Overview
Explainable AI framework for ECG models (ExECG) is a Python framework designed to address variations in practical pipelines and reporting conventions across ECG Explainable AI (XAI) studies. The framework provides a three-stage pipeline to standardize operations related to ECG diagnostic models, which utilize deep learning for tasks such as arrhythmia classification and abnormality detection.
Research Context
Deep learning models have demonstrated robust performance in ECG diagnostics, particularly for arrhythmia classification and abnormality detection. However, their utility in clinical deployment is constrained by the absence of explanations detailing why a specific output was generated. This limitation affects justification, error analysis, and trust in these models. Although ECG XAI has undergone extensive investigation and continuous improvements, the lack of standardized practices for pipelines and reporting conventions across different studies hinders their reuse and reproducibility.
Approach
The ExECG framework was developed with a three-stage pipeline:
- Wrapper: This stage is designed to standardize access across heterogeneous ECG formats and various intermediate representations. Its function is to provide a consistent interface regardless of the data’s original structure.
- Explainer: This component unifies diverse XAI methods under a shared execution protocol. This standardization aims to streamline the application and comparison of different interpretability techniques.
- Visualizer: The Visualizer supports consistent cross-method comparison within a unified interface. This enables researchers to compare results from various XAI methods systematically.
The framework’s end-to-end usage is demonstrated through concise examples and two case studies. These demonstrations highlight its capacity to facilitate interoperable and reproducible ECG explainability.
Findings
The ExECG framework provides a structured approach for integrating and comparing various XAI methods for ECG models. It standardizes the processes from data access to explanation visualization, thereby addressing issues of heterogeneity and inconsistency prevalent in existing ECG XAI research. The framework's design facilitates interoperability and reproducibility in ECG explainability. The presented case studies demonstrate its functionality and practical application for these objectives.