Automatic Multichannel Electrocardiogram Record Classification Using XGBoost Fusion Model.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Increasing demand for automated cardiac arrhythmia detection from 12-lead electrocardiogram (ECG) signals.
• Varied contributions of different ECG leads and temporal segments to classification accuracy.
• Need for advanced models that can effectively fuse spatial and temporal features.

Purpose of the Study:

• To propose an automatic classification model for 12-lead ECG signals using model fusion.
• To focus on representative spatial and temporal features for enhanced cardiac arrhythmia detection.
• To evaluate the performance of the proposed CBi-DF-XGBoost model against existing methods.

Main Methods:

• Feature extraction using a convolutional neural network (CNN) for local features.
• Temporal feature extraction via bi-directional long short-term memory (BiLSTM).
• Fusion of CNN, BiLSTM, and domain-specific features using eXtreme Gradient Boosting (XGBoost).

Main Results:

• Achieved a macro-average F1 score of 0.825 and micro-average AUC of 0.919 for classifying nine ECG signal categories.
• Demonstrated superior performance compared to common network structures and advanced ECG classification algorithms, particularly in F1 score.
• Ablation studies confirmed the benefits of 12-lead complementary information and domain-specific features.

Conclusions:

• The XGBoost-based fusion model (CBi-DF-XGBoost) is effective and feasible for classifying 12-lead ECGs into nine common heart rhythms.
• Findings hold clinical importance for early arrhythmia diagnosis and suggest avenues for further research.
• The multichannel feature fusion algorithm has potential applications in other physiological signal analyses.