A PSO-Based Hyperparameter Optimized Deep Neural Network for Accurate Heart Disease Classification

Abstract

Cardiovascular diseases rank among the foremost causes of death worldwide, underscoring the necessity for precise predictions through effective predictive models. This study introduces an Optimized Deep Neural Network (ODNN) aimed at forecasting heart disease by utilizing electronic health record (EHR) databases sourced from the UCI machine learning repository. The methodology employed in this research is delineated as follows: initially, the research process encompasses data preprocessing, feature extraction, feature optimization, and ultimately, heart disease classification via deep learning algorithms. To begin with, missing data were addressed using K-Nearest Neighbor (KNN), while dimensionality reduction was achieved through Singular Value Decomposition (SVD) and Z-score normalization techniques. Subsequently, Principal Component Analysis (PCA) was applied to extract significant features from the dataset, which were then optimized using Particle Swarm Optimization (PSO). Furthermore, to mitigate overfitting concerns, the oversampling technique was adapted to one that is more suitable for educational contexts. Ultimately, the PRelu activation function and the move entropy loss function were employed to detect coronary heart diseases, facilitated by the optimized function. The experiments utilized the Cleveland and ORDBA datasets. The findings reveal that the proposed model attained classification accuracies of 99.14% and 98.45% for these datasets. These outcomes suggest that the proposed framework is capable of effectively and reliably predicting heart disease, thereby serving as a valuable tool for diagnosing the early stages of heart conditions through medical decision support systems.