Chronic pain is associated with disrupted cortical activity, yet individual variability in these neural patterns remains poorly understood. Electroencephalography (EEG) provides a noninvasive means of characterizing these dynamics and may help identify patient subtypes relevant to spinal cord stimulation (SCS) outcomes. This study applied unsupervised machine learning to preoperative EEG to determine whether chronic pain patients exhibit distinct neurophysiological clusters and whether these groups differ in clinical characteristics or patient-reported outcomes (PROs). PROs included measures of pain intensity and related domains such as disability, catastrophizing, and mood, assessed using the Numerical Pain Rating Scale (NRS), Oswestry Disability Index (ODI), Pain Catastrophizing Scale (PCS), McGill Pain Questionnaire (MPQ), and Beck's Depression Inventory (BDI). Resting-state scalp EEGs were recorded from 16 patients scheduled for SCS implantation. After standard preprocessing, spectral features were extracted and a k-means clustering (K = 3) was applied to identify the structures within the EEG feature space. Three clusters emerged. Cluster 1 was characterized by globally reduced alpha and low-beta power along with lower theta power and entropy, suggesting a pattern of reduced oscillatory activity and spectral complexity within sensorimotor-related regions. In contrast, Cluster 2 showed elevated alpha and low-beta power, consistent with a distinct oscillatory profile that may reflect altered network dynamics. Cluster 3 exhibited moderate alpha and low-beta power alongside the highest theta entropy, indicating greater spectral complexity and variability in neural activity patterns, distinguishing this cluster from the other clusters and potentially relevant to pain processing. Demographic variables were similar across groups, but NRS "worst" scores differed significantly (p = 0.046). Feature-importance analysis identified peak low-beta and alpha power in primary somatosensory (S1), secondary somatosensory (S2), primary motor (M1), and parieto-occipital (PO) regions as the strongest contributors to cluster separation. Peak low-beta power across S1, S2, M1, and PO showed the most robust between-cluster differences (all p≤0.004), with peak alpha power in S1 and M1 also differing significantly (p≤0.008). Within clusters 2 and 3, multiple EEG features correlated significantly with postoperative improvements in PROs, suggesting that potential neural markers of symptom change. Cluster-specific EEG features were associated with postoperative improvement, supporting the potential utility of EEG for identifying chronic pain phenotypes and informing individualized neuromodulation approaches.
Dunn et al. (Fri,) studied this question.