With advancements in neurointerventional techniques–particularly after the publication of five large randomized controlled trials in the New England Journal of Medicine in 2015–mechanical thrombectomy (MT) has become the standard treatment for acute-phase anterior circulation acute ischemic stroke (AIS) (1–5). However, meta-analyses show that only 46% of patients achieve favorable outcomes, with a significant proportion of patients still experiencing poor outcomes or death (6). This underscores the urgent need for reliable predictive models to identify appropriate candidates for MT. Several traditional scoring systems have been developed using logistic regression to predict MT outcomes. These include: The Pittsburgh Response to Endovascular therapy (PRE) score (7), which considers age, National Institutes of Health Stroke Scale (NIHSS), and Alberta Stroke Program Early CT Score (ASPECTS); The Stroke Prognostication using Age and NIH Stroke Scale (SPAN) score (8), based on age and NIHSS; The Totaled Health Risks in Vascular Events (THRIVE) score (9), which includes age, NIHSS, and chronic comorbidities; and The Houston Intra-Arterial Therapy (HIAT) score (10), which incorporates age, NIHSS, and admission blood glucose, and its updated version, HIAT2 (11), which also includes ASPECTS. The area under the receiver operating characteristic (ROC) curve for these scores ranges from 0.56 to 0.79, highlighting the need for improved predictive accuracy.

Machine learning (ML) enables high-precision analysis of large clinical datasets using artificial intelligence algorithms. A recent meta-analysis of MT outcome prediction models found that machine learning-based models generally outperform traditional models in predicting the outcomes of neurointerventional procedures, while emphasizing the importance of external validation for model generalizability. However, existing machine learning models for predicting MT outcomes are predominantly based on post-procedural clinical data (12–14). Only a limited number of studies have incorporated only preprocedural clinical data and the simple imaging score ASPECTS, generally overlooking the key imaging features of arterial calcified plaques (15). Intracranial artery calcification (IAC), a common imaging finding on non-contrast cranial computed tomography (CT), is closely associated with atherosclerotic plaques and potentially exerts a more direct impact on MT outcomes (16). Although previous studies have explored the association between IAC and MT outcomes, their conclusions have shown significant discrepancies (17, 18). Therefore, we conducted a comprehensive quantitative assessment of IAC and evaluated its potential as an imaging biomarker for MT prognosis. Additionally, we aimed to develop machine learning models using emergency preprocedural data to aid in the precise identification of AIS patients most likely to benefit from MT.

The machine learning models developed and validated in this study, based on preoperative clinical data and imaging features of arterial calcification, demonstrated strong predictive performance. To our knowledge, this is the first study to incorporate preoperative emergency data–including IAC–into machine learning frameworks to predict MT outcomes. By limiting feature selection to preoperative data, the models maximized clinical applicability and provided actionable guidance for emergency physicians in making personalized diagnostic and therapeutic decisions.

Among the 11 machine learning models constructed, the Extra Trees model demonstrated superior predictive performance, achieving AUC values exceeding 0.8 across the training, internal validation, and external test sets. The primary advantage of machine learning models over traditional logistic regression lies in their ability to automatically capture nonlinear relationships and feature interactions, thereby reducing collinearity effects and exhibiting stronger predictive capabilities in high-dimensional data (22). The Extra Trees algorithm employs ensemble learning with decision trees, balancing performance and efficiency in binary classification tasks through dual randomness–random feature selection and random split points–making them particularly more suited for handling high-dimensional noise and interaction effects in complex medical data (23). To enhance interpretability, the SHAP method was used to explain the model results and demonstrate feature importance rankings. The three most critical predictors were a history of hypertension, serum albumin level, and TCV, followed by age, creatinine level, admission blood glucose level, platelet count, NIHSS score, neutrophil-to-lymphocyte ratio, and onset-to-admission time. Among these, age and NIHSS score were commonly included in previous scoring systems, while emergency admission blood glucose aligned with the HIAT score metrics. Both the HIAT2 score and the machine learning model constructed by Nishi et al. (15) included only the simplified semi-quantitative imaging indicator ASPECTS, without incorporating advanced neuroimaging features. The results of this study suggest that TCV evaluation of IAC is a more suitable imaging predictor than ASPECTS. Hypertension, a traditional risk factor for cerebrovascular disease, may influence MT outcomes through multiple mechanisms (24). Prolonged, uncontrolled preoperative hypertension can exacerbate atherosclerosis, increase vascular tortuosity and thrombus burden, and thereby hinder the passage of thrombectomy devices and prolong procedural times. Intraoperative and postoperative hypertension may trigger cerebral hyperperfusion syndrome, leading to blood–brain barrier disruption, increased risk of brain edema, and a higher probability of hemorrhagic transformation. Albumin may exert protective effects on MT outcomes through mechanisms such as antioxidative stress, anti-inflammation, maintenance of osmotic pressure, protection of vascular endothelial function, and stabilization of blood–brain barrier integrity (25). In Yao et al.’s (12) MT machine-learning prediction model, which was based on postoperative laboratory indicators, albumin was the core predictor. This study further validated the predictive role of albumin level using preoperative data. The remaining model features consisted of routine clinical characteristics and laboratory results available in most hospitals, thereby ensuring clinical practicality.

IAC was previously considered a marker of irreversible vascular aging; however, recent studies have revealed that it involves multiple pathophysiological mechanisms and represents a dynamically regulated process (26) that closely associated with the onset, development, and prognosis of AIS (27, 28). Vascular lumen stenosis and hardening caused by calcification may more directly influence MT treatment outcomes in patients with AIS. Nevertheless, current research conclusions remain inconsistent. A semiquantitative study conducted by Haussen et al. (17) suggested that extensive calcification of the intracranial internal carotid artery does not affect the clinical outcomes of endovascular treatment. In contrast, Hernández-Pérez et al. (29) reported that intracranial carotid artery calcification volume was associated with MT outcomes at 90 days post-treatment, based on a cohort of 194 patients who were either unresponsive to thrombolytic therapy or had contraindications. Compagne et al. (30) conducted a post hoc subgroup analysis of 344 patients with anterior circulation stroke from the 2015 MR CLEAN trial, finding better efficacy in those with medial calcification patterns, though no significant correlation between IAC volume and outcomes. A 2023 study by Rodrigo-Gisbert et al. involving 393 patients with AIS qualitatively indicated that symptomatic IAC could predict endovascular treatment outcomes (18). A recent prospective study by Sierra-Gómez et al. (31), utilizing a semi-quantitative visual calcification score, indicated that carotid artery calcification was associated with larger infarct volumes and poorer outcomes following MT. Despite numerous studies, the conclusions remain heterogeneous, with most relying on qualitative or semi-quantitative assessment and lacking comprehensive quantitative evaluation of IAC indicators. In this study, we conducted a thorough quantitative assessment of IAC, identifying it as an important predictor of MT outcomes. These findings not only clarify the relationship between IAC and MT success but also offer valuable insights for constructing high-performance machine learning prediction models incorporating plaque imaging features.

The potential mechanisms underlying the predictive value of IAC involve several pathways, including a direct increase in plaque hardness as a marker of atherosclerosis, thereby contributing to increased surgical difficulty (32). Additionally, microcalcifications arising from the necrosis or apoptosis of lipid core cells may increase the risk of plaque rupture (33). Additionally, IAC may impair vascular endothelial function and reduce the vessel’s buffering capacity and compensatory capacities in response to blood flow changes (34, 35). The superiority of TCV over ICV and other calcification metrics is attributed to its more accurate reflection of overall IAC distribution. In large vessel occlusion of the anterior circulation, cerebral blood flow critically depends on collateral supply from the contralateral anterior circulation and the posterior circulation via the anterior and posterior communicating arteries. Thus, assessing calcification in all major intracranial arteries, including the contralateral anterior circulation arteries, the vertebral arteries, and the basilar arteries, provides a more comprehensive evaluation of the impact of IAC on collateral compensatory flow compared to conventional approaches that evaluate only the occluded vessel.

The machine learning models developed in this study can assist physicians in predicting treatment outcomes more accurately, this enabling more personalized treatment plans. However, our study had some limitations. First, as a retrospective study, it may be subject to inherent biases. Second, the study included a relatively small number of patients in the model development cohort. Third, although internal and external validation was performed, the model’s generalizability requires further testing in diverse populations (e.g., across different races or countries). Future large-scale, multicenter prospective studies across various cohorts, including non-East Asian populations, are needed to refine and optimize the prediction models. Despite these limitations, establishing a reliable model for predicting MT outcomes holds significant value for guiding clinical decision-making.

In conclusion, total IAC volume is a critical predictor of MT outcomes in patients with anterior circulation AIS. The machine learning models developed using preoperative emergency data demonstrated strong predictive performance, offering a robust theoretical foundation for clinicians to more accurately identify patients with AIS who are suitable candidates for MT.