Perioperative stroke is a serious complication that can occur within 30 days after surgery, leading to significant neurological dysfunction, prolonged hospitalization, and increased mortality rates (Vlisides and Mashour, 2016). The incidence of perioperative stroke varies across different surgery types, ranging from 0.1% to 1.9% in non-cardiac surgeries to 1.9%–9.7% in cardiovascular surgeries (Ng et al., 2011; Marcucci et al., 2023). While several risk factors for perioperative stroke have been identified, traditional statistical-based prediction models struggle with nonlinearity and variable selection issues (Chahine et al., 2023). Recently, machine learning has emerged as a promising tool for predicting perioperative stroke in patients undergoing non-cardiac surgeries (Fernandes et al., 2021; Abraham et al., 2023).

This study aims to collect preoperative and intraoperative data, including relevant variables and demographics, from patients undergoing non-cardiac surgeries to develop an interpretable risk prediction model for perioperative stroke. The results will enhance physicians’ understanding of the mechanisms behind perioperative stroke, ultimately supporting informed decision-making and improving stroke prevention and prediction.

A total of 106,328 adult patients met the inclusion criteria during the study period, and after applying the exclusion criteria, the study cohort consisted of 70,718 patients. We analyzed data from 199 perioperative stroke patients, allocating 140 to the development cohort and 59 to the validation cohort. Tables 1,2 present descriptive statistical analyses: Table 1 compares patients with and without ischemic stroke, while Table 2 contrasts the development and validation cohorts.

To identify predictive factors and develop a corresponding model, we employed the LASSO logistic regression algorithm. Out of 179 factors, 31 with nonzero coefficients were selected based on an optimal lambda value of 0.0003226 (Figures 2A,B). LASSO regression enhances model interpretability by reducing the number of predictors, focusing on the most clinically relevant factors. Figure 3 and Table 3 display the correlation matrix, variance inflation factors (VIF), and tolerance values for each variable, respectively.

Ultimately, the final model included 9 significant factors: age (OR: 1.07, 95% CI: 1.06–1.09), drinking history (OR: 2.20, 95% CI: 1.44–3.37), diabetes (OR: 1.89, 95% CI: 1.26–2.84), angina pectoris (OR: 8.00, 95% CI: 2.80–22.87), history of cerebrovascular disease (OR: 3.60, 95% CI: 2.42–5.36), intraoperative MAP ≤60 mmHg for 5–10 min (OR: 3.12, 95% CI: 1.42–6.86), preoperative blood sodium level (OR: 0.95, 95% CI: 0.91–0.99), preoperative ACEI drugs (OR: 2.03, 95% CI: 1.31–3.14), perioperative non-steroidal drugs (OR: 4.35, 95% CI: 1.77–10.67) (Table 4).

After validating the development and validation cohorts, the model achieved an AUC of 0.864 (95% CI 0.839–0.890) in the development cohort and 0.869 (95% CI 0.827–0.910) in the validation cohort (Figures 4A,B). At the optimal threshold of 0.002, sensitivity and specificity were 0.771 and 0.807 in the development cohort, and 0.794 and 0.814 in the validation cohort. The corresponding F1 scores were 0.785 and 0.802, respectively, demonstrating the model’s ability to distinguish between stroke and non-stroke patients despite the imbalanced dataset.

Calibration analysis showed strong agreement between predicted and observed probabilities, with Brier scores of 0.002 in both cohorts. Hosmer–Lemeshow test results indicated good calibration (development cohort: χ² = 5.680, P = 0.771; validation cohort: χ² = 6.877, P = 0.650). Calibration plots further confirmed the model’s reliability in both cohorts (Figures 4C,D).

Decision curve analysis (DCA) was performed to assess the clinical utility of the predictive model. This method calculates the net benefit of applying the model across a range of threshold probabilities, comparing it to two baseline strategies: treating all patients or treating none. The threshold probabilities, ranging from 0% to 94% in the development cohort and 0%–92% in the validation cohort, represent the minimum predicted risk at which a clinical decision (e.g., initiating treatment) would be made. The model showed a positive net benefit within these ranges, supporting its utility for patients undergoing non-cardiac, non-vascular, and non-neurosurgical surgeries (Figures 5A,B).

Figure 6A presents the SHAP (SHapley Additive exPlanations) summary plot for Logistic regression model, where the X-axis represents SHAP values (higher values indicate stronger contributions to stroke prediction). Feature magnitude is represented by a color gradient, ranging from purple (high values) to yellow (low values).

The top four predictors of stroke risk were age, a history of cerebrovascular disease, diabetes, and the use of perioperative non-steroidal anti-inflammatory drugs (NSAIDs). Advanced age, a prior history of cerebrovascular disease, diabetes, and the perioperative use of NSAIDs were strongly associated with an elevated risk of stroke. Other significant factors included a history of alcohol consumption, abnormal serum sodium levels, low usage rates of ACE inhibitors, intraoperative hypotension (mean arterial pressure [MAP] ≤ 60 mmHg for 5–10 min), and a history of angina pectoris.

These findings were further supported by the feature importance rankings in Figure 6B. The interpretability of the model was also demonstrated through individualized case analyses. Figure 6C shows SHAP force plots for a representative non-stroke patient, highlighting feature-specific contributions to the prediction.

Our thorough validation process confirmed the model’s effectiveness in clinical settings. We developed and validated a logistic regression model based on patient characteristics and stroke risk factors, employing various graphical methods for evaluation.

Perioperative strokes in non-cardiac and non-neurological surgery patients are relatively rare, occurring in approximately 0.1%–0.8% of patients. Notably, the incidence rises to 7% among patients older than 65 years, with many strokes remaining undetected (Lindberg and Flexman, 2021). Although less frequently diagnosed following non-cardiac, non-carotid, or non-neurological surgeries, these strokes are significant postoperative complications, contributing to increased perioperative mortality, morbidity, longer hospital stays, and higher healthcare costs (Ng et al., 2011; Lewis et al., 2019).

Preventing perioperative strokes is challenging due to complex and diverse risk factors. Common comorbidities in these patients, such as atrial fibrillation, diabetes, and hypertension, are all significant stroke risk factors (Reinert et al., 2021). Additionally, perioperative stress responses, medication adjustments, and the surgery itself can further elevate stroke risk (Vlisides and Mashour, 2016).

Clinical prediction models are essential for accurately forecasting perioperative strokes. These models, which integrate factors such as age, medical history, and surgery type, are effective at identifying high-risk patients and outperform traditional assessment methods (Goeller and Bartels, 2021). Unlike traditional statistical models, machine learning models capture complex, nonlinear relationships between risk factors, improving prediction accuracy. This is especially useful in clinical practice, where risk factors often interact in ways that are difficult for traditional models to fully account for. Developing personalized treatment plans based on these models can enhance treatment efficacy and reduce unnecessary interventions. However, the effectiveness of these treatments depends on data quality and adaptation to various clinical scenarios (Woo et al., 2021).

Applying these models in resource-limited settings remains challenging, necessitating continuous improvements to ensure their relevance in clinical practice and bridge the gap between clinical and research settings (Goeller and Bartels, 2021; Woo et al., 2021).

This study focuses on ischemic stroke and its risk factors during the perioperative period. Through statistical analysis and comprehensive evaluation, the model’s reliability in predicting perioperative stroke risk in non-cardiac surgery patients was confirmed (Goeller and Bartels, 2021; Woo et al., 2021). The model can help clinicians develop early, tailored treatment strategies.

Our model incorporates nine predictive factors (Vlisides and Moore, 2021): age, drinking history, diabetes, angina pectoris, cerebrovascular disease history, intraoperative MAP ≤60 mmHg for 5–10 min, preoperative blood sodium level, preoperative ACE inhibitor use, and perioperative NSAID use. By incorporating clinically relevant variables and their coefficients, along with the SHAP (Shapley Additive Explanations) interpretability model, the model improves interpretability. This approach enables clinicians to assess the relative importance of each predictor in perioperative stroke risk. The inclusion of SHAP further facilitates informed decision-making by offering clear insights into the contribution of each factor, thereby identifying actionable variables for intervention.

Preoperatively, stroke risk is positively correlated with advanced age, alcohol consumption, diabetes, angina, and cerebrovascular disease history (Pai et al., 2017). Additionally, low serum sodium levels and cessation of ACE inhibitors during the perioperative period increase stroke risk (Roshanov et al., 2017; Khan et al., 2021). Intraoperatively, prolonged low MAP and NSAID use are key risk factors (Pai et al., 2017; Vlisides and Moore, 2021).

Wu and Fang (2020) developed a stroke prediction model for individuals over 60, using SMOTE technology to balance the data, identifying sex, LDL cholesterol, blood glucose, hypertension, and uric acid as key predictors. (Wu and Fang, 2020). Another study involving patients over 65 years old undergoing non-cardiac surgery used various machine learning techniques to evaluate factors such as age, chest pain history, heart failure symptoms, high-risk surgeries, intraoperative blood pressure, serum creatinine levels, left ventricular ejection fraction, and perioperative transfusions to predict adverse cardio-cerebral events (Wu et al., 2023). These findings highlight the value of machine learning in risk factor analysis.

Unlike previous studies that focused on older populations or specific patient groups, our study included adult patients of all ages undergoing non-cardiac, non-vascular, and non-neurosurgical procedures, improving the generalizability of the findings. By using advanced machine learning techniques, including comprehensive imputations and integrating logistic and LASSO regression, we addressed challenges such as nonlinearity and multicollinearity, resulting in a highly interpretable and clinically relevant model. These methods, supported by previous research comparing machine learning to traditional logistic regression in prognostic modeling (Liew et al., 2022), produced exceptional AUC values of 0.864 and 0.869 in the development and validation cohorts, respectively. Internal validation confirmed the model’s robustness and its potential to guide early interventions and improve perioperative stroke prevention strategies.

This study has several limitations. First, as a single-center investigation, the findings require external validation (Riley et al., 2021). Additionally, the retrospective design introduces potential biases, such as selection bias, and limits the ability to account for changes in clinical practices over time. To mitigate these limitations, we employed random sampling, robust statistical methods, and thorough model validation to minimize bias.

The model primarily focuses on preoperative and intraoperative variables, such as age, medical history, and surgical conditions. It is designed for perioperative risk stratification and guiding postoperative management rather than preoperative decision-making. While retrospective data lacks prospective validation, it offers an efficient approach for developing and validating predictive models using large, real-world datasets. These findings provide critical insights for optimizing perioperative management and lay the groundwork for future multi-center, prospective studies to integrate preoperative predictors and enhance clinical decision-making.

This study developed and validated a robust predictive model for perioperative stroke risk in non-cardiac, non-vascular, and non-neurosurgical patients. Despite its retrospective design, the model demonstrated strong discriminatory performance and clinical relevance. These findings provide a solid foundation for future multi-center, prospective studies to refine the model, incorporate additional variables, and expand its applicability to diverse patient populations.