A clinically-oriented XAI framework for arrhythmia triage via Hierarchically-Decomposed Neural Attribution: From...

Area of Science:

• Computational Biology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning models excel in biomedical signal analysis but lack interpretability, limiting critical application deployment.
• Explainable AI (XAI) is crucial for understanding and trusting AI decisions in healthcare.

Purpose of the Study:

• To present the Hierarchically-Decomposed Neural Attribution (HDNA) framework, a novel XAI methodology for interpretable biomedical data analysis.
• To apply HDNA to electrocardiogram (ECG) signal processing for enhanced transparency and performance.

Main Methods:

• HDNA utilizes a dual-stream architecture: a rhythmicity stream for temporal event identification and a morphology stream for feature localization.
• A "Where-then-What" reasoning pipeline with structural gating reduces computational overhead by 90% compared to traditional gradient methods.
• The framework achieves real-time inference (2.4 ms latency) on standard CPU hardware with a lightweight design (1.4M parameters).

Main Results:

• HDNA demonstrated robust performance in cardiac arrhythmia detection on PTB-XL (Macro F1: 0.760 ± 0.008) and MIT-BIH (Weighted F1: 0.89 ± 0.01) datasets.
• High explanation faithfulness was achieved, with 93.0% pointing accuracy and 0.78 ± 0.04 IoU.
• The method offers efficient explanation generation suitable for resource-constrained systems and real-time monitoring.

Conclusions:

• HDNA provides a transparent and interpretable machine learning methodology for biomedical signal analysis.
• The framework's efficiency and accuracy make it suitable for real-time applications and diverse computational biology tasks.
• This work advances the deployment of trustworthy AI in critical biomedical applications.