Skip to content

Abstract 7027: Deep Learning-Based Digitization of Paper Ventricular Tachycardia ECGs for Downstream VT Exit Localization

Nov 2026 · Circulation · 0 citations

Abstract

Background: Many VT ECGs remain available only as paper printouts, scanned documents, or image-based records, limiting their reuse in computational systems that require waveform-level QRS morphology and inter-lead spatial information. Hypothesis: We hypothesized that digitized paper VT ECGs would preserve QRS morphology and inter-lead spatial information sufficient to maintain downstream VT exit localization predictions relative to reference digital ECGs. Methods: A two-stage deep learning and deterministic reconstruction pipeline was developed. Model A segmented paper ECG images into waveform trace, grid, and artifact probability maps. Model B estimated row-wise trace-centerline maps for the three short-lead rows and the long-lead II rhythm strip. These outputs were integrated to identify valid signal windows, extract and repair row-wise trajectories, estimate grid-based paper scale, and reconstruct calibrated 12-lead ECG waveforms. This validation assumed a template-defined standard 12-lead layout. Waveform fidelity was evaluated by comparing digitized ECGs with paired reference digital ECGs using signal-to-noise ratio (SNR) on external UK-VT and VT-ECGs test sets. Downstream utility was assessed by applying a previously developed automated 12-lead ECG-based VT exit localization system to 25 matched reference-digitized VT ECG pairs from 12 patients. Results: The final reconstruction pipeline achieved mean 12-lead SNRs of 16.16 dB (95% CI, 15.93-16.38) on UK-VT (n = 1,043) and 17.88 dB (95% CI, 17.24-18.55) on VT-ECGs (n = 32). In the downstream localization cohort, against paired reference digital ECGs, digitized ECGs achieved a mean patient-level MSE of 0.003 and a mean SNR of 17.19 dB. Patient-level SNR ranged from 15.79 to 19.91 dB. Across 12 patients comprising 25 matched reference-digitized VT ECG pairs, digitized ECGs produced the same patient-level VT exit localization prediction as the paired reference digital ECGs in all 12 patients. Conclusions: Automated paper-ECG digitization recovered waveform-level VT signals across external test sets and preserved downstream VT exit localization prediction agreement with paired reference digital ECGs in all evaluated VT patients. These findings support a step toward input-format-agnostic ECG-based VT exit localization by converting standard-layout paper, scanned, or PDF/image-based VT ECGs into waveform-level signals usable by downstream localization workflows.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.