulation:

  • SHAP beeswarm + bar (global) and waterfall (local) for a chosen patient.
  • What‑if function perturbs features (e.g., systolic/diastolic BP, smoking, activity), recomputes calibrated risk, and reports absolute/relative change.

Demo:

  • app.py (Streamlit) loads the trained, calibrated model + scaler and runs the “what‑if” scenarios interactively.

Challenges

  • Probability quality: getting risk calibrated (not just high AUC) so “what‑if” deltas are meaningful.
  • Data cleanliness: ensuring physiologically plausible vitals (BP relation, BMI bounds) to avoid corrupting the learner.
  • Explaining to non‑experts: choosing SHAP visualizations and copy that are useful to clinicians and accessible to students/newcomers.

Accomplishments we’re proud of

  • A calibrated risk model with clear, clinician‑friendly explanations.
  • Working demo with one‑click “Lower BP to 120/80” and “Quit Smoking & Increase Activity” scenarios.
  • End‑to‑end reproducibility (seeded CV, saved artifacts, single‑run notebook).

What we learned

  • Calibration matters: Brier score + calibration curves change the conversation from “model is accurate” to “probabilities you can trust.”
  • SHAP can be both global (feature ranking) and local (patient‑level rationale) without being overwhelming if plotted well.

What’s next

  • Add optional ECG‑derived features (Modal B) for a multimodal ensemble.
  • Add fairness slices by age/sex to study performance parity.
  • Deploy on Streamlit Community Cloud for a stable, always‑on demo link.

How we evaluated

  • ROC‑AUC & PR‑AUC for discrimination; F1 for balance; Brier score + calibration curve for probability quality.
  • SHAP confirms domain‑plausible drivers: systolic BP, age, BMI, cholesterol.
  • A patient case study shows risk reduction when systolic BP is lowered toward guideline targets.

How to run (summary)

  1. Open the notebook, upload cardio_train.csv, run all cells to train + save artifacts.
  2. Launch the Streamlit app: streamlit run app.py
  3. Try the built‑in scenarios in the UI (“Lower BP to 120/80”, “Quit Smoking & Increase Activity”).

Built With

Share this project:

Updates

Submission history