Three-Dimensional Point-Cloud Learning for Patient-Specific Deep Brain Stimulation Motor Response Prediction in...

Objective: To develop a patient-specific 3-dimensional (3D) point-cloud deep learning framework for predicting motor improvement after deep brain stimulation (DBS) in Parkinson's disease. Impact Statement: This study introduces an electric-field-aware point-cloud representation that preserves patient anatomy, electrode geometry, and local electric-field distributions for individualized DBS motor response prediction. Introduction: DBS response varies because of complex interactions among electrode location, anatomy, stimulation parameters, and electric field propagation. Existing volume-of-tissue-activated-, atlas-, sweet-spot-, stimulation-map-, connectomic-, and feature-engineered approaches have advanced outcome modeling, but many compress these interactions into thresholded volumes, voxel grids, or global summary features. Methods: We integrated preoperative MRI, postoperative computed tomography, electrode reconstruction, tissue conductivity, and finite element electric field simulations into patient-specific 3D point clouds. The screening database contained 561 stimulation records corresponding to 116 initially eligible patients. After exclusion of 37 patients with incomplete or unusable imaging, stimulation, outcome, registration, reconstruction, or point-cloud data, 79 independent patients formed the modeling cohort; one stimulation-outcome record was retained per patient. Patient-specific source point clouds were generated and subsequently downsampled for network processing. Results: On the original fixed test set, PointNet++ produced 15 of 24 predictions within ±5 Movement Disorder Society-sponsored revision of the Unified Parkinson's Disease Rating Scale, Part III (MDS-UPDRS-III) points (62.5%), while the voxelized convolutional neural network produced 8 of 24 (33.3%). Anatomical and stimulation features showed modest predictive value individually, whereas target- and parameter-response analyses revealed substantial interindividual heterogeneity. Given the small test set, these fixed split estimates do not support claims of generalization, superiority, or clinical utility. Conclusion: By preserving irregular electrode-tissue geometry and spatially resolved electric field information, the framework provides a spatially faithful basis for individualized DBS motor response prediction and future comparison of candidate stimulation settings.