An initial machine learning model applied to local field potential data from the subthalamic nucleus to detect...

Freezing of gait (FOG) is an unpredictable and debilitating symptom of Parkinson's disease (PD). Current approaches to treating FOG are lacking; adaptive deep brain stimulation (aDBS) holds promise in treating FOG. However, a necessary precursor to using aDBS is understanding changes in the subthalamic nucleus (STN) prior to and during FOG. The aim of this project was to develop a machine learning model to detect FOG episodes using local field potential (LFP) data from the STN in individuals with PD.

The LFP data were collected from five individuals with PD using the Medtronic Percept DBS platform while in the off-therapy state (off-DBS and off-medication) as they walked through virtual reality environments designed to induce FOG by manipulating levels of physical, anxiety, and cognitive load. The LFP time-series data were z-score normalized within each participant and canonical frequency band powers, rolling means, maxima, minima, and standard deviations were calculated. In addition, alpha-beta burst dynamics were calculated and included for analysis. The model was trained and subsequently tuned with a transfer learning component before testing its capacity to detect FOG episodes.

The deep learning model achieved an average weighted F1 score of 0.67, a macro F1 score of 0.62 across eight trials from five participants and successfully detected 24/29 (80%) of FOG episodes. The predictive importance of the twelve LFP channels varied across participants and FOG triggers, underscoring the need for an individualized, adaptable approach to modelling FOG using only LFP data.

Initial results indicate that FOG detection using LFP data gathered while walking is feasible. Future aDBS applications should contemplate a patient- and environmental-specific approach to optimize FOG treatment potential.

clinicaltrials.gov, Identifier (NCT05103384).