Ensemble to Deep Learning Framework Transition for Enhanced Heart Disease Prediction

Abstract

Cardiovascular diseases remain a leading cause of mortality worldwide, which underscores the need for accurate and interpretable risk stratification tools. We propose a hybrid framework that merges gradient-boosted decision trees with a deep neural network to better heart disease classification and systolic blood pressure regression. The methodology first employs an ensemble stacking architecture comprising Random Forest, XGBoost, and LightGBM models to extract robust feature representations and reduce noise from multidimensional patient data. This ensemble phase yields a transformed feature vector capturing non-linear interactions and complex relationships. This intermediate representation then feeds directly into a three-layer Multi-Layer Perceptron, which learns higher-order correlations without being exposed to raw data noise. The progression from ensemble outputs to neural network inputs directs the stochastic optimization process, thereby improving convergence stability and predictive accuracy. In binary classification, the final features are transformed into a probability score via a sigmoid activation function; for regression, a linear output layer optimized by mean squared error is adopted. The principal innovation resides in the layered association of tree-structured models with deep learning, wherein deterministic patterns from the ensemble guide the neural network’s learning process. This design achieves superior accuracy without sacrificing computational efficiency. Furthermore, the framework possesses inherent explainability by means of the ensemble’s feature importance metrics. Our experiments indicate that this hybrid approach outperforms standalone models, thereby presenting a pragmatic solution for resource-constrained clinical settings.