Establishment of a machine learning prediction model for Wallerian degeneration after ischemic stroke

Wallerian degeneration (WD) is a common and clinically significant complication of ischemic stroke (IS). Due to the multifactorial and nonlinear characteristics of its underlying mechanisms, accurately identifying high-risk patients early remains challenging. This study aimed to develop and validate an interpretable machine learning (ML) model to predict WD after IS.

We retrospectively analyzed clinical data from 269 patients with IS, all admitted to the Xinhua Hospital of Dalian University. The patients were randomly divided into a training set (70%) and an internal validation set (30%). Thirty demographic, imaging, and laboratory variables were assessed, and predictive features were selected through Least Absolute Shrinkage and Selection Operator (LASSO) regression, followed by confirmation using multivariate logistic regression. Nine ML algorithms were constructed and compared. The best-performing model was interpreted using Shapley Additive Explanations (SHAP).

Among the 269 patients, 35.32% (95/269) of IS patients developed WD. LASSO regression selected eight candidate predictors (six reaching statistical significance in multivariate logistic regression; MCA and PCA retained based on LASSO selection and biological relevance): smoking history, hyperlipidemia, standard antiplatelet therapy, achieving LDL target with oral statins, maximum cross-sectional area of the stroke, middle cerebral artery (MCA), posterior cerebral artery (PCA), and the number of stroke-affected layers. Within the study population, the Random Forest model showed internally favorable predictive performance (training AUC = 0.946; validation AUC = 0.856) and reasonable internal consistency, surpassing AdaBoost, Logistic Regression, Lasso, Decision Tree, KNN, GaussianNB, XGBoost, and LightGBM. Through SHAP analysis, this study quantified and visualized the contribution of each predictive variable to the Random Forest model's prediction of the occurrence of WD, identifying key factors such as smoking history, MCA, and PCA, and revealing their interactions, thereby enhancing the model's interpretability for research purposes.

We developed and validated an interpretable Random Forest model with potential for predicting the occurrence of WD. By integrating demographic, imaging, and laboratory features, this model provides an internally validated framework that shows promise for early risk assessment, with potential to support personalized management pending external validation.