Machine learning-based detection of cardiovascular disease using ECG signals: performance vs. complexity.

Area of Science:

• Cardiology and Medical Informatics
• Artificial Intelligence in Healthcare

Background:

• Cardiovascular disease is a major health concern, necessitating efficient diagnostic tools.
• Electrocardiogram (ECG) interpretation is crucial but requires expertise and time.
• Developing automated methods for early cardiac abnormality detection is vital for improving patient care.

Purpose of the Study:

• To evaluate modern approaches for classifying cardiac diseases using ECG recordings.
• To compare the performance and efficiency of different machine learning models.
• To investigate energy consumption and model interpretability for ECG analysis.

Main Methods:

• Utilized Poincaré representation with deep learning image classifiers.
• Applied one-dimensional convolutional neural networks (1D CNNs) to raw ECG signals.
• Employed XGBoost models for time-series feature prediction.

Main Results:

• The 1D ResNet model achieved the highest F1 scores (85% on CinC 2017, 71% on CinC 2020), outperforming challenge-winning solutions.
• 1D convolutional models demonstrated high specificity and superior energy efficiency (lower power consumption and CO2 emissions).
• Poincaré methods showed promise for Atrial Fibrillation (AF) but not other arrhythmias; XGBoost had long inference times.

Conclusions:

• 1D convolutional models, particularly 1D ResNet, are highly effective and efficient for cardiac disease classification from ECGs.
• Residual connections in 1D CNNs maintain performance while simplifying models.
• Analysis of power consumption and model interpretation provides insights into computational mechanisms and efficiency.