Grey wolf optimized neural network with hybrid feature extraction for heart disease prediction

With an increasing mortality rate due to heart diseases, there is a critical need for early and reliable cardiovascular disease prediction. However, when it comes to health records data, traditional approaches have primarily utilized a single prediction model. This study proposes an advanced approach that leverages Support Vector Machine (SVM) in independent parallel streams with Convolutional Neural Networks (CNN) for extracting features, combined with an ANN enhanced via Grey Wolf Optimization (GWO) for predictive classification. First, the input data is passed into the SVM and CNN simultaneously to extract relevant and important features by each method. Both extractors receive identically preprocessed inputs simultaneously; their outputs are concatenated post-extraction. The SVM extracts probability-based discriminative features reflecting class separation confidence, while CNN extracts hierarchical convolutional features capturing local patterns and non-linear attribute interactions. Then, the attributes extracted by SVM and the attributes extracted by CNN are concatenated to get a comprehensive set of features. The concatenation enriches the feature space with diverse representational characteristics: SVM contributions provide regularized decision confidence scores, while CNN contributions offer multi-scale pattern descriptors. This heterogeneous feature integration enables the ANN to learn more robust classification boundaries by accessing both discriminative and descriptive feature domains simultaneously. Second, the concatenated features are sent to the ANN for prediction. The GWO algorithm is used to configure ANN parameters like channel size, drop-out rate, and learning rate. The fitness function maximizes stratified 5-fold cross-validation accuracy. The experiments were conducted on the Heart Disease dataset collection, collected from the UCI Machine Learning Repository (Cleveland, Hungarian, Switzerland, and Long Beach), which includes 14 variables for heart disease prediction. All results represent mean ± standard deviation from nested cross-validation. The results demonstrate that the proposed SVM-CNN+ANN-GWO model achieves a high classification accuracy of 91.80% ± 1.2% with the Cleveland, 91.53% ± 1.5% with Hungarian, 96.08% ± 0.8% with Switzerland, and 87.69% ± 2.1% with VA Long Beach databases, respectively. Statistical significance was confirmed via paired Wilcoxon tests ([Formula: see text]). It outperforms existing baseline approaches and other recent approaches from previous studies based on predictive accuracy. It achieves promising performance in heart disease prediction across all four databases and hence generalizes well on different datasets.