Explainable AI model for ECG signal assessment

• Main Author: Low, Stefanie Jing Ting
• Other Authors: Vidya Sudarshan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/166107
• Institution: Nanyang Technological University

The use of artificial intelligence (AI) systems can automate the onerous task of manually interpreting ECG parameters to detect cardiac diseases. However, to instill confidence in such systems, it is crucial to ensure performance, transparency and interpretability in these systems. This report presents a novel hybrid approach that leverages on the combination of Residual Network (ResNet) and Discrete Wavelet Transformation (DWT) to achieve accurate predictions of cardiac diseases based on ECG signals. The proposed methodology is evaluated on the PTX-XL dataset, the largest publicly accessible 12-lead ECG dataset. The hybrid model, which incorporated features extracted from DWT, achieved higher true positive rates for cardiac diseases compared to the end-to-end ResNet. This is especially crucial as misidentifying a positive patient as negative could delay the treatment process and have high costs. Our findings demonstrate the efficacy of the hybrid approach in enhancing performance, especially in accurately predicting ECG samples of patients with cardiac diseases. This can ultimately have significant clinical implications in terms of early detection and treatment of these conditions. To enhance transparency and interpretability of the model's predictions, we have incorporated Shapley Additive explanations (SHAP) and interpret the predictions at both the local and global level. Our detailed SHAP analysis highlighted how the learned features from ResNet and features extracted using DWT work synergistically to capture complex patterns and characteristics, enabling more accurate identification of cardiac diseases. The SHAP analysis also effectively identified key predictors for the predictions, which we attempted to validate by comparing them to medical knowledge. This enabled us to justify the usefulness of incorporating DWT features and proposing our hybrid approach. Overall, our SHAP analysis enhances the interpretability of our proposed hybrid approach, enhancing trust when using such AI systems.