Abstract
Background
This study introduces an interpretable machine learning model, derived from patient data, to address the notable lack of perioperative stroke prediction tools for adults undergoing noncardiac, nonvascular, and nonneurosurgical procedures, thereby improving clinical decision-making.
Methods
A retrospective cohort study encompassed 106,328 patients aged 18 years or older who underwent non-cardiac, non-vascular, and non-neurosurgical surgeries in our institution. The training cohort included 74,429 patients with 140 perioperative stroke incidents, and the validation cohort comprised 31,899 patients with 59 incidents. Risk factors for perioperative stroke were identified using univariable logistic regression analysis. The Least Absolute Shrinkage and Selection Operator (LASSO) regression method screened variables, followed by the development, validation, and performance evaluation of the prediction model through multivariate logistic regression analysis.
Results
The established prediction model, leveraging 16 variables including demographic information, medical history, and pre- and post-operative data, demonstrated robust discriminatory capability in forecasting perioperative stroke (AUC = 0.919; 95% CI, 0.896–0.942). It also showed an excellent fit with the validation cohort (Hosmer–Lemeshow test, χ²=4.085, P = 0.906). Decision curve analysis affirmed the model's substantial net benefit.
Conclusion
Through the analysis of patients aged 18 and above undergoing specified surgeries, this study successfully identified risk factors for perioperative stroke. Subsequently, it developed and validated effective prediction models that exhibit notable predictive accuracy, thereby serving as a pivotal tool for clinicians in decision-making processes. These insights lay the groundwork for the prevention and enhanced perioperative management of stroke, marking a significant stride in patient care optimization.