TCT-449 Automated SYNTAX Score Prediction Using Deep Learning Outperforms Expert Readers in Coronary Angiography Assessment
Recommended Citation
Asslo G, El-Yamani N, Harrabi S, Marquis-Gravel G, Nosair M, Abdelhai O, Tanguay J, Delfrate J, Avram R. TCT-449 Automated SYNTAX Score Prediction Using Deep Learning Outperforms Expert Readers in Coronary Angiography Assessment. J Am Coll Cardiol 2025; 86(17 Supplement):B197-B198.
Document Type
Conference Proceeding
Publication Date
10-28-2025
Publication Title
J Am Coll Cardiol
Abstract
Background: SYNTAX score calculation remains subjective and time-consuming despite its critical role in revascularization decisions. We developed an automated deep learning system for end-to-end SYNTAX score prediction from multi-view coronary angiography. Methods: DeepCORO-CLIP (pretrained on 169,000 video-text pairs, Montreal Heart Institute 2017-2024) was fine-tuned on CardioSyntax dataset (1,844 patients, multi-view angiograms) to predict continuous SYNTAX scores (global/left/right). Split: 1,475 training, 369 test. Six interventionalists annotated exams using majority voting, categorizing: no disease (SYNTAX=0), mild (<18), moderate (18-27), severe (>27). AI outputs recategorized using optimized thresholds (≤2.23, 2.23-18.50, 18.50-22.95, >22.95). Metrics: Pearson correlation, Cohen's kappa, accuracy. Results: AI achieved strong correlations with ground truth SYNTAX scores (Pearson r=0.824 global, 0.795 left, 0.781 right). AI's categorical agreement (κ=0.558) significantly outperformed human raters (κ=0.478; p<0.001) and majority voting (73.8% accuracy vs 67.0% accuracy, p<0.001). Overall accuracy: AI 73.8% vs individual raters 65.9%, 66.5%, 55.9%. By severity, AI achieved 83.6% for no disease, 65.6% mild, 45.0% moderate, 67.0% severe. [Formula presented] Conclusion: Deep learning SYNTAX prediction achieved physician-level performance with statistically significant superiority over individual raters and majority voting. Notably, for moderate disease (SYNTAX 18-27) where clinical stake making for revascularisation is highest, AI achieved 45.0% accuracy, highlighting potential to standardize assessment in this challenging category. Categories: CORONARY: Artificial Intelligence: Coronary
Volume
86
Issue
17 Supplement
First Page
B197
Last Page
B198
