パーキンソン病患者における深部脳刺激療法の転帰予測のためのビデオベース機械学習モデル

Current levodopa challenge test (LCT) for deep brain stimulation (DBS) candidate screening in Parkinson's disease (PD) relies on subjective clinical scales, limiting its predictive capacity for postoperative motor outcomes. We developed video-based machine learning models using quantified kinematic metrics during preoperative LCT in seventy PD patients who underwent DBS surgery. Objective multi-domain motor features were extracted via validated motor assessment software. Binary classification defined patients' outcomes as DBS+ (≥30% improvement in MDS-UPDRS Part III) or DBS- (<30%). Ternary classification further categorized outcomes as DBS + + (≥ 60%) and DBS+ - (30-60%). Results show: (1) For binary classification (DBS + /DBS - ), Linear Discriminant Analysis (LDA) achieved an F1 score of 0.87 (Receiver Operating Characteristic Area Under Curve (ROC AUC) = 0.77, accuracy = 0.8). (2) For ternary efficacy stratification, LDA attained a weighted F1 score of 0.67 (average ROC AUC = 0.67, accuracy = 0.67). (3) Models combining video-derived features with conventional clinical predictors significantly outperformed the baseline logistic regression model that included only conventional clinical predictors. (4) Clinical interpretation: Velocity-driven domains demonstrated key contributions in both binary and ternary outcome predictions, while amplitude- and stability-related metrics also played a supporting role. Axial parameter aided in identifying DBS responsiveness, and asymmetric levodopa response patterns were found to stratify efficacy tiers. Although linear models performed well, non-monotonic relationships between specific metrics and motor outcomes were identified. This analytical approach serves as a complementary tool for specialists, strengthening preoperative screening through objective motor-responsiveness profiles derived from LCT video, potentially promoting data-driven patient selection and personalized surgical consultation in the future.