The Necessity of Multiple Data Sources for ECG-Based Machine Learning Models.

Area of Science:

• Medical Informatics
• Machine Learning
• Cardiology

Background:

• Machine learning (ML) shows growing interest in medicine, yet clinical relevance is hindered by data quality and interoperability issues.
• Standardized electrocardiogram (ECG) datasets are expected to be interoperable, but variations may exist.
• The generalization capability of ML models trained on single-site ECG data is a critical concern.

Purpose of the Study:

• To examine site- and study-specific differences in publicly available standard ECG datasets.
• To investigate whether subtle study peculiarities affect the stability of trained ML models.
• To assess the generalization of ML results from single-site ECG studies across different datasets.

Main Methods:

• Performance evaluation of modern network architectures on diverse ECG datasets.
• Application of unsupervised pattern detection algorithms across different datasets.
• Comparative analysis of ML model stability using varied ECG data sources.

Main Results:

• Identified significant site- and study-specific differences in standard ECG datasets.
• Demonstrated that minor data peculiarities can notably affect ML model stability.
• Observed variations in the generalization performance of ML models across datasets.

Conclusions:

• Data quality and interoperability issues in ECG datasets pose challenges for ML model generalization in medicine.
• Even seemingly minor variations in study design or data collection can impact the reliability of ML models.
• Further research is needed to ensure the robustness and clinical applicability of ML models trained on ECG data.