Why the study?
Stanford type-A aortic syndrome and STEMI can show overlapping ECG findings while requiring different treatments, yet labeled ECG resources for this task remain scarce and direct paper-ECG image classification risks exploiting non-physiological image cues.
Population
728 first-admission paper-based 12-lead ECG records from 344 TAAS and 384 STEMI patients
Comparison
Stanford type-A aortic syndrome vs STEMI differentiation via bias-resilient waveform modeling
Design
Cohort-based proof-of-concept machine learning model development and evaluation
Answers are grounded in this paper and cite their sources.
Demonstrates rapid differentiation of aortic syndrome from STEMI using digitized ECG waveforms, indicating reduced image-bias risk.
Stanford type-A aortic syndrome (TAAS) and ST-segment elevation myocardial infarction (STEMI) are time-critical cardiovascular emergencies that can show overlapping electrocardiographic findings but require fundamentally different treatment strategies. Differentiation at the first ECG assessment is therefore critical, yet labeled ECG resources for this task remain scarce, especially for physician-confirmed TAAS cases from emergency-care settings. Here, we introduce the Anzhen Dataset, a paper-based 12-lead ECG cohort comprising 728 first-admission records from 344 TAAS and 384 STEMI patients, providing a clinically grounded resource for this underexplored differential diagnosis task. We also identify a shortcut-learning risk in direct paper-ECG image classification, in which visual models exploit non-physiological image cues rather than electrophysiological morphology, limiting their reliability in real-world application scenarios. To reduce this risk, we propose a bias-resilient waveform modeling framework that converts paper ECGs into standardized 12-lead one-dimensional signals through perspective correction, lead localization, grid and noise removal, waveform digitization, and interpolation-based compensation. A lightweight ConvNeXt-Femto model then combines morphological analysis, low-rank feature compression, and lead-region relational attention to capture local waveform differences and cross-lead pathological consistency. In the TAAS–STEMI task, the model achieved 0.900 accuracy and 0.906 F1-score with 61,850 trainable parameters and a classifier inference latency of 4.010 ms per sample. Multi-class experiments, ablation studies, and visualization analyses further supported the model’s representation ability, module contribution, and physiological interpretability. These proof-of-concept findings support waveform-centered modeling as a lightweight approach that reduces image-related bias in TAAS–STEMI differentiation from paper-based ECGs; the model is intended only for preliminary risk stratification and cannot replace confirmatory imaging.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: