Post Hoc Sample Size Estimation for Deep Learning Architectures for ECG-Classification.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Deep learning models for time series analysis, particularly in electrocardiograms (ECGs), demand substantial training data.
• Traditional sample size estimation methods are inadequate for machine learning applications in ECG analysis.

Purpose of the Study:

• To develop and evaluate a sample size estimation strategy for binary classification tasks using deep learning on ECG data.
• To benchmark sample size requirements across various deep learning architectures and specific ECG conditions.

Main Methods:

• Utilized the large, publicly available PTB-XL dataset (21,801 ECG samples).
• Evaluated binary classification for Myocardial Infarction (MI), Conduction Disturbance (CD), ST/T Change (STTC), and Sex.
• Benchmarked estimations across XResNet, Inception-, XceptionTime, and Fully Convolutional Network (FCN) architectures.

Main Results:

• Identified trends in the required sample sizes for different ECG classification tasks.
• Demonstrated varying sample size needs based on the chosen deep learning architecture.
• Provided data-driven insights into the feasibility of deep learning studies with limited ECG data.

Conclusions:

• The proposed strategy offers valuable orientation for sample size determination in future ECG-related deep learning studies.
• Results aid in assessing the feasibility of deep learning projects based on available data and chosen architectures.
• This work contributes to optimizing the design and resource allocation for machine learning in cardiovascular diagnostics.