An ensemble based lightweight deep learning model for the prediction of cardiovascular diseases from electrocardiogram images

Cardiovascular Disease (CVD) has become a serious reason of death all over the world. According to the Australian Institute of Health and Welfare report, CVD was the underlying cause of 42,700 deaths (25% of all) in 2021. Therefore, accurate detection of CVD is crucial for early treatment. An Electrocardiogram (ECG) signal is used to measure the rhythms and electrical activity of the heart. Traditionally, paper-based ECG images are used to detect CVD which may not be accurate always. The existing pre-trained models such as transfer learning can be used for automatic diagnosis. However, they fall short of multiclass classification due to complex variations of the ECG signals. For this research, four cardiac abnormalities including Abnormal Heartbeats, Myocardial Infarction, History of Myocardial Infarction, and Normal Heartbeats are predicted using the combination of machine learning and deep learning techniques. The deep learning method is used for feature extraction and the machine learning algorithm is used for classification. A lightweight Convolutional Neural Network (CNN) model is designed to extract the prominent features from ECG images and use these features for the prediction of CVDs. However, only a classifier may not fully utilize extracted features for classification. Therefore, the proposed method developed an optimized weighted average ensemble model for classifier selection. The proposed ensemble-based lightweight model showed its ability to extract informative features when tested on real ECG datasets and outperformed the baseline methods by achieving the highest accuracy of 99.29%. • Applied pre-processing methods to remove artifacts and improve the quality of the ECG image. • Investigated four renowned transfer learning models to predict CVD & evaluated their performance. • Proposed a lightweight deep CNN for extracting prominent features from ECG images. • Applied classification algorithms with the best parameters for individual prediction. • Developed an optimized weight model for the ensemble of classifiers for final prediction.