Heart Failure Survival Prediction
Clinical ML system predicting patient survival using GMM-based synthetic augmentation and ensemble classifiers.
The Problem
Small Clinical Dataset
Heart failure has high mortality, but clinical datasets are small. The challenge was building an accurate survival prediction model despite limited training data (299 patients).
The Approach
Synthetic Augmentation + Ensemble
Tested 21 model configurations across Logistic Regression, KNN, Random Forest, Gradient Boosting, and Neural Networks. Addressed small-sample limitations with GMM-based synthetic data augmentation (5,299 generated samples) and threshold tuning to minimize missed high-risk cases.
Technologies & Methods
The Results
85% Prediction Accuracy
Achieved 85% test accuracy with GMM-augmented Gradient Boosting and cut test loss by 42% (0.99 to 0.27). Tuned the decision threshold (0.25–0.30) to prioritize recall, reaching up to 75% recall and reducing missed high-risk patients.