Explainable machine learning model predicting neurological deterioration in Wilson's disease via MRI radiomics and clinical features.

Background: This study aims to build a machine learning (ML) model to predict the deterioration of neurological symptoms in Wilson's disease (WD) patients during short-term anti-copper therapy. The model combines brain T1WI MRI radiomics with clinical features and employs SHapley Additive exPlanations (SHAP) to interpret the contributions of the features.

Methods: Automated segmentation techniques were used to delineate regions of interest (ROI) in routine brain T1WI MRI scans from 107 WD cases. Radiomics features were extracted and screened using LASSO regression. Six ML models were trained to develop a predictive model for symptom deterioration during short-term anti-copper therapy. The SHAP method was applied to identify and explain the importance of the features in the ML models.

Results: Significant correlations were observed between UWDRS-N scores, Course of disease, age, and radiomics features of the brainstem, caudate, and corpus callosum (p < 0.05). The ML models incorporated 18 radiomics and 9 clinical features. Six ML models were used in the training and test set, the best performing model was XGBoost, with AUC values of 0.96 and 0.94, respectively. SHAP analysis revealed that the five most important features were the UWDRS-N score, age, right putamen GrayLevelNonUniformityNormalized, right caudate ZoneEntropy, and central corpus callosum SmallDependenceEmphasis. The SHAP force plot illustrated how the XGBoost model predicted neurological symptom deterioration at the patient level.

Conclusion: An explainable XGBoost model was successfully developed using brain T1WI MRI radiomics and clinical features. This model identifies WD patients at risk of neurological symptom deterioration during anti-copper therapy and provides insight into the contributions of individual features.