Classification of Pain Intensity using Functional Connectivity Networks Derived from Intracranial Electroencephalography in Humans

Abstract

Pain is an innate response most commonly arising from sensory and emotional stimuli reflecting injury or illness. Brain structures such as the anterior cingulate cortex, insula, and thalamus have been previously implicated in neural pain networks. However, the strength and nature of the relationships between pain structures have yet to be fully characterized. Intracranial electroencephalography (iEEG) is a method to record neural electrophysiology from cortical and subcortical regions of the brain, and its high spatial and temporal resolution has been used to identify neurophysiological networks associated with various conditions and cognitive states. Here, we collected pain reports using the 0 to 10 clinical visual analog scale and recorded multi-day iEEG in patients undergoing clinical seizure monitoring (n=5). For each pain report, 5-minute windows of iEEG from more than 87 channels were extracted, and neural features such as total power-in-band, correlation, and coherence were calculated and used as inputs in subject-specific pain classification models. Across subjects, multi-channel logistic regression models had a 76 + 9% accuracy in classifying binarized pain states (low vs. high pain) and multi-channel random forest models had 64 + 21% accuracy in classifying discrete VAS scores (0–10). The best-performing model had a 95% classification accuracy of discrete VAS scores. Our results indicate that subject-specific pain networks can be constructed from multi-day iEEG and be used to predict reported pain intensities. The machine learning paradigm described here can help inform closed-loop neuromodulation and pharmacological approaches for treating pain.