A multimodal machine learning model for predicting postoperative worsening of FOGQ in Parkinson's disease following STN-DBS.

Objective: To develop and validate a multimodal machine learning model to predict postoperative worsening of freezing of gait questionnaire (FOGQ) scores in patients with Parkinson's disease (PD) undergoing subthalamic nucleus deep brain stimulation (STN-DBS). Methods: This retrospective study analyzed data from 134 patients with PD who underwent bilateral STN-DBS. The model integrated four data modalities: clinical scale assessments, structural neuroimaging features derived from voxel-based morphometry (VBM), stereotactic electrode localization data via Lead-DBS analysis, and radiomics features extracted from preoperative MRI. Following standardization, feature selection was conducted using LASSO, Boruta and recursive feature elimination with cross-validation (RFECV) methods to identify the most relevant predictors. Multiple machine learning algorithms were evaluated. Model development and internal validation were conducted using a 5-fold nested cross-validation framework. Model performance was assessed using ROC curves, calibration curves, and decision curve analysis, and model interpretability was analyzed using SHAP values. Results: The LightGBM model achieved the highest AUC of 0.917 for predicting FOGQ deterioration. The analysis emphasized the importance of multimodal data integration, combining clinical, structural, and radiomic features to enhance predictive accuracy. Conclusion: This multimodal LightGBM model achieved robust discrimination between patients with and without postoperative FOGQ deterioration, highlighting the value of integrating clinical, structural, and radiomic features for preoperative risk stratification in PD patients undergoing STN-DBS. These findings may inform personalized patient selection, early identification of high-risk individuals, and treatment planning, though external validation in prospective multicenter cohorts remains a necessary next step.