Supplementary code for DBS pointnet
<p dir="ltr">Personalizing deep brain stimulation (DBS) for Parkinson disease (PD) remains challenging due to trial-and-error programming and feature-engineered models that fail to capture critical 3D spatial-field interactions.Traditional volume of tissue activated (VTA) models focu...
Tallennettuna:
| Päätekijä: | |
|---|---|
| Julkaistu: |
2025
|
| Aiheet: | |
| Tagit: |
Lisää tagi
Ei tageja, Lisää ensimmäinen tagi!
|
| Yhteenveto: | <p dir="ltr">Personalizing deep brain stimulation (DBS) for Parkinson disease (PD) remains challenging due to trial-and-error programming and feature-engineered models that fail to capture critical 3D spatial-field interactions.Traditional volume of tissue activated (VTA) models focus solely on the spatial extent of activated tissue,losing continuous electric field distributions, gradients, and directional properties, which limits efficacyprediction accuracy. Here, we develop and validate PointNet++, a point cloud framework integratingsubmillimeter imaging, biophysical modeling, and artificial intelligence to overcome these limitations andenable precision prediction of DBS efficacy in PD. This retrospective multicenter cohort study included 561individuals with PD implanted with directional DBS leads, who underwent preoperative 3T MRI andpostoperative CT scans, with motor outcomes assessed via MDS-UPDRS-III in OFF-medication/ON-stimulation states across approximately 640 amplitude settings. The PointNet++ framework integrates patient-specific 3D point clouds (~227,500 points/subject) containing MRI-derived anatomy of deep-brain nuclei, high?resolution electric-field vectors from finite-element simulations, and electrode geometry with active contactlocations, processed via a hierarchical architecture to predict MDS-UPDRS-III motor score improvements.Results show that PointNet++ achieved 62.5% accuracy in predicting optimal stimulation parameters,representing a 37.5% relative improvement over conventional feature-based methods (support vector regressionand lasso regression: 25% accuracy each). Predicted MDS-UPDRS-III improvements demonstrated significantcorrelation with clinical outcomes, and the framework outperformed linear and traditional machine learningmodels (ridge regression and multilayer perceptron: 15% accuracy) by preserving high-resolution 3D field?tissue interactions. This approach enables precise, patient-specific prediction of DBS efficacy in PD byintegrating 3D point cloud imaging, biophysical modeling, and AI, shifting DBS programming frompostoperative trial-and-error to precision planning, potentially reducing treatment optimization time andfacilitating personalized selection of effective parameter combinations for PD management.</p> |
|---|