Background: Cerebral hyperperfusion after carotid artery stenting (CAS) can precipitate early hemorrhagic complications. Using a teacher-student artificial intelligence framework, we examined whether an integrated model combining pre-operative computed tomography perfusion (CTP) and intra-operative time-domain near-infrared spectroscopy (TD-NIRS) would highlight contributors to risk, and whether that insight could be translated into a CTP-free, NIRS-only surrogate designed for intra-operative surveillance. Methods: In this single-center retrospective cohort, we analyzed 44 consecutive CAS procedures (6 hyperperfusion) and 15 balloon occlusion tests between February 2022 and March 2025. The teacher model (multilayer perceptron; leave-one-out cross-validation) used 61 features from TD-NIRS time-series statistics, pre-operative CTP parameters (cerebral blood flow, cerebral blood volume, time-to-peak), and the balloon occlusion test score. A NIRS-only student model (62 TD-NIRS features) was trained via knowledge distillation and evaluated in an independent test set of 11 cases (2 hyperperfusion). Feature attribution was conducted post hoc using permutation importance and SHapley Additive Explanations. Results: The teacher achieved a leave-one-out cross-validated area under the receiver operating characteristic (ROC) curve of 0.978 (sensitivity 0.67; specificity 1.00). The NIRS-only student achieved an area under the ROC curve of 0.78 (sensitivity 1.00; specificity 0.56) in the small independent test set. Attribution ranked lower contralateral-hemisphere cerebral blood flow on pre-operative perfusion imaging, larger intra-occlusion TD-NIRS StO 2 drop, and longer ipsilesional time-to-peak among the top contributors. Conclusions: Model attribution suggests that contralateral hemodynamic reserve may be relevant to hyperperfusion risk after carotid artery stenting. A NIRS-only student model showed high sensitivity in a small test set and may offer a preliminary CTP-free surrogate for intra-operative risk surveillance. These single-center exploratory findings with few events require external validation.
Maruyama et al. (Thu,) studied this question.