Developed as a Bioengineering + Machine Learning project focused on clinically reliable prediction systems.
A machine learning-based clinical decision support system for predicting cardiovascular disease risk using patient clinical data.
Designed as a clinical decision support prototype for cardiovascular risk prediction, prioritizing high recall to minimize missed diagnoses and providing calibrated probability estimates for reliable decision-making.
👉 https://clinical-risk-predictor.streamlit.app
Early detection of heart disease is critical. In clinical settings, missing a positive case (false negative) can lead to severe consequences.
This project focuses on:
- Maximizing recall (reducing missed diagnoses)
- Providing reliable probability estimates for decision-making.
Dataset: UCI Heart Disease dataset (clinical tabular data).
-
Model: Calibrated Logistic Regression
-
Threshold tuning: 0.3 (recall-focused decision boundary)
-
Calibration: Ensures predicted probabilities reflect real-world likelihood
-
Evaluation:
- ROC-AUC
- Precision-Recall Curve
- Confusion Matrix
- Cross-validation
Although Random Forest achieved near-perfect performance, it showed signs of overfitting due to dataset size.
Logistic Regression was chosen because:
- Better generalization
- Interpretability (important in healthcare)
- Supports calibrated probabilities
- Improves recall
- Reduces false negatives
- Aligns with clinical priority of early detection
Used CalibratedClassifierCV to ensure:
- Probability outputs are trustworthy
- Model can support risk-based decision making
- High recall (0.91) for disease class, prioritized to reduce missed diagnoses (false negatives).
- Well-calibrated probabilities suitable for clinical decision-making
- Model Performance Comparison
Logistic Regression:
- ROC-AUC: 0.93
- Recall (Disease Class): 0.91
- Precision (Disease Class): 0.76
- F1-score: 0.83
Random Forest:
- ROC-AUC: 0.999
- Recall (Disease Class): 1.00
- Precision (Disease Class): 0.98
- F1-score: 0.99
Although Random Forest achieves near-perfect performance, this is likely due to overfitting given the limited dataset size.
Logistic Regression was selected as the final model due to better generalization, interpretability, and more reliable calibrated probabilities, making it more suitable for clinical use.
Threshold tuning (0.3) was applied to improve recall (0.91) and align predictions with clinical priorities.
Key contributing features include:
- Chest pain type
- Maximum heart rate achieved
- ST depression (oldpeak)
- Number of major vessels
These align with known clinical indicators of cardiovascular disease risk.
-
Risk stratification:
- Low Risk (< 0.3)
- Medium Risk (0.3 – 0.6)
- High Risk (> 0.6)
-
Human-readable clinical reasoning (heuristic-based explanations)
-
Recommendation system based on risk level
-
Deployed interactive Streamlit web application (public access)
- Dataset is relatively small
- Categorical variables treated as ordinal; proper encoding (e.g., OneHotEncoder) can improve performance.
- No external clinical validation
- SHAP-based explainability
- Proper categorical encoding (OneHotEncoder)
- Validation on real-world clinical datasets
- Python
- scikit-learn
- pandas, numpy
- matplotlib, seaborn
- Streamlit (for deployment)
- joblib (model serialization)
- app.py → Streamlit web app
- notebook.ipynb → model development & analysis
- model.pkl → trained model
- columns.pkl → feature order used for prediction
- requirements.txt → dependencies
- Clone the repository:
git clone https://github.com/TarunaJ2006/Clinical-Risk-Prediction-System.git
cd Clinical-Risk-Prediction-System- Install dependencies:
pip install -r requirements.txt- Run the app:
streamlit run app.pyThis project is for educational purposes and is not a substitute for medical diagnosis.