Development and validation of a machine learning model for predicting adverse prognosis in Wallerian degeneration
Zhiqi Yu1, Xujie Wang2, Ruiqi Tian3
1Department of Neurology, Xinhua Hospital Affiliated with Dalian University, Shahekou, Dalian, Liaoning, China.
Background:
Wallerian degeneration (WD) is a neurodegenerative change that often leads to irreversible neurological dysfunction in the central nervous system, resulting in poor prognosis. Currently, there is a lack of predictive tools capable of estimating the risk of poor prognosis in WD patients. This study aims to develop a machine learning-based predictive model to assess the risk of poor prognosis in WD patients, with the outcome measure of modified Rankin Scale (mRS) assessed at least 3 months after WD diagnosis by imaging.
Methods:
The data for this study were sourced from Xinhua Hospital, affiliated with Dalian University. Clinical patient data were randomly split into a training set and a validation set in a 7/3 ratio. Based on the AIC/BIC criteria, the Boruta algorithm and multivariable logistic regression were used for variable selection, and eight machine learning models were constructed. The models were evaluated for discrimination, predictive accuracy, and clinical benefit using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). Additionally, SHAP values were used to systematically evaluate feature importance in the best-performing machine learning model.
Results:
A total of 285 WD patients were included in the study, with a median age of 63 years. Among them, 157 (55.09%) were male and 128 (44.91%) were female, and 161 (56.49%) patients had poor prognosis. Ten candidate predictors for poor prognosis in WD were identified through analysis: diabetes, hypertension, hyperlipidemia, atrial fibrillation, NIHSS score, medulla oblongata, periventricular region, centrum semiovale, subcortical volume 1,000, and the severity of WD. Seven of the ten predictors showed statistical significance, while three demonstrated a borderline association. Among the evaluated machine learning models, the AdaBoost model (AUC = 0.880) demonstrated the most stable performance in terms of discrimination, calibration, and clinical benefit. SHAP analysis indicated that NIHSS score, atrial fibrillation, and hypertension made significant contributions to predicting poor prognosis in WD patients.
Conclusion:
Our study successfully developed and validated a machine learning predictive model that integrates clinical indicators and imaging data to estimate the risk of poor prognosis in WD patients. This model may assist clinicians in identifying high-risk individuals and provide evidence-based guidance for early intervention and personalized management.