Predictive modeling of sensory responses in deep brain stimulation

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Deep brain stimulation (DBS) typically aims to avoid stimulation-induced sensations, but these can be intentionally induced for applications like computer-brain interfaces.
• Selecting optimal DBS parameters to control these sensations is challenging due to the vast parameter space.

Purpose of the Study:

• To develop a machine learning model for predicting the occurrence and somatic location of paresthesias evoked by thalamic DBS.
• To streamline the selection of DBS parameters for both avoiding and inducing sensations.

Main Methods:

• Utilized a dataset of 3,359 paresthetic sensations from 18 thalamic DBS leads across 10 individuals.
• Modeled the Volume of Tissue Activation (VTA) for each stimulation.
• Trained a machine learning model using stimulation parameters and VTA data to predict sensation occurrence and location.

Main Results:

• The model demonstrated fair to substantial agreement (Kappa 0.31-0.72) with ground truth for predicting paresthesia presence and location.
• Prediction accuracy for the presence of paresthesias was similar for seen and unseen cases (Kappa 0.72 vs. 0.60).
• Prediction accuracy for specific somatic locations was lower for unseen cases (Kappa 0.53 vs. 0.31).

Conclusions:

• Machine learning holds potential for optimizing DBS parameter selection, improving postoperative management.
• Predictive models can guide clinical DBS programming and the tuning of DBS-based computer-brain interfaces.