Introduction Stroke remains a leading global cause of morbidity and mortality. Conventional tools, such as the Framingham Stroke Risk Score, offer limited flexibility in complex cases. We developed STRIKE ( STroke RIsK Estimator ), a machine learning‐based risk calculator using XGBoost (XGB), to predict 10‐year stroke risk from routinely collected clinical variables. We compared its performance to a simplified Framingham‐style score (age, systolic blood pressure, diabetes status, smoking), and externally validated the model on an independent U.S. cohort. Methods We extracted data from the National Health and Nutrition Examination Survey (NHANES) 2017‐2018, a survey research program conducted by the National Center for Health Statistics to assess the health and nutritional status of adults and children in the United States, and to track changes over time, regarding adults with complete information on age, sex, systolic and diastolic blood pressures, diabetes status, and smoking history. Stroke status was based on self‐report. A balanced dataset (n = 1,020) was constructed via random under‐sampling and split evenly into training and test sets. STRIKE was trained using six input features. Area Under the Curve (AUC) was computed for internal test performance and compared to a simplified Framingham‐style additive score (0‐4 points). For external validation, we used NHANES 2015‐2016 (n = 5,057). Figure 1 shows the STRIKE calculator interface. Results STRIKE achieved an AUC of 0.71 on the internal test set and 0.96 on training data. The simplified risk score yielded a comparable AUC of 0.73. External validation demonstrated an AUC of 0.79 (Figure 2). Conclusion STRIKE could serve as a practical and generalizable machine learning‐based stroke risk model that matches or exceeds traditional tools using simple clinical inputs. The external cohort confirmed strong discriminative performance. While retrospective and based on self‐reported outcomes, STRIKE shows promise for integration into clinical decision tools. image image
Salkhadi et al. (Sat,) studied this question.