Adversarial Multi-Task Learning for Robust End-to-End ECG-based Heartbeat Classification.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Manual diagnosis of heart arrhythmias is time-consuming and prone to errors.
• Automated arrhythmia classification offers potential for improved efficiency and accuracy.
• Existing methods may struggle with patient-to-patient generalization due to biological variations.

Purpose of the Study:

• To develop an automated heart arrhythmia classifier with improved generalization across patients.
• To introduce an adversarial multitask learning framework for heartbeat classification.
• To reduce classification errors in detecting various types of heart arrhythmias.

Main Methods:

• An end-to-end deep neural network (DNN) system was designed, comprising a generator and two discriminators (heartbeat-type and subject).
• Adversarial multitask learning was employed, where the generator was trained to be "friendly" to heartbeat classification and "hostile" to subject discrimination.
• The network was trained on raw ECG signals to classify five heartbeat types: normal, right bundle branch blocks (RBBB), premature ventricular contractions (PVC), paced beats (PB), and fusion of ventricular and normal beats (FVN).

Main Results:

• The proposed adversarial multitask learning method achieved a 17% reduction in classification error compared to a baseline fully-connected DNN classifier.
• The system demonstrated improved generalization capabilities in heartbeat arrhythmia classification.
• The approach was validated using the MIT-BIH arrhythmia dataset.

Conclusions:

• Adversarial multitask learning is an effective strategy for enhancing the generalization of automated arrhythmia classifiers.
• The developed subject-independent system can assist clinicians in the heart arrhythmia diagnosis process.
• This research validates the potential for robust, automated detection of heart rhythm abnormalities.