Worldwide, Acute ischemic stroke (AIS) ranks as a crucial contributor of mortality and disability, with an increasing burden due to aging populations and rising incidence among younger individuals (1, 2). Tandem lesions (TL) are a more severe form of cerebrovascular disease, characterized by significant stenosis or occlusion in extracranial vessels, coupled with distal intracranial vessel occlusion (3, 4). Endovascular thrombectomy (EVT) has been proven to be beneficial for AIS patients with TL (5, 6). However, compared to patients with isolated stenosis, AIS patients with TL have a more complex pathological mechanism, greater difficulty in achieving vessel recanalization, and a worse prognosis. They also face an increased risk of intracranial hemorrhage (ICH) after EVT, along with higher rates of disability and mortality (3, 7–10).

ICH is a common complication following EVT that has been established as an independent risk factor affecting patient prognosis, and symptomatic intracranial hemorrhage (sICH) represents a more severe form (6, 10–12). Identifying risk factors influencing the occurrence of ICH and sICH after EVT in TL patients can aid clinicians in timely recognition and intervention, thereby controlling the progression of surgical outcomes. This can prevent situations where hemorrhage after EVT impacts antiplatelet therapy or necessitates surgical intervention due to sICH, both of which severely compromise patient outcomes (5, 9–11, 13, 14).

Currently, no predictive model exists for assessing the risk of ICH and sICH after EVT in TL patients. This study aims to construct a machine learning (ML) model to predict the risk of ICH and sICH following EVT in TL patients. We ultimately utilized a nomogram to intuitively display the relative importance of various independent variables in the model, facilitating the control of risk factors and improving patient prognosis.

Based on the inclusion and exclusion criteria, a total of 200 patients with TL involving anterior circulation large vessel occlusion were enrolled in this study between June 2020 and November 2024 (Figure 1). The cohort comprised 141 males (70.5%) and 59 females (29.5%). After EVT, ICH occurred in 92 patients (46%), and sICH occurred in 24 patients (12%). Patient characteristics, including baseline demographics, preoperative clinical data, and procedural details, are presented in Table 1. Correlations between the variables are illustrated in Figure 2.

LASSO regression was applied to select features influencing ICH after EVT in TL patients, reducing 19 initial features to 10 potential predictors (Supplementary Figures 1, 2). The results of the multivariate LR analysis for these predictors are presented in Table 2. Among them, diabetes, drinking history, and low mTICI score were significantly associated with an increased risk of ICH. Based on the LR model, a nomogram for ICH risk prediction was constructed (Figure 3). The final model demonstrated good fit (overdispersion test: p = 0.287) and low variable collinearity (VIF range: 1.077–1.734). The Hosmer-Lemeshow test for the model produced a p-value of 0.633. The model exhibited good risk assessment performance (AUC = 0.712, 95% CI: 0.641–0.784), with the ROC curve shown in Figure 4. The adjusted AUC via the bootstrap method was 0.654. Calibration curve and decision curve analysis for the ICH nomogram are presented in Supplementary Figures 3, 4.

LASSO regression identified 8 potential predictors influencing sICH after EVT in TL patients (Supplementary Figures 5, 6). The results of the LR analysis for these predictors are presented in Table 3. Among them, low GCS score, low mTICI score, and low ASPECTS were significantly associated with an increased risk of sICH. Based on the LR model, a nomogram for sICH risk prediction was constructed (Figure 5), showing good model fit (overdispersion test: p = 0.435) and low collinearity (VIF range: 1.022–1.725). The Hosmer-Lemeshow test for the model produced a p-value of 0.638. The model exhibited good risk assessment performance (AUC = 0.830, 95% CI: 0.741–0.919), with the ROC curve shown in Figure 6. The adjusted AUC was 0.773. Calibration curve and decision curve analysis for the sICH nomogram are presented in Supplementary Figures 7, 8.

SVM models were constructed using all feature variables. The SVM model predicting ICH risk after EVT in TL patients achieved an AUC of 0.861 (95% CI, 0.812–0.909), and its variable importance plot is shown in Supplementary Figure 9. The SVM model predicting sICH risk achieved an AUC of 0.688 (95% CI, 0.589–0.786), with the variable importance plot presented in Supplementary Figure 10. A summary table of the performance metrics for each model was provided in Supplementary Table 1.

In the LR model predicting hemorrhage risk after EVT in TL patients, diabetes, drinking history, and low mTICI score were significantly associated with an increased risk of ICH; low GCS score, low mTICI score, and low ASPECTS were significantly associated with an increased risk of sICH. The nomogram of the LR model converted multivariate predictors into a visual scoring system, quantifying variable contributions and linking total scores to outcome probabilities. This tool democratizes predictive analytics, supporting real-time, evidence-based decisions during consultations (18, 23, 24).

According to our data characteristics (Table 1), the median baseline ASPECTS for TL patients was seven, lower than previously reported for general AIS populations. The incidence of sICH after EVT in TL patients was 14%, higher than reported in several studies focusing on general AIS patients (15, 18, 22, 25, 26). Furthermore, baseline ASPECTS was a significant risk factor for sICH occurrence. This validates the perspective that patients with TL present with more severe disease and carry a higher risk of sICH complications compared to general AIS patients (3, 7–9). The inherent complexity of TL may compromise reperfusion success, as evidenced by a low mTICI score observed during the procedure (3, 8, 9). Our study also found that low mTICI score was a significant risk factor for both ICH and sICH occurrence (p < 0.05). This further suggests that patients with TL may have a higher risk of hemorrhage after EVT (7, 8).

Similar to findings in general AIS populations, diabetes also increased the risk of ICH after EVT in TL patients. Hyperglycemia may contribute to hemorrhage through various biological mechanisms, including impaired cellular metabolism, disruption of vascular integrity, and increased blood–brain barrier (BBB) permeability (7, 16, 20, 22, 27–29). Our study did not find a significant association between hyperglycemia and sICH risk. However, previous studies have shown that hyperglycemia significantly increases the likelihood of sICH after reperfusion therapy and is associated with poor functional outcomes and mortality (16, 25–27). Strict pre-stroke glycemic control (HbA1c ≤ 7.0%) may benefit neurological recovery in patients undergoing EVT (30). Prior research indicates that AIS patients with persistent postoperative hyperglycemia (blood glucose levels >140 mg/dL) have a significantly increased risk of sICH and poor functional outcome after EVT. Another study demonstrated that maintaining fasting blood glucose levels below 11.5 mmol/L on postoperative day 1 reduces the incidence of poor prognosis (31, 32). Therefore, for diabetic patients with TL, early identification and control of perioperative blood glucose levels are crucial for improving outcomes after EVT.

This study proposes a novel finding that drinking history may be an independent risk factor for ICH after EVT in TL patients. Previous research in stroke patients has shown that excess alcohol consumption is an independent risk factor for spontaneous hemorrhagic transformation (33). A study in rats has indicated that blood alcohol concentrations >300 mg/dL can induce irreversible spasm and even rupture with hemorrhage in the cortical microvessels (34). Alcohol may promote hemorrhagic events through mechanisms such as elevated blood pressure, vascular wall damage, exacerbation of neuroinflammation, and BBB disruption (20, 29, 35, 36).

Previous research has established that in AIS patients undergoing EVT, the mTICI score correlates with both ICH and sICH, with low mTICI score identified as an independent predictor of hemorrhage (15, 25). Low mTICI score may exacerbate injury in the infarcted area, increase the risk of hemorrhagic transformation, and facilitate it by damaging the BBB and increasing vascular permeability (20, 29, 37). Operator caution, potentially driven by concerns that frequent device manipulation (e.g., multiple stent retriever passes) could disrupt the BBB, might contribute to suboptimal reperfusion (7, 15, 20, 22, 26, 28). However, this study did not find that procedural factors like the number of stent retriever passes or intraarterial thrombolysis significantly influenced hemorrhage risk. This suggests that proactive surgical intervention during EVT for TL patients may reduce postoperative hemorrhage risk and benefit patients.

In contrast to some prior studies, low GCS score emerged as a significant predictor of sICH after EVT in TL patients. Although NIHSS score was not significant in the multivariate LR (Table 3), it was retained in the nomogram (Figure 5) after selection by LASSO regression. NIHSS score can complement GCS in predicting outcomes and identifying hemorrhagic transformation (26, 38, 39). The lack of significance for NIHSS in the LR analysis might be related to its strong negative correlation with GCS score (Figure 2). Patients with low GCS scores often present with high NIHSS scores, and this high correlation may have suppressed the independent effect of NIHSS in the multivariate model, preventing it from reaching statistical significance (23, 38, 40). Therefore, while GCS demonstrates a robust association with poor outcomes, incorporating variables such as NIHSS from the nomogram into a comprehensive analysis enhances the reliability of clinical decision-making.

Similar to findings in general AIS populations, this study found that low ASPECTS also increases the risk of sICH after EVT in TL patients (16, 17, 22, 28). ASPECTS serves as a surrogate marker for infarct volume; a lower ASPECTS value indicates greater irreversible ischemic brain tissue damage and is associated with an increased risk of hemorrhagic transformation through mechanisms involving vascular injury, BBB disruption, and inflammatory responses (16, 20, 29, 41). A multicenter retrospective study found that the potential benefit of functional independence following EVT in some TL patients with ASPECTS of 0–5 was non-negligible (17). Likely due to this phenomenon, operators at our center, after thorough informed consent, did not decline EVT for these patients. This underscores the clinical relevance of our study: when encountering a single high-risk factor for sICH, a comprehensive assessment incorporating other factors from the nomogram enables personalized prognosis evaluation, guiding clinical decision-making.

The application of ML algorithms for variable selection and nomogram construction to predict hemorrhage risk after EVT in TL patients represents a novel approach. In the variable importance plots of the SVM models, the top nine variables associated with ICH and the top seven variables associated with sICH all belonged to the key variables identified via LASSO regression (Supplementary Figures 9, 10), further validating the reliability of the LASSO variable selection. The constructed nomograms demonstrated good discriminative ability, providing TL patients with relatively accurate predictive tools for assessing postoperative ICH risk.

We acknowledge several limitations in this study. Firstly, the study was retrospective. Despite the implementation of strict inclusion and exclusion criteria, it remained challenging to entirely eliminate biases among the outcomes. Due to the retrospective nature of the data, imaging data such as precise infarct core volume and relative cerebral blood flow ratios were unavailable; we utilized ASPECTS as a surrogate marker for infarct volume (16, 20). Additionally, we did not perform dynamic monitoring of perioperative blood pressure and blood glucose levels in TL patients (31, 32, 42). Secondly, the model was developed using data derived from Chinese patients. Its applicability to populations in other countries has not been validated, necessitating enhanced collaboration with other international stroke centers to improve the model’s generalizability (25, 43). Thirdly, although the creation of predictive models demands “big data,” there is currently no standardized criteria to determine an appropriate sample size. Increasing the sample size or applying data balancing techniques in the future may enhance the model’s predictive performance (18, 26, 44). Despite these limitations, the final predictive model from this study exhibited good performance and did not appear to be adversely affected by them.

This study demonstrates that in AIS patients with anterior circulation TL, diabetes, drinking history, and low mTICI score significantly increase the risk of ICH following EVT. Meanwhile, low GCS score, low mTICI score, and low ASPECTS significantly increase the risk of sICH. Furthermore, the nomograms constructed using ML models in this study quantify the contribution of variables and link total scores to outcome probabilities, thereby assisting clinicians in rapidly assessing hemorrhage risk for personalized prognosis evaluation and treatment guidance.