Detection and classification of arrhythmia using an explainable deep learning model.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Early arrhythmia detection is crucial for treatment and mortality prevention.
• Existing deep learning models for arrhythmia detection lack transparency.
• Explainable AI is needed to build trust and facilitate clinical adoption.

Purpose of the Study:

• To develop an explainable deep learning model (XDM) for arrhythmia classification.
• To validate the XDM's performance and explainability using diverse datasets.
• To compare the XDM with conventional deep learning models.

Main Methods:

• Developed an XDM using a neural network-backed ensemble tree with six feature modules.
• Utilized the Sejong dataset (86,802 ECGs) for internal validation.
• Externally validated the XDM on 36,961 ECGs from four independent datasets.

Main Results:

• Achieved high average AUCs of 0.976 (internal) and 0.966 (external) for arrhythmia classification.
• Demonstrated strong explainability with AUCs ranging from 0.925-0.991.
• Outperformed a previous simple multi-classification deep learning model.

Conclusions:

• The XDM accurately classifies arrhythmia from diverse ECG formats and provides clear decision rationale.
• Explainable deep learning improves accuracy and transparency over conventional methods.
• The XDM shows potential for enhanced clinical application due to its transparency.