Digitizing ECG image: new fully automated method and open-source software code

Abstract

Background: We aimed to develop and validate an automated, open-source code ECG-digitizing tool and assess agreements of ECG measurements across three types of median beats, comprised of digitally recorded, simultaneous and asynchronous ECG leads, and digitized asynchronous ECG leads. Methods: We used the data of clinical studies participants (n=230; mean age 30 ± 15 y; 25% female; 52% had the cardiovascular disease) with available both digitally recorded and printed on paper and then scanned ECGs, split into development (n=150) and validation (n=80) datasets. The agreement between ECG and VCG measurements on the digitally recorded time-coherent median beat, representative asynchronous digitized, and digitally recorded beats was assessed by Bland-Altman analysis. Results: Agreement between digitally recorded and digitized representative beat was high [area spatial ventricular gradient (SVG) elevation bias 2.5(95% limits of agreement [LOA] -7.9-13.0) degrees; precision 96.8%; inter-class correlation [ICC] 0.988; Lin concordance coefficient ρc 0.97(95% confidence interval [CI] 0.95-0.98)]. Agreement between digitally recorded asynchronous and time-coherent median beats was moderate for area-based VCG metrics (spatial QRS-T angle bias 1.4(95%LOA -33.2-30.3) degrees; precision 94.8%; ICC 0.95; Lin concordance coefficient ρc 0.90(95%CI 0.82-0.95)], but poor for peak-based VCG metrics of global electrical heterogeneity. Conclusions: We developed and validated an open-source software tool for paper-ECG digitization. Asynchronous ECG leads are the primary source of disagreement in measurements on digitally recorded and digitized ECGs.

A fully-automated paper ECG digitisation algorithm using deep learning

“…Although horizontal anchor points may be identified by dividing the ECG image in half, this approach may not be accurate in ECG configurations where the leads are not equidistant and would only work for 6 × 2 ECG configurations. Other digitisation tools also require manual labelling of anchor points [21][22][23][24]29 www.nature.com/scientificreports/ and restricted in their application by ECG configurations. They are also user-dependent, requiring manual selection of each lead prior to the digitisation process.…”

“…A number of attempts have been made to develop 12-lead ECG digitisation tools [21][22][23] . For example, ECGscan 21 was the first such application to be commercialised but requires significant user input to identify the regions of the ECG that require digitisation.…”

“…For example, ECGscan 21 was the first such application to be commercialised but requires significant user input to identify the regions of the ECG that require digitisation. Similarly, other digitisation tools [22][23][24] also require manual input to ensure that the ECG leads are correctly identified by the end-user. Others have developed ECG digitisation tools to work directly on segmented single-lead ECG image [25][26][27] .…”