Hygeia: A Multilabel Deep Learning-Based Classification Method for Imbalanced Electrocardiogram Data.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Electrocardiograms (ECGs) are crucial for heart disease diagnosis.
• Current ECG analysis often provides only general or limited classifications.
• Detailed ECG analysis presents a complex multilabel classification challenge.

Purpose of the Study:

• To develop Hygeia, a deep learning method for comprehensive ECG classification.
• To address the multilabel nature of detailed ECG analysis.
• To improve diagnostic accuracy and sensitivity for various heart conditions.

Main Methods:

• A guidance model transforms multilabel classification into interrelated binary classifications.
• Data generation and mixed sampling techniques address sample imbalance in ECG datasets.
• The Hygeia model is trained to classify 55 types of ECG abnormalities.

Main Results:

• For 11 imbalanced ECG types, performance improved significantly: average accuracy rose from 87.74% to 92.68%, sensitivity from 43.11% to 96.92%.
• The F1 score increased from 0.3929 to 0.9287, and accuracy from 0.3929 to 99.47%.
• For 44 abnormal ECG types, the model achieved 99.69% average accuracy, 95.81% sensitivity, 0.9758 F1 value, and 99.72% accuracy.

Conclusions:

• Hygeia offers a robust deep learning solution for detailed ECG analysis.
• The method effectively handles multilabel classification and data imbalance.
• This approach significantly enhances the accuracy and sensitivity of diagnosing diverse heart conditions from ECGs.