Artificial intelligence identifies spectral biomarkers for use in adaptive deep brain stimulation in Parkinson's disease
Lauren Hammer1, Carina Oehrn2, Clay Smyth3,4, Stephanie Cernera2, Ro'ee Gilron2, Simon Little1, Philip Starr2
1Neurology, 2Neurological Surgery, 3Bioengineering, University of California, San Francisco, 4Bioengineering, University of California, Berkeley
Objective:
To use machine learning with standardized neural signal recordings to identify patient-specific spectral biomarkers of parkinsonian clinical state during active deep brain stimulation (DBS).
Background:
Adaptive DBS (aDBS) is an investigational approach that modulates stimulation according to a patient’s clinical state (more/less parkinsonian), which is often estimated from spectral properties of neurophysiologic signals (“spectral biomarkers”). Most previous aDBS studies in Parkinson’s disease (PD) have used subthalamic or pallidal beta (13-30 Hz) spectral power with simple thresholding to estimate a patient’s clinical state. Prior studies have suggested machine learning may be useful in estimating clinical state, but were only studied using perioperative recordings with externalized leads and without active stimulation.
Design/Methods:
Three patients with PD who were implanted with investigational sensing neurostimulators connected to subcortical DBS leads and subdural paddle electrodes over primary sensorimotor cortex participated (with more being recruited). Subcortical local field potentials (LFPs) and electrocorticography were recorded during high- and low-levodopa states of a medication cycle, each at two stimulation settings: a high-amplitude setting optimized for the low-medication state, and a low-amplitude setting for the high-medication state. Spectral biomarkers most predictive of medication state (despite fluctuation in stimulation amplitude) were identified by partitioning the subcortical and cortical power spectra and using forward-feature selection with linear discriminant analysis.
Results:
The most discriminative biomarker for predicting clinical state varied across patients. For some patients, despite the presence of subcortical LFP beta peaks when OFF-stimulation, beta power was not the most discriminative biomarker when ON-stimulation. Use of two patient-specific frequency bands (instead of one) sometimes improved clinical state prediction accuracy.
Conclusions:
Subcortical beta power is not always the most discriminative biomarker of parkinsonian clinical state during DBS stimulation, even if beta peaks are seen during stimulation-OFF conditions. Machine learning may be useful to identify data-driven patient-specific frequency bands that better predict clinical state during active stimulation.