Electrocardiogram Abnormality Detection Using Machine Learning on Summary Data and Biometric Features.

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Electrocardiogram (ECG) data are crucial for diagnosing cardiovascular diseases.
• Manual ECG interpretation is labor-intensive and error-prone.
• Machine learning (ML) offers automated ECG abnormality classification.

Purpose of the Study:

• To evaluate ML models for classifying ECG abnormalities.
• To utilize a dataset combining clinical and ECG biometric data, excluding time-series information.
• To identify key features for ECG abnormality classification.

Main Methods:

• Data preprocessing included handling class imbalance, outliers, feature scaling, and categorical encoding.
• Five ML models (Gaussian Naive Bayes, SVM, Random Forest, Extremely Randomized Trees, Gradient Boosted Trees) and an ensemble were trained.
• Stratified k-fold cross-validation and a reserved testing set were used for model optimization and evaluation.

Main Results:

• Extremely Randomized Trees achieved the highest performance (66.79% accuracy, 66.79% recall, 62.93% F1-score).
• Key predictive features included ventricular rate, QRS duration, and QTC (Bezet).
• Class imbalance and feature overlap presented challenges, particularly for borderline cases.

Conclusions:

• ML models, especially Extremely Randomized Trees, show potential for ECG abnormality classification using non-time-series data.
• The exclusion of time-series ECG signals limits current diagnostic accuracy.
• Future research should integrate time-series data and deep learning for enhanced clinical relevance.