Despite the remarkable success of deep brain stimulation (DBS) in alleviating Parkinson's disease (PD) symptoms, complexities arising from inherent inter-individual variability and the vast array of available methodologies for functional brain imaging data processing and interpretation have resulted in substantial heterogeneity across published reports. Within this context, advanced modelling approaches offer a promising conceptual framework. However, the optimal criteria and methodological strategies yielding reliable outputs remain to be established. Leveraging a substantial dataset of 104 PD patients managed with subthalamic nucleus DBS, the present study applied nine machine learning algorithms to distinguish between DBS ON and OFF states. The input features were derived from global and local connectivity metrics and BOLD fluctuation amplitudes obtained from resting-state functional magnetic resonance imaging (fMRI) data. Model performance was evaluated using a 5-fold cross-validation with hyperparameter optimization, and the efficacy of various feature maps was systematically compared. The generalizability of classification models was further tested through validation in an independently acquired cohort of 34 additional PD patients. Global connectivity measures when combined with linear modelling approaches - namely logistic regression and linear discriminant analysis - or with support vector classifiers employing nonlinear kernels demonstrated superior classification performance. These models achieved area under receiver operating characteristic curve values of up to 0.82, with comparable performances observed within the validation cohort. Overall, this investigation not only identifies the most promising fMRI metrics and machine learning algorithms for future DBS-fMRI research but also reinforces the prevailing view of network-wide modulation standing at the core of DBS effects.
Kiakou et al. (Thu,) studied this question.