A deep learning model to predict atherectomy device use during PCI from preprocedural CCTA

Background Severe coronary calcification remains a major challenge in percutaneous coronary intervention (PCI), frequently requiring atherectomy-based debulking devices that increase procedural complexity and periprocedural risk. However, the decision to use atherectomy devices is typically made intra-procedurally, and objective preprocedural tools to predict device necessity are lacking. Purpose We aimed to develop and externally validate a deep learning (DL) model using preprocedural cardiac computed tomography angiography (CCTA) to predict the use of atherectomy devices at the lesion level during PCI and to evaluate the incremental value of clinical variables. Methods The dataset included CCTA images of 498 moderate-to-severely calcified coronary lesions from 415 patients who underwent PCI with preprocedural CCTA in 14 cardiovascular hospitals. The dataset was annotated with clinical data including PCI reports. Data from 12 hospitals (401 lesions) were used for model training and internal validation with stratified four-fold cross-validation, and data from 2 independent hospitals (97 lesions) were used as an external validation cohort. We developed a novel transformer-based DL model incorporating a convolutional neural network (CNN) as a feature extractor in a multi-instance learning (MIL) framework and optimized hyperparameters to maximize the AUC. We integrated clinical variables into a multimodal extension of the DL model. We evaluated the performance of the DL model and compared it with that of human readers, based on majority voting of three general cardiologists and three PCI experts, using the McNemar test in the external validation cohort. Results We found that the CCTA-based model achieved an AUC of 0.79 (95% confidence interval [CI], 0.70–0.88) for predicting atherectomy device use at the lesion level in the external validation. The addition of clinical variables resulted in an AUC of 0.80 (95% CI, 0.71–0.89), suggesting the importance of anatomical characteristics to determine the need for atherectomy devices (Figure 1). In a non-clinical competition with blinded human readers, the DL model demonstrated superior accuracy compared with general cardiologists (p=0.005) and comparable accuracy to experts (p=0.093) (Figure 2). Conclusion We demonstrate that a novel DL model enables lesion-level prediction of atherectomy device use in PCI from preprocedural CCTA. This image-based approach may support objective preprocedural risk stratification, facilitate procedural planning, and improve informed decision-making in patients with calcified coronary lesions.