Heart Disease Prediction Using Machine Learning

Cardiovascular disease is one of the leading causes of death worldwide, and limited doctor-to-patient ratios in developing regions spur low-cost machine-learning-based screening tools. This article does a review and case study of supervised learning approaches to heart disease prediction using UCI/Cleveland dataset. The five previously reported technique families in the literature are K-Nearest Neighbour, Decision Tree, Naive Bayes, Support Vector Machine and Logistic Regression and untuned Random Forest. These families are reviewed and benchmarked against a proposed Random Forest that is rigorously tuned. The pipeline that is proposed extends the initial 13 clinical parameters to 21 clinical parameters by the addition of 8 domain-informed engineered features. A custom RandomizedSearchCV with 5-fold stratified cross-validation (30 combinations, 150 fits) is applied. The system is deployed as a real-time Flask web application. An ablation study analyzes the contribution of feature engineering and hyperparameter tuning to the final result. The tuned model is able to achieve a test accuracy of 90.24% and a ROC-AUC of 92.64%. The accuracy achieved is higher than or equal to that of the best performing models reported during the literature review, while at the same time providing 5-fold cross-validated performance (91.59% ± 1.94%), a full complement of multi-metrics evaluation, Gini-based feature-importance ranking, and end-to-end deployment interface all of which were largely absent from the reviewed literature.