A Hybrid GA-KNN Framework For Cardiovascular Disease Prediction Using Optimized Clinical Feature Selection

Abstract

Background Study: Background Study: Cardiovascular disease (CVD) causes millions of fatalities each year and places a heavy financial strain on healthcare systems. Better patient outcomes, prompt clinical intervention, and lower healthcare costs all depend on early and precise cardiovascular disease prediction. Through the analysis of massive amounts of clinical data, machine learning algorithms have considerable potential in helping doctors identify diseases. Problem Statement: High-dimensional clinical datasets, repetitive and irrelevant features, and the difficulty to consistently identify the most discriminative risk factors are common problems for current machine learning-based techniques for cardiovascular disease prediction. These problems limit the robustness and generalizability of prediction models, raise computing costs, and decrease classification accuracy. This is particularly true for distance-based classifiers, such as K-Nearest Neighbor (KNN). Developing an efficient approach that combines accurate classification with suitable clinical feature selection remains a critical research problem for improving early cardiovascular disease prediction and enabling reliable clinical decision-making. Purpose: In a medical decision support system, the prediction of cardiovascular disease is an important task, as early detection can help in minimizing the risk of mortality, delay in treatment, and cost of healthcare. Methods: In this study, an optimized hybrid approach of the genetic algorithm and K-nearest neighbor method for cardiovascular disease prediction is proposed. The clinical attributes are selected using the genetic algorithm, and the final classifier is KNN. Four datasets, the Cleveland Processed Heart Dataset, the CRPF Ranchi Clinical Heart Dataset, the Cleveland Hungarian Statlog Dataset, and the Heart Failure Clinical Record Dataset, were used for evaluating the model. Initial experiments were conducted with k-fold values of 5, 10, 15, 20, and 25 folds, and then an optimized 10-fold GA-KNN approach with feature selection, normalization, binary target conversion, and hyperparameter tuning of KNN was executed. Results: The optimized model achieved accuracies of 78.19%, 75.71%, 92.10%, and 81.98%, respectively, with ROC-AUC values of 0.8612, 0.7603, 0.9665, and 0.8313. Conclusion: It is demonstrated that optimized GA-KNN can achieve both  feature dimensions for the initial screening of cardiovascular diseases. The proposed GA-KNN framework is simple, interpretable, and computationally efficient for preliminary cardiovascular disease screening.