RiskVision
RiskVision predicts stroke and heart disease risk using clinical features. The model was trained on public health datasets and achieves over 80% accuracy on held-out test data.
Problem
Early detection of stroke and heart disease is critical but requires specialist analysis. The goal was to build a tool that triages risk from basic clinical inputs — age, blood pressure, glucose, BMI and a handful of lifestyle flags.
Approach
Trained gradient-boosted and neural network classifiers on UCI and Kaggle health datasets:
- SMOTE to handle severe class imbalance in the positive cases
- SHAP values for per-prediction interpretability
- Stratified k-fold cross-validation to keep the minority class represented
from imblearn.over_sampling import SMOTE
X_res, y_res = SMOTE(random_state=42).fit_resample(X_train, y_train)
model.fit(X_res, y_res)
Interpretability was non-negotiable — a risk score a clinician can't interrogate is a risk score they won't use.
Outcome
Deployed as a FastAPI service with a React frontend. The model consistently outperformed the baseline logistic regression by 14 percentage points, holding above 80% accuracy on held-out test data.