A multi-module algorithm for heartbeat classification based on unsupervised learning and adaptive feature transfer.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Machine Learning

Background:

• Deep learning for heartbeat classification faces challenges due to scarce annotated data.
• Existing transfer learning (TL) methods often ignore domain distribution discrepancies between source (SD) and target (TD) databases.
• Inconsistent tasks between SD and TD databases further complicate effective transfer learning.

Purpose of the Study:

• To address the challenge of limited labeled data in heartbeat classification models.
• To develop an effective method for eliminating domain discrepancy between SD and TD databases.
• To improve heartbeat classification performance by adapting features from a different task domain.

Main Methods:

• Proposed a multi-module heartbeat classification algorithm utilizing unsupervised feature extractors.
• Introduced a novel adaptive transfer method to eliminate domain discrepancy between pre-training (PTF-SD) and fine-tuning (FTF-TD) features.
• Employed unsupervised learning and adaptive feature transfer for model pre-training and fine-tuning.

Main Results:

• Achieved an overall accuracy of 96.7% in classifying heartbeats into normal, supraventricular ectopic beats (SVEBs), and ventricular ectopic beats (VEBs).
• Demonstrated high performance for SVEBs with sensitivity (Sen) of 0.802, positive predictive value (PPV) of 0.701, and F1 score of 0.748.
• Showcased excellent results for VEBs with Sen of 0.976, PPV of 0.840, and F1 score of 0.903.

Conclusions:

• The proposed multi-module algorithm effectively mitigates labeled data scarcity in heartbeat classification.
• Unsupervised learning and adaptive feature transfer are key components for successful cross-domain heartbeat classification.
• The method demonstrates robust performance even when source and target domains are derived from inconsistent tasks.