Publication

Abstract : Deep learning models have been prominently applied for the automatic detection of various cardiovascular conditions using ECG signals. The concept of explainable artificial intelligence is all about finding whether the deep learning framework has captured the appropriate characteristics for the detection task instead of learning some unsolicited approaches (learning features that do not reflect the properties of the instance being classified). So that the health care professionals who use the artificial intelligence system can be confident about the diagnosis obtained from the model. In this chapter, we establish an explainable artificial intelligence method through the multi-label ECG classification. Our proposed method employs a convolutional neural network (CNN) trained using 2D matrices constructed from various leads of ECG recordings. In this work, we show that training the CNN with only a single label per ECG recording is sufficient for the CNN to capture the characteristics of multi-label ECG points. The proposed model when tested with ECG signals containing multi-label information was found that the output probabilities of the correct labels obtained from the softmax layer are in the same order of magnitude. This leads to an explainable framework of how the proposed network correctly gets activated when it sees multiple features pertaining to different heart diseases in the same ECG signal despite having trained only with a single label for each ECG recording. This establishes the fact that the CNN has captured the correct features for the categorization of ECG instead of some undesirable features which are local to our dataset. Further, thresholding is applied to the probabilities from the softmax layer of the proposed CNN, leading to the multilabel categorization of ECG containing up to two labels. The number of ECG records in the training and test set is 6,311 and 280, respectively. The proposed model is evaluated with common performance metrics analyzed in the multi-label classification problem. The model achieved the following scores in different metrics considered: subset accuracy - 0.962, hamming loss - 0.037, precision - 0.986, recall - 0.949 and F1-score - 0.967. Thus, our proposed model is an explainable artificial intelligence-based approach for the multi-label classification of ECG recordings of patients which can be used for computeraided diagnosis of heart diseases confidently.