Heart Disease Prediction with Feature Engineering, SMOTE Augmentation, and Interpretable Deep Learning Models

Abstract

Cardiovascular disease is still a leading destination in all of the universe, and that’s why this serious matter demands the development of effective and accurate forecasting methods to address the issue both at early stage and prevent it in the case of concern. This work utilizes the machine learning and deep learning paradigms to manipulate the Cleveland Heart Disease dataset (CHDD) in order to make easy and credible predictions of heart disease. Working with the data included pre-processing such as missing value replacements, outlier detection and removal, standardisation, as such – the authors balanced the classes using SMOTE (Synthetic Minority Over-sampling Technique). Apart from that, various techniques such as Analysis of variance (ANOVA), Chi-square, and Mutual Information were employed for feature selection lead to improved subsets (SF-1, SF-2, SF-3) which were classifiers subjected to tests including Logistic Regression, Support Vector Machines, K-Nearest Neighbours and Voting Ensembles, among others. The outcomes revealed that the unsophisticated together with single classifiers were able to obtain around 90-91% accuracies. On the other hand, the cutting-edge Regularized Deep Feedforward Neural Network (DNN) with Swish activation, AdamW optimizer and SMOTE oversampling boosted the accuracy rate substantially, registering an impressive accuracy figure of 98% with balanced precision, recall and F1-score. SHAP explainer was used to improve the model interpretability and the final model was packaged as a mobile application for medical professionals to use in real time. The understanding is that in contradiction to individual ML or DL models, there is potential of hybrid ML-DL pipelines in implementing applications, which use cardiovascular risk prediction and are reliable, affordable, and can be easily scaled up.