A Combined‐Mode Machine Learning Model for Predicting Stroke Recurrence During Hospitalization in Patients with Acute Minor Ischemic Stroke

FIGURE 1. Study inclusion and exclusion flowchart. Abbreviations: CSCA, Chinese stroke center alliance; NIHSS, National Institutes of Health Stroke Scale. Created using draw.io‐desktop version 14.6.13.

FIGURE 2. The Shapley additive explanations values of the benchmark model (GLM) (A) and the best prediction model (LGB) (B) in individual‐mode in training set. Abbreviations: GLM, generalized linear model; LGB, light gradient boosting; NIHSS, National Institutes of Health Stroke Scale; INR, international normalized ratio; HCY, homocysteine; ICH, intracerebral hemorrhage; PVD, peripheral vascular disease; SHAP, Shapley additive explanations. Note: The horizontal axis illustrates the SHAP contribution values of each feature to the model output for each case, while the vertical axis signifies the features involved in the perdition model. The histogram below the horizontal line of each feature displays the distribution of SHAP values for individuals who did not experience stroke recurrence during hospitalization, and the histogram above the line represents those who did. Both histograms are presented on a logarithmic scale. Colors depict the feature values for each case.

FIGURE 3. Predictive performance evaluation for the developed GLM model, machine learning model and scoring scale: (A) the ROC curves for the GLM and LGB models in combined‐mode and RMS score in the test set; (B) the decision curves for the default strategies and for the GLM and LGB models before and after combinatorial optimization in the test set. Abbreviations: GLM, generalized linear model; LGB, light gradient boosting; RMS, in‐hospital stroke recurrence prediction scale for patients with acute minor ischemic stroke; AUC, area under ROC curve; ROC, receiver operation characteristic; CI, confidence interval.

FIGURE 4. Model performance of the GLM (A, B, and C) and LGB (D, E, and F) models after combinatorial optimization in the test set, on subgroups based on hospitalization duration (A and D), gender (B and E), and age (C and F). Legend: (specificity, sensitivity). Abbreviations: GLM, generalized linear model; LGB, light gradient boosting; AUC, area under ROC curve; ROC, receiver operation characteristic; CI, confidence interval.

FIGURE 5. Block diagram of model training and testing processes for both individual‐mode and combined‐mode. Abbreviations: HCI, previous history of cerebral ischemic stroke. Created using draw.io‐desktop version 14.6.13. †The threshold for predicting stroke recurrence during hospitalization in the test dataset is maintained consistent with that identified in the training dataset's ROC curve, which is determined by maximizing the Youden index.