Explainable AI (XAI) for Arrhythmia Detection in ECG

Background: Advancements in deep learning have enabled highly accurate arrhythmia detection in electrocardiogram (ECG) signals. However, these models are often considered "black-box" systems, limiting their clinical adoption due to a lack of interpretability. Explainable AI (XAI) techniques aim to bridge this gap by providing insights into model decisions. Despite the progress in XAI, most techniques are designed for image and feature-based models rather than time series data like ECG signals. This study aims to explore explainability of AI models for arrhythmia detection

Methods: A model was developed utilizing the MIT-BIH arrhythmia dataset. The study employs deep learning techniques, specifically Long Short-Term Memory (LSTM) networks. Data preprocessing includes segmentation based on R-peak, detected using Pan-Tompkins algorithm, and handling class imbalance through Synthetic Minority Over-sampling Technique (SMOTE). To assess performance on a large dataset, an additional 12-lead ECG dataset was included. Various eXplainable AI (XAI) techniques were analyzed and
tested, leading to the use of SHapley Additive exPlanations (SHAP), a method based on game theory. Four different SHAP-based techniques were implemented, tested, and compared (Permutation importance, KernelSHAP, gradients, and Deep Learning Important FeaTures (DeepLIFT). Additionally, medical professionals were consulted to gather their perspectives on the explainability of these models.

Findings: The deep learning model achieved a high validation accuracy of 98.3% on the MIT-BIH dataset. However, performance dropped significantly when trained on the combined dataset, highlighting the importance of consistent measurement environments. Of the four tested methods, permutation importance and KernelSHAP exhibited similar behavior, producing cluttered output. The remaining two methods produced more promising results, showing the points of interests more clearly. These methods provided clear indications of important waveform segments used in classification, aligned with clinical knowledge. However, the patterns they highlighted differed between the two methods and also varied across samples, raising concerns about reliability. Additionally, medical professionals preferred heatmap-style visualizations for interpretability.

Discussion: This study highlights the potential of XAI in improving transparency in arrhythmia detection models. While the deep learning model demonstrated strong classification performance, its reliance on single-heartbeat segmentation limited the detection of arrhythmias that require rhythm context. The study underscores the need for domain-specific adaptations in XAI methods for ECG analysis. While the explainable methods sometimes produced incorrect highlight areas, the arrhythmias were still correctly identified. Future work should focus on refining XAI techniques for time-series data and validating their clinical applicability.