Abstract 11016: ESeeG: Vision-Based Pipeline for Image-Domain Wave Delineation and Transparent Clinical Interpretation of Electrocardiograms
Abstract
Background: The 12-lead electrocardiogram (ECG) is the most widely used cardiac diagnostic test. Although deep learning models have demonstrated high performance for ECG interpretation, their generalizability across different ECG formats and clinical environments remains limited. Large-scale, rule-based ECG analysis performed directly on ECG images — without access to the underlying electrical signal — has remained an unmet challenge. Research Question: Can P, QRS, and T wave boundaries be accurately delineated directly on ECG images, and do the resulting measurements support transparent, rule-based clinical interpretation at the level of cardiologist annotations? Methods: ESeeG, a three-stage interpretable pipeline, was developed to process raw ECG images without a standardized layout requirement. Stage 1 (Layout Detector) employs a YOLO-based object detection model (YOLOv8m) trained on 10,000 ECG images (70% train/15% validation/15% test splits) rendered from the PTB-XL database (a publicly available 12-lead ECG dataset, with 21,799 recordings). Stage 2 (Wave Detector) applies a second object detection model to localize P-wave, QRS complex, and T-wave boundaries directly on 35674 (70%/15%/15%) single-lead image crops — requiring no signal digitization. Stage 3 (ECG Interpreter) generates rule-based diagnoses with explicitly cited AHA/ACC guideline thresholds. Validation was performed on 200 ECGs from the LUDB (Lobachevsky University ECG Database), an external dataset absent from all training steps, and on 25 real-world ECG photographs and screenshots. Results: The Layout Detector achieved a mean average precision at 50% IoU threshold (mAP50) of 0.990, with lead assignment accuracy of 99.6% (99.2%-99.8%) across five standard ECG layout formats. The Wave Detector achieved a QRS onset error of –0.2 ± 11.2 ms against expert cardiologist annotations across 2,392 leads (–0.4 to 0.0 ms), within published human inter-observer variability. Stage 3 multi-class accuracy was 98.5% for rate (95.6%-99.5%), 95.3% for rhythm (91.3%-97.5%), 99.5% for electrical axis (97.1%-99.9%), and 96.9% for intraventricular conduction (93.4%-98.6%) on 200 LUDB ECGs, with proof-of-concept generalization demonstrated on 25 real-world ECG images. Conclusions: ESeeG is, to the author's knowledge, the first layout-agnostic, image-domain ECG interpretation model that performs wave boundary delineation directly on ECG images without intermediate signal digitization, providing an auditable foundation for transparent, guideline-based AI-assisted ECG analysis.