Abstract Details

Title Automated Deep Brain Stimulation Programming Combining Electrophysiology and Artificial Intelligence

Topic Movement Disorders

Presentation(s) P1 - Poster Session 1 (8:00 AM-9:00 AM)

Poster/Presentation
Number 3-002

Objective To predict the best clinical contact in deep brain stimulation therapy using electrophysiology and artificial intelligence.

Background Deep brain stimulation (DBS) is an effective therapy for movement disorders such as Parkinson’s disease (PD). However, DBS programming is a manual and time consuming process done by clinical experts over many in-clinic visits. Modern DBS technology allows for local field potential (LFP) recording in implanted DBS systems. We sought to leverage this LFP data through machine-learning (ML) to develop automated algorithms that facilitate DBS programming workflow.

Design/Methods We conducted a retrospective study using the INFORM clinical research database at the University of Florida of PD patients implanted with Medtronic Percept PC neurostimulator. We selected patients who completed 4 months of DBS programming and had LFPs obtained on initial visit. We evaluated several ML-models designed to predict “dorsal” or “ventral” contact use at the 4th month programming visit based on LFPs obtained from initial visit.

Results For 76 patients who met inclusion criteria (9 STN, 66 GPi, 1 VIM), power spectral density (PSD) was calculated by applying short-time Fourier transform on their LFPs. The PSDs were separated into predefined frequency bands: alpha 8-12Hz, low-beta 13-20Hz, high-beta 21-30Hz, gamma 30Hz or greater. Average power in these bands were input features for the ML-models, which were evaluated across 5-fold cross validation testing design. Ensemble based Extra-Trees classifier, with mean (SD) area under receiver operating characteristic curve (AUC-ROC) of 67.5 (16.6) and mean (SD) F1-score of 73.9 (8.8) had best performance. Gradient boosting based CatBoost with mean (SD) AUC-ROC of 64.3 (15.4) and mean (SD) F1-score of 71.2 (7.9) had second best performance.

Conclusions We demonstrate that LFP data can inform contact selection for DBS programming. This sets the foundation towards an automated algorithm that can significantly decrease clinical burden and DBS programming time in patients with newly implanted DBS.

Authors/Disclosures
Venkat S. Lavu, MBBS (University of Arkansas Medical Sciences) PRESENTER	Dr. Lavu has nothing to disclose.
Jackson Cagle	Jackson Cagle has nothing to disclose.
Tiberio Correia de Araujo	No disclosure on file
Coralie De Hemptinne, PhD	The institution of Mrs. De Hemptinne has received research support from Alpha Omega. Mrs. De Hemptinne has received personal compensation in the range of $0-$499 for serving as a grant review with Michael J Fox foundation.
Joshua Wong, MD (University of Florida College of Medicine - Neurology)	The institution of Dr. Wong has received research support from NIH.

��ɫ��

��ɫ��