Explainable machine learning on resting-state magnetoencephalography power spectra reveals neural alterations in knee osteoarthritis.

This study aimed to investigate abnormal patterns of resting-state magnetoencephalography (MEG) power spectra in patients with knee osteoarthritis (KOA) and to identify the spectral features most relevant to group classification using explainable machine learning approaches. A total of 80 patients with KOA who met the diagnostic criteria and 80 age- and sex-matched healthy controls were recruited. Resting-state MEG signals were recorded for all participants and preprocessed through filtering, artifact rejection, and independent component analysis. Power spectral density was estimated using the Welch method, and relative power in theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz), gamma (30-40 Hz) bands, and peak alpha frequency were extracted by brain region. In addition, a template-based source-level validation was performed using the FreeSurfer standard template brain. Compared with healthy controls, patients with KOA showed significantly reduced theta power in 3 regions and increased gamma power in 5 regions (FDR-P < 0.05). In the template MRI-based source-level validation, the overall direction of change was consistent with the sensor-level results. Pain intensity was negatively correlated with theta-band power. Among the 8 machine learning algorithms tested, the Logistic Regression model achieved the best classification performance (AUC = 0.803, ACC = 0.770). SHapley Additive exPlanations identified theta and gamma relative power in frontocentral region as most influential. The theta-band model alone yielded the highest single-band accuracy (AUC = 0.765). These findings reveal disrupted oscillatory dynamics in KOA, characterized by decreased theta and increased gamma power in frontocentral networks, suggesting central neural reorganization linked to pain.