Source-Only Transportability of Engineered ECG Features for Healthy-Versus-Myocardial Infarction Classification.

Area of Science:

• Cardiovascular Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Electrocardiogram (ECG)-based myocardial infarction (MI) classifiers often exhibit high performance on internal validation data.
• However, their performance frequently diminishes when applied to external datasets, posing a challenge for real-world clinical deployment.
• This study focuses on the "transportability" of ECG features for MI classification across different data sources.

Purpose of the Study:

• To evaluate the source-only transportability of engineered 12-lead ECG feature families for binary healthy-versus-MI classification.
• To compare performance across a cardiologist-annotated hospital dataset and the PTB-XL dataset.
• To assess the gap between internal validation and cross-dataset transfer performance.

Main Methods:

• Utilized a hospital dataset (1749 recordings) and a matched PTB-XL subset (14,982 recordings) for binary classification.
• Compared eleven engineered feature families and five classifier families using preprocessing, patient-aware splitting, and source-validation.
• Employed bootstrap uncertainty estimation and analyzed both ranking and operating-threshold transportability.

Main Results:

• Internal validation performance exceeded strict source-only transfer performance in both datasets.
• Highest balanced accuracy was 0.775 (hospital) and 0.898 (PTB-XL) internally; source-only transfer yielded lower, asymmetric results (max 0.632).
• Operating-threshold transportability diverged significantly, with hospital-to-PTB-XL transfer showing poor sensitivity (0.164) at the source-selected threshold.

Conclusions:

• A notable transportability gap exists for ECG-based MI classifiers between different data sources.
• Neither hospital-to-PTB-XL nor PTB-XL-to-hospital transfer achieved performance comparable to internal validation.
• Source-only operating points are insufficient for clinical MI screening without further calibration and validation.