Machine Learning-Based Algorithm to Predict Procedural Success in a Large European Cohort of Hybrid Chronic Total...

CTOs are frequently encountered in patients undergoing invasive coronary angiography. Even though technical progress in CTO-PCI and enhanced skills of dedicated operators have led to substantial procedural improvement, the success of the intervention is still lower than in non-CTO PCI. Moreover, the scores developed to appraise lesion complexity and predict procedural outcomes have shown suboptimal discriminatory performance when applied to unselected cohorts. Accordingly, we sought to develop a machine learning (ML)-based model integrating clinical and angiographic characteristics to predict procedural success of chronic total occlusion (CTO)-percutaneous coronary intervention(PCI). Different ML-models were trained on a European multicenter cohort of 8904 patients undergoing attempted CTO-PCI according to the hybrid algorithm (randomly divided into a training set [75%] and a test set [25%]). Sixteen clinical and 16 angiographic variables routinely assessed were used to inform the models; procedural volume of each center was also considered together with 3 angiographic complexity scores (namely, J-CTO, PROGRESS-CTO and RECHARGE scores). The area under the curve (AUC) of the receiver operating characteristic curve was employed, as metric score. The performance of the model was also compared with that of 3 existing complexity scores. The best selected ML-model (Light Gradient Boosting Machine [LightGBM]) for procedural success prediction showed an AUC of 0.82 and 0.73 in the training and test set, respectively. The accuracy of the ML-based model outperformed those of the conventional scores (J-CTO AUC 0.66, PROGRESS-CTO AUC 0.62, RECHARGE AUC 0.64, p-value <0.01 for all the pairwise comparisons). In conclusion, the implementation of a ML-based model to predict procedural success in CTO-PCIs showed good prediction accuracy, thus potentially providing new elements for a tailored management. Prospective validation studies should be conducted in real-world settings, integrating ML-based model into operator decision-making processes in order to validate this new approach.