A BIO-INSPIRED DEEP HYBRID FRAMEWORK FOR CARDIOVASCULAR DISEASE PREDICTION: SYNERGIZING MODIFIED RANDOM FOREST FEATURE SELECTION WITH DENSENET-GNM REPRESENTATION

Abstract

Cardiovascular diseases (CVDs) remain the leading cause of global mortality, necessitating diagnostic models that balance high accuracy with computational efficiency. While Deep Learning (DL) models like DenseNet and Gated Network Models (GNM) offer superior feature representation, they often suffer from noise sensitivity and suboptimal hyperparameter tuning via static Grid Search methods. Conversely, traditional Machine Learning approaches benefit from robust feature selection but lack the capacity for deep latent representation. This paper proposes a novel Optimized Dual-Stream Ensemble Framework. First, we employ a Modified Random Forest (MRF) algorithm using Cohen’s kappa coefficient to eliminate redundant clinical features. Second, the selected optimal features are projected into a high-dimensional latent space using a DenseNet-based Feature Augmentation module. Finally, classification is performed via a hybrid ensemble of a GNM and a Deep Multi-Layer Perceptron (DMLP). Crucially, the hyperparameters and ensemble weights are dynamically optimized using a Sobel Sequence Brownian Random Walk-based Dragonfly Algorithm (SSBRWDOA) to avoid local optima. Experimental results on the Cleveland dataset demonstrate that this synergistic approach achieves 99.12% accuracy, outperforming state-of-the-art methods.