Diagnostic value of transcranial ultrasonography and clinical features for Parkinson's disease based on XGBoost model and SHAP visualization analysis: a retrospective study.

Abstract

Objectives: Parkinson's disease (PD) requires early diagnosis for optimal management. This study aims to evaluate whether combining transcranial ultrasonography (TCS) and clinical data using an interpretable machine learning model improves diagnostic accuracy.

Materials and methods: In this retrospective study, data from patients who underwent TCS between May 2023 and December 31, 2024, were retrospectively collected. Key clinical and TCS features were identified using the Boruta algorithm. An XGBoost model (an advanced gradient boosting algorithm) was developed based on these features, and Shapley Additive Explanations (SHAP, a method for interpreting machine learning predictions) was applied to visualize their contributions to PD diagnosis.

Results: The study included 599 patients (480 training, 119 validation) and achieved area under the curve (AUC) values of 0.863 and 0.811 in training and validation datasets, respectively. SHAP analysis revealed that bilateral substantia nigra hyperechoic (SNHA) and the substantia nigra/midbrain ratio (S/M) were the most influential predictors.

Conclusion: Integrating TCS with clinical data via XGBoost and SHAP provides high diagnostic performance and clear interpretability, supporting early PD diagnosis.

Key points: Question Can TCS features combined with machine learning provide reliable diagnostic support for PD? Findings XGBoost model integrating TCS and clinical features achieved a high diagnostic performance (AUC = 0.811) and interpretable outputs via SHAP visualization analysis. Clinical relevance This interpretable AI model supports early PD diagnosis and individualized decision making using non-invasive imaging and routine clinical parameters.