Key result
Deep learning on angiographic video predicts in-stent restenosis with ~20.4-fold higher odds, outperforming cardiologists.
Why the study?
Existing in-stent restenosis prediction models require post-hoc structured data unavailable during PCI, precluding real-time intra-procedural risk stratification. No deep learning framework had previously attempted to predict in-stent restenosis directly from routine coronary angiographic video.
Does an end-to-end deep learning framework operating on routine coronary angiographic video accurately predict in-stent restenosis risk compared to conventional deep learning architectures, clinical tabular models, and expert interventional cardiologists in patients undergoing percutaneous coronary intervention?
Observational (n=431)
Blinded evaluation for interventional cardiologists
No
Does an end-to-end deep learning framework operating on routine coronary angiographic video accurately predict in-stent restenosis risk compared to conventional deep learning architectures, clinical tabular models, and expert interventional cardiologists in patients undergoing percutaneous coronary intervention?
An end-to-end deep learning model analyzing routine pre- and post-PCI coronary angiographic videos accurately predicts in-stent restenosis in real time, significantly outperforming clinical models and expert interventional cardiologists.
Background In-stent restenosis (ISR) remains the principal long-term complication of percutaneous coronary intervention (PCI), affecting 5–10% of drug-eluting stent recipients. All existing prediction models require post-hoc structured data unavailable during the procedure, precluding real-time intra-procedural risk stratification. Deep learning (DL) has transformed cardiovascular image analysis, yet no framework has attempted to predict ISR directly from routine angiographic video. Purpose We aimed to develop and temporally validate an end-to-end DL framework that predicts ISR risk directly from coronary angiographic video at two peri-procedural time points, without any structured clinical input. Methods This retrospective study enrolled 431 patients (518 target lesions) from a tertiary cardiovascular centre. Lesions were divided into a training set (n=278, ISR 110 [39.6%]; 2017–2019), a development set (n=120, ISR 32 [26.7%]; same period) and a temporally independent external-validation set (n=120, ISR 60 [50.0%]; 2021–2022). An attention-based global–local feature-extraction architecture operating on coronary digital subtraction angiography video was used to build two models: pre-PCI video and post-PCI video. Discrimination was assessed by AUC with DeLong 95% confidence intervals (CIs). Three-level head-to-head comparisons were conducted: (1) alternative DL architectures (DenseNet-121, ResNet-50, Swin Transformer) under identical protocols; (2) five clinical models built from clinical structured variables; and (3) six board-certified interventional cardiologists scoring ISR risk from identical video under blinded conditions. Model interpretability was assessed by gradient-weighted class-activation mapping (Grad-CAM). Results In the temporal external-validation set, pre-PCI achieved an AUC of 0.856 (95% CI 0.782–0.920) and post-PCI 0.870 (0.807–0.929). Conventional DL backbones achieved AUCs of only 0.611–0.682 (all P<0.001). The best clinical tabular model reached 0.655 (all P<0.001 vs. our model). In blinded evaluation, the model significantly outperformed pooled expert cardiologists: pre-PCI 0.856 vs. 0.683 (P=0.023) and post-PCI 0.870 vs. 0.723 (P=0.035), with substantially higher sensitivity (0.917 vs. 0.567). Risk stratification yielded odds ratios of 8.27 (95% CI 3.67–19.94; P < 0.001) for pre-PCI and 20.43 (7.65–65.73; P < 0.001) for post-PCI. Grad-CAM revealed model attention concentrated on vessel and stent zones. Conclusion An end-to-end DL model operating on routine coronary angiographic video predicts ISR with clinically meaningful accuracy, significantly outperforming conventional architectures and expert cardiologists. The dual-time-window design enables real-time intra-procedural risk stratification and may guide stent optimisation decisions at the point of care.Figure 1Figure 2
No takes yet. Share an insight, caveat, or question.
Ba et al. (2026) conducted an observational in In-stent restenosis after percutaneous coronary intervention (n=431). End-to-end deep learning model on pre- and post-PCI coronary angiographic video vs. Conventional DL architectures, clinical tabular models, and six board-certified interventional cardiologists was evaluated on In-stent restenosis (ISR) prediction and risk stratification (OR 20.43, 95% CI 7.65-65.73, p=<0.001). An end-to-end deep learning model on angiographic video predicted in-stent restenosis with an AUC of 0.870 post-PCI (OR 20.43; 95% CI 7.65-65.73; P<0.001), outperforming cardiologists.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: