Functional Connectivity and Volumetrics Improve Outcome Prediction for Deep Brain Stimulation in Parkinson's Disease

Abstract

Background: Deep brain stimulation (DBS) targeting the subthalamic nucleus (STN) can effectively treat motor symptoms of Parkinson's disease (PD). However, optimal patient selection remains challenging due to the inadequacy of outcome predictors. Most clinicians rely on levodopa response to predict DBS motor outcomes, yet previous studies have identified other MRI-based predictors including resting-state functional connectivity (FC) and volumetric measures.

Objectives: To compare the predictive value of functional and volumetric MRI data with levodopa response alone for motor outcomes in STN DBS for PD.

Methods: We analyzed 65 participants who underwent STN DBS for PD at Washington University in St. Louis. We used relaxed LASSO regression to select factors including clinical, volumetric, and FC measures to generate predictive models for relative changes in motor scores post-DBS and assessed cross-validated performance. We then compared the relative influence of the predictive factors with standardized regression coefficients.

Results: Addition of MRI predictors (FC and volumetric) improved model fit and cross-validated model performance over levodopa response alone (levodopa alone: R² = 0.191, RMSE 13.6; MRI + levodopa: R² = 0.31, RMSE 12.6). Within the optimized model, aggregated FC and levodopa response exhibited the highest influence on motor outcome prediction.

Conclusions: Including MRI-based predictors significantly enhances the prediction of motor outcomes in STN DBS compared to levodopa response alone, with FC predictors demonstrating the greatest influence in the optimized model. External validation studies are necessary to confirm the clinical utility of these predictors in routine practice.

Keywords: Parkinson's disease; clinical prediction; deep brain stimulation; functional connectivity; volumetric MRI.

Figure 1

Model generation strategy. Predictor sets began with percent levodopa response, followed by adding one of three additional data types (expanded clinical, volumetrics, functional connectivity). All data types were combined to create the comprehensive predictor set. Model fitting was performed for each predictor set independently, using crossvalidated relaxed LASSO to select a sparse set of predictors. Leave‐one‐out crossvalidation was used to estimate out‐of‐sample performance on each model for comparison between predictor sets.

Figure 2

Comprehensive (optimized) model performance. Actual versus raw predicted percent motor response for (A) comprehensive model (R ² = 0.581, 95% CI 0.410–0.753) and (B) levodopa response model (R ² = 0.2335, 95% CI 0.0242–0.4428). Leave‐one‐out crossvalidation was then used to generate scatterplots of actual vs crossvalidated predictions for the (C) comprehensive model (R ² = 0.315) and (D) levodopa response model (R ² = 0.191). (E) Comparison of absolute value standardized regression coefficients for data types included in comprehensive model following factor selection.

Figure 3

Functional connectivity profile associated with DBS response, organized by canonical resting state network by seed pair. (A) Functional connectivity loadings by principal component (PC) included in comprehensive model following factor selection. (B) Sign‐adjusted weighted average (by β) of all included PCs in the comprehensive model, indicating overall connectivity pattern associated with DBS motor response.