Explainable SHAP- XGBoost with DAT and clinical data for freezing of gait detection in Parkinson disease

Freezing of gait (FOG) is a common and debilitating symptom in Parkinson's disease (PD) that requires early detection for timely intervention. In this study, we developed an explainable SHAP-XGBoost model integrating clinical assessments and dopamine transporter (DAT) imaging to identify L-dopa responsive FOG. The internal cohort included 516 participants, with the model trained on a subset and validated on both internal and external test sets (Parkinson's Progression Markers Initiative, PPMI). The model demonstrated strong predictive performance, achieving AUCs of 0.90, 0.89, and 0.75 on the internal training, internal test, and external PPMI sets, respectively. SHAP analysis revealed that Hoehn & Yahr (H&Y) staging and DAT availability in the contralateral anterior putamen were the most influential features. Threshold analyses identified key cutoffs: around 5 years for disease duration, 64 years for age, and 35.7 for MDS-UPDRS Part III score. Notably, among patients with milder motor symptoms (H&Y ≤ 2.5), a contralateral anterior putamen SBR below 1.25 was associated with a higher FOG risk compared to those with H&Y > 2.5. In summary, our explainable model effectively detects L-dopa responsive FOG by leveraging clinical and DAT imaging data, emphasizing the contralateral anterior putamen as a critical region in FOG pathophysiology.