Evaluating the Harmonization of Native Digital and Digitized ECGs for ECG-AI Research
In the authors' words
Large epidemiologic studies have historical electrocardiogram (ECG) collections limited to paper tracings or scanned images rather than native digital files, limiting ECG-AI applications. We compared 15 ECGs available as native digital XML files and as PDF tracings digitized using ECGScan software. Both versions were processed using an ensemble of five convolutional neural networks predicting 10-year heart failure risk. Predictions were strongly correlated (Pearson r=0.804; Spearman {rho}=0.893). Predicted probabilities were (mean {+/-} SD) 0.195 {+/-} 0.048 for native digital ECGs and 0.171 {+/-} 0.042 for digitized ECGs. These preliminary findings support further evaluation of digitized ECGs for ECG-AI applications.
Appeared: Wednesday, September 23. medRxiv. Preprint, not yet peer-reviewed.