Reinforcement Learning-Guided Optimization of ECG Features for Enhanced Heart Disease Prediction using Stacked CNNs

Abstract

Accurate and timely heart disease diagnosis through intelligent ECG signal processing is essential to reducing death rates and improving patient quality of life. However, efficient feature extraction and development of sound classification models are an enormous challenge in this regard. To overcome these challenges, in this study, a novel method based on a combination of deep learning and reinforcement learning techniques is proposed. The proposed method performs the diagnosis process in four key steps: In the first step, the representation and extraction of ECG signal features are performed using three separate feature sets including Short-time Fourier transform (STFT), Gramian Angular Fields (GAF), and statistical feature set. In the second step, the STFT and GAF features are processed by Convolutional Neural Network (CNN) models to locally diagnose the condition by each of these models. The third phase includes reducing the statistical features using a new algorithm based on reinforcement learning approach to identify a set of features that are more relevant to the presence of the condition. Finally, in the fourth step, this selected set of features is merged with the diagnoses made by CNN models to perform the final diagnosis using a meta-learner based on the Multilayer Perceptron (MLP) structure. The proposed method, after careful evaluation on the MIT-BIH dataset with four target classes, achieved significant results. The results of this evaluation showed that the proposed method is capable of diagnosing heart diseases with very high accuracy, so that the Accuracy and F-measure reached 99.67% and 0.9958, respectively. The findings demonstrate the high efficiency of the suggested approach in extracting features related to heart diseases and their optimal use in the classification process.