Comparison of Hyperparameter Optimization Techniques in Hybrid CNN-LSTM Model for Heart Disease Classification

Abstract

Heart disease, which causes the highest number of deaths worldwide, recorded about 17.9 million cases in 2019, or about 32% of total global deaths, according to the World Health Organization (WHO). The significance of early detection of heart disease drives research to develop effective diagnosis systems utilizing machine learning. The advancement of machine learning in healthcare currently primarily serves as a supporting role in the ability of clinicians or analysts to fulfill their roles, identify healthcare trends, and develop disease prediction models. Meanwhile, deep learning has experienced rapid development and has become the most popular method in recent years, one of which is detecting diseases. The main objective of this research is to optimize the hybrid convolutional neural network (CNN) and long short-term memory (LSTM) model for classifying heart disease by comparing hyperparameter optimization using grid search and random search. Although random search requires less time in hyperparameter tuning, the classification performance results of grid search show higher accuracy. In the test, the hybrid CNN and LSTM model with grid search achieved 91.67% accuracy, 89.66% recall (sensitivity), 93.55% specificity, 92.86% precision, 91.23% f1-score, and 0.9310 AUC value. These results confirm that using a hybrid CNN and LSTM model with a grid search approach is better suited for classifying heart disease.