Explainable AI for CNN-LSTM Network in PCG-Based Valvular Heart Disease Diagnosis

• Published in: 2024 14th International Conference on Cloud Computing, Data Science & Engineering (Confluence) pp. 92 - 97
• Main Authors: Divakar, Chelluri, Harsha, Ramanu, Radha, Kodali, Rao, Dulipalla Venkata, Madhavi, Nadakuditi, Bharadwaj, Thiruveedhula
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18.01.2024
• Subjects: CNN-LSTM; Convolutional neural networks; Decision making; Explainable AI; Feature extraction; Grad-CAM; Phonocardiogram signals; Predictive models; Real-time systems; Rendering (computer graphics); VHD; XAI
• ISSN: 2766-421X
• DOI: 10.1109/Confluence60223.2024.10463207

Globally, valvular heart diseases (VHDs) account for a major portion of deaths and illnesses. An accurate and timely identification of VHDs is essential for directing proper treatment and enhancing patient outcomes. Phonocardiogram (PCG) signals provide a non-invasive and affordable means of capturing acoustic information about the cardiac cycle, rendering them suitable for VHD detection. The proposed method provides an explainable artificial intelligence (XAI) framework for PCG-based VHD diagnosis using convolutional neural network (CNN) - long short-term memory (LSTM) (CNN-LSTM) network. The proposed framework leverages the strengths of deep learning to achieve high diagnostic accuracy while providing interpretability using XAI for the model's predictions. Data augmentation techniques are utilized to augment the PCG signals. Mel-spectrograms are used to extract the relevant features from the PCG signals. The learning model consists of CNN architecture and a layer of LSTM making a CNN-LSTM architecture. The proposed CNN-LSTM model will be a 5-class classifier with classes named aortic stenosis, mitral regurgitation, mitral stenosis, mitral valve prolapse, and normal. The XAI technique employed is gradient-weighted class activation mapping (Grad-CAM), enabling the visualization of model decision-making by generating heatmaps. An impressive accuracy of 97.5% has been achieved using CNN-LSTM model. The integration of XAI ensures a comprehensive interpretation of the model, enhancing transparency for potential real-time clinical deployment.