Towards Quantitative Precision for ECG Analysis: Leveraging State Space Models, Self-Supervision and Patient Metadata

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models, particularly convolutional neural networks, dominate automatic electrocardiogram (ECG) analysis.
• Existing methods face challenges in capturing long-term dependencies and leveraging diverse data effectively.

Purpose of the Study:

• To improve the quantitative accuracy of deep learning-based ECG analysis systems.
• To investigate the impact of structured state space models (SSMs), self-supervised learning, and demographic metadata on ECG analysis performance.

Main Methods:

• Exploration of structured state space models (SSMs) for enhanced time series analysis.
• Application of self-supervised learning (contrastive predictive coding) to improve feature representation.
• Integration of basic demographic metadata with ECG signals as input.

Main Results:

• SSMs improve performance over convolutional models, with no benefit from >100 Hz sampling rates or >3 second input windows for standard tasks.
• Self-supervised learning further boosts SSM performance by learning robust features.
• Inclusion of demographic metadata consistently enhances predictive performance.

Conclusions:

• SSMs, self-supervised learning, and demographic metadata are key components for advancing ECG analysis algorithms.
• Optimal data parameters for ECG analysis include 100 Hz sampling rate and 3-second input windows.
• Future ECG analysis systems should integrate these components for improved accuracy and insights.