AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Quantifying the Effects of ECG Shortening and Digitization on Multitask ECG-AI Performance
By Ahmad S, Syed F
Background: Electrocardiogram (ECG)-based machine learning models are commonly developed using digital 10-second waveforms, whereas many clinical archives store ECGs as images or PDF reports. Prior studies have shown that ECG images can be digitized and analyzed by AI, but the incremental loss attributable to ECG shortening versus subsequent digitization remains incompletely characterized. We quantified performance changes across digital 10-second, 2.5-second digital and digitized ECGs. Methods: We trained otherwise identical ResNet-based multitask models in PTB-XL (n=21,799) using locked patient-level train, validation, and held-out test splits: one on full 12-lead 10-second digital ECG waveforms and the other on 12-lead 2.5-second digital segments derived from the same source ECGs. Models jointly predicted age and 8 binary endpoints: normal ECG, atrial fibrillation, premature ventricular complex, conduction disturbance, myocardial infarction, ST/T abnormality, left ventricular hypertrophy, and female sex. For the digitized experiment arm, held-out 2.5-second ECGs were rendered as clinical-format 3х4 PDFs at 25 mm/s, 10 mm/mV, and 600 dpi, reconstructed with a PDF-to-waveform digitizer, and evaluated using the trained 2.5-second model without retraining or fine-tuning. Digitization fidelity was assessed against the source digital waveforms. Results: Internal held-out testing was performed on 2,073 ECGs. The digitization pipeline achieved a mean signal-to-noise ratio of 19.39 dB and a mean correlation of 99.2% versus original digital waveforms. Age prediction performed the best with 10-second digital (mean absolute error [MAE]: 7.36 years; R ² : 0.691), relative performance declined with 2.5-second digital and digitized 2.5-second ECGs by 0.29 and 0.61 MAE years, and -0.033 and -0.054 R ² , respectively. Across the 3 groups, mean (range) AUROC for the eight binary tasks were 0.932 (0.898 to 0.992), 0.925 (0.881 to 0.986) and 0.922 (0.867 to 0.987), respectively. The mean (and single largest) AUROC loss was 0.007 (0.017) after shortening and 0.010 (0.014) after shortening plus digitization, relative to 10-second digital ECGs. Conclusions: In an internal PTB-XL held-out test set, 2.5-second digitized ECGs largely preserved multitask ECG-AI performance relative to original digital waveforms. Most performance loss was attributable to shortening from 10-second to 2.5-second inputs, with a smaller additional loss from PDF rendering and digitization. These findings quantify the relative impact of temporal shortening and digitization across age prediction and ECG phenotype classification.