EEG functional connectivity features for predicting phantom limb pain severity in lower-limb amputees using machine learning

Phantom limb pain (PLP) is experienced by a significant number of people after lower-limb amputation. It is difficult to measure objectively as clinical evaluation continues to be mostly based on self-reported pain. Electroencephalography (EEG) functional connectivity (FC), which quantifies the temporal correlation between oscillatory signals in distributed cortical areas, has been suggested as a physiological correlate of chronic and phantom pain and machine learning (ML) provides a potential means to integrate several connectivity features into one predictor of the severity of pain. The aim of this study was to model the correlation between the resting-state EEG functional connectivity features with the PLP severity as measured by a self-reported questionnaire, using descriptive statistics, inferential statistics and supervised ML classification methods. The sample was described and the connectivity features were described descriptively. Associations between connectivity features and PLP severity were tested using Pearson correlations, independent-samples t-tests, one-way ANOVA, chi-square tests and multiple linear regression. The participants were classified into mild, moderate and severe PLP bands with a Random Forest and a Support Vector Machine. Random Forest and Support Vector Machine were trained to classify the participants into four different PLP bands. Reduced interhemispheric sensorimotor alpha connectivity (r = -0.70) and increased frontal theta connectivity (r = 0.82) exhibited the strongest correlations with PLP severity and a seven-feature regression model accounted for 85.2% of variance in NRS scores (R² = 0.852, F(7,88) = 72.61, p < 0.001) in the sample. The Random Forest classifier's hold-out accuracy was 70.8% (five-fold cross-validated mean accuracy of 74.9%).