Cardiovascular Disease Risk Prediction: Full Project Report

📋 1. Executive Summary

The primary objective of this project was to develop a machine learning model capable of identifying patients at high risk for cardiovascular disease. Through a systematic evaluation of linear (SVM) and non-linear (Decision Tree) algorithms, the analysis revealed a significant challenge with overfitting.

While initial unconstrained models achieved near-perfect training scores, they failed to generalize to new data. By implementing Hyperparameter Tuning via Grid Search, the model was stabilized, moving from "memorizing" the training set to identifying broader clinical patterns. This project serves as a detailed case study in balancing model complexity with real-world reliability.

---

🛠️ 2. Project Workflow & Methodologies

2.1 Data Preprocessing & Cleaning

Raw clinical data was transformed to ensure biological and statistical validity:

• Feature Categorization: Features were grouped into Objective (Age, Height), Examination (Blood Pressure, Cholesterol), and Subjective (Smoking, Alcohol) categories.
• Feature Engineering: Age was converted from days to years, and BMI was calculated to provide standard clinical context.
• Outlier Management: Instances where blood pressure readings were physically impossible (e.g., diastolic higher than systolic) were removed to ensure data integrity.

2.2 Baseline Modeling: Linear SVM

We first established a performance baseline using a Linear SVM:

• Outcome: The model achieved a 0.7232 training accuracy but suffered from a "no-skill" result on the test set (AUC 0.49).
• Finding: A linear boundary was insufficient to capture the complex relationships between these health indicators.

2.3 Advanced Modeling: Decision Tree

To capture non-linear interactions, we implemented a Decision Tree:

• The Overfitting Issue: Without constraints, the tree hit 99.4% training accuracy.
• Generalization Failure: Testing accuracy dropped to 48.4%, indicating the model was learning noise rather than transferable medical patterns.

2.4 Optimization via Grid Search

We used Grid Search with 5-fold cross-validation to prune the tree and find a stable middle ground:

• Optimized Parameters: max_depth: 5 and min_samples_leaf: 20.
• Stabilization: Training accuracy was brought down to a more honest 72.3%, narrowing the gap between training and testing performance.

---

📈 3. Final Performance Comparison

Metric	Linear SVM	Unconstrained DT	Optimized Decision Tree
Training Accuracy	0.7232	0.9943	0.7232
Testing Accuracy	0.4947	0.4844	0.4947
Testing Recall	1.0000	0.7446	1.0000

---

🧪 4. Conclusion & Clinical Insights

1. Generalization over Accuracy: High training scores (99%) are often a warning sign of overfitting. Real clinical utility requires depth constraints to handle noisy medical data.
2. Prioritizing Sensitivity: The final model achieved a 1.00 Recall on the test set, ensuring that no high-risk patients are missed—a critical priority in healthcare settings.
3. Future Recommendations: Moving toward Ensemble Methods like Random Forest or Gradient Boosting is recommended to further improve precision while maintaining the stability achieved through Grid Search

---

🛠️ How to Run

1. Clone this repository.
2. Install dependencies: pip install -r requirements.txt
3. Open Cardiovascular_Risk_Analysis.ipynb in Jupyter Notebook.