Seminar Overview
This graduate seminar explored the application of ensemble decision trees (Random Forests) for patient cardiovascular risk prediction using clinical and demographic indicators. The project investigated the trade-offs between model interpretability, feature importance ranking, and classification robustness in medical diagnostic decision support.
Methodology & Classification Pipeline
- Bagging & Random Subspace: Evaluated bootstrap aggregation combined with random feature subspace sampling to minimize individual tree variance without inducing bias.
- Splitting Criteria: Compared Information Gain (Entropy), Gini Impurity, and Gain Ratio metrics across continuous and categorical clinical predictors (e.g., resting blood pressure, serum cholesterol, ST depression).
- Ensemble Aggregation: Analyzed majority voting and soft probability thresholding across 100+ decision tree estimators.
- Hyperparameter Optimization: Explored the impact of tree depth constraints, minimum sample split boundaries, and maximum feature proportions on out-of-bag error rates.
Experimental Insights
The ensemble approach demonstrated strong resilience against local noise in clinical measurements. Feature importance analyses identified key predictive physiological markers while highlighting how imbalanced diagnostic datasets require class-weighted loss functions rather than raw accuracy maximization.
References
- Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5-32.
- UCI Machine Learning Repository & Kaggle Heart Disease Dataset.