A PointNet++-Based Deep Learning Approach Using Patient-Specific 3D Point Clouds for Personalized DBS Efficacy Prediction in Parkinson’s Disease

Abstract Personalizing deep brain stimulation (DBS) for Parkinson’s disease remains challenged by trial-and-error programming and feature-engineered models that fail to capture critical 3D spatial-field interactions, as traditional volume of tissue activated (VTA) models focus exclusively on the spatial extent of activated tissue and lose substantial information such as continuous electric field distributions, gradients, and directional properties. To overcome these limitations, we developed DeepPoint-DBS, a point cloud framework integrating submillimeter imaging, biophysical modeling, and AI for precision efficacy prediction. In a retrospective cohort study, we analyzed 561 individuals with Parkinson’s disease implanted with directional DBS leads (Medtronic 3389 and Pins L301), who underwent preoperative 3T MRI (T1/T2-weighted) and postoperative CT scans, with motor outcomes assessed via MDS-UPDRS-III in OFF-medication/ON-stimulation states across approximately 640 parameter combinations (amplitude: 1.0–4.0 V, 130 Hz fixed frequency). Patient-specific point clouds (~ 227,500 points/subject) integrated MRI-derived anatomy of deep-brain nuclei including STN and GPi, high-resolution electric-field vectors from finite-element simulations, and electrode geometry with active contact locations, processed via a hierarchical PointNet + + architecture to predict MDS-UPDRS-III improvements. DeepPoint-DBS achieved 62.5% accuracy in identifying optimal stimulation parameters, representing a 37.5% relative improvement over conventional feature-based approaches such as SVR and LR, with predicted MDS-UPDRS-III improvements demonstrating correlation with clinical outcomes. By preserving 3D field-tissue interactions at high resolution, this clinically deployable framework shifts DBS programming from postoperative trial-and-error to precision planning, potentially halving treatment optimization time and facilitating personalized selection of effective parameter combinations for Parkinson’s disease management.