Magnetoencephalographic Spectral Abnormalities and Explainable Machine Learning of Neuropathic-Like Pain in Knee Osteoarthritis

Abstract

Objective

Neuropathic-like pain (NLP) in knee osteoarthritis (KOA) is associated with greater pain severity, poorer function, and less favorable outcomes, but its neurophysiological basis remains unclear. This study aimed to characterize magnetoencephalographic (MEG) spectral differences between NLP and nociceptive pain (NoCP) in KOA and evaluate whether interpretable machine-learning models could discriminate these pain phenotypes.

Methods

In this cross-sectional study, 192 patients with KOA were classified as having NLP (n=55) or NoCP (n=137) using modified painDETECT questionnaire scores. All participants underwent pain-evoked MEG recording. Forty-five spectral features were extracted using the Welch method, including theta-, alpha-, beta-, and gamma-band relative power and peak alpha frequency (PAF). Between-group differences were assessed using Mann–Whitney U-tests and multivariable linear regression adjusted for age, sex, Visual Analogue Scale score, and Western Ontario and McMaster Universities Osteoarthritis Index score. Eight machine-learning models were evaluated using 10 repeats of stratified 10-fold cross-validation. Model interpretability was assessed using Shapley additive explanations.

Results

Compared with NoCP, NLP was associated with higher Visual Analogue Scale and Western Ontario and McMaster Universities Osteoarthritis Index scores, while other clinical characteristics were similar. NLP showed increased theta- and gamma-band relative power, decreased alpha-band relative power, and slower PAF, mainly in central, frontocentral, and frontal regions. These patterns remained largely unchanged after adjustment. Logistic regression performed best, with an area under the receiver operating characteristic curve of 0.768 (95% CI, 0.745–0.789) and balanced accuracy of 0.699 (95% CI, 0.675–0.722).

Conclusion

Patients with NLP defined using the modified painDETECT questionnaire scores showed distinct MEG spectral alterations compared with those with NoCP. Interpretable MEG-based machine-learning models showed moderate ability to distinguish these questionnaire-defined pain phenotypes.

Plain Language Summary

Knee osteoarthritis often causes long-term pain, but not all people experience the same type of pain. Some people have symptoms that resemble nerve-related pain, such as burning, tingling, numbness, or electric shock-like sensations. This is often called neuropathic-like pain. People with this pain pattern may have more severe symptoms and may respond differently to treatment. However, this pain type is usually identified using questionnaires, and its brain-related features are still not fully understood. In this study, we examined brain activity in people with knee osteoarthritis using magnetoencephalography, a non-invasive method that records magnetic signals produced by brain activity. Participants were divided into two groups according to their questionnaire scores: those with neuropathic-like pain and those with nociceptive pain, which is pain mainly related to joint tissue damage. We compared brain rhythm patterns between the two groups and tested whether machine learning could help distinguish them. We found that people with neuropathic-like pain showed different brain rhythm patterns, especially in brain regions related to sensory processing and pain control. These differences included higher slow-frequency and high-frequency activity, lower alpha activity, and slower peak alpha frequency. A machine-learning model using these brain features showed moderate ability to distinguish the two pain groups. These findings suggest that neuropathic-like pain in knee osteoarthritis may involve changes in central pain processing. Brain activity measures may help support pain phenotype identification in future research, although larger studies are needed before clinical use.

Introduction

Knee osteoarthritis (KOA) is a major cause of chronic pain and functional limitation and represents a growing global public health burden, affecting approximately 374.7 million people worldwide in 2021, with prevalent cases more than doubling from 1990 to 2021.¹ Major risk factors include aging, female sex, obesity, previous knee injury, malalignment, muscle weakness, and excessive mechanical loading of the knee.^2,3 KOA is increasingly recognized as a whole-joint disease involving multiple articular and periarticular tissues, including cartilage degradation, subchondral bone remodeling, synovial inflammation, meniscal degeneration, infrapatellar fat pad inflammation, and biomechanical alterations.^4,5 For example, the infrapatellar fat pad is not merely a passive cushioning structure, but an active osteoarthritic joint tissue that can interact with cartilage, synovium, subchondral bone, menisci, ligaments, and nervous tissue through inflammatory cytokines, adipokines, and other molecular mediators, thereby contributing to local inflammation, joint degeneration, and pain.⁶ Pain is the predominant clinical symptom of KOA and a key determinant of healthcare utilization, treatment choice, and quality of life.⁷ Traditionally, KOA pain has been conceptualized primarily as nociceptive pain (NoCP), arising from peripheral nociceptive input associated with these joint-wide pathological changes. However, peripheral structural abnormalities alone do not adequately account for the clinical presentation of pain in KOA. Pain severity often shows only a weak or inconsistent relationship with imaging-defined joint damage,^8,9 and a subset of patients experience burning pain, electric shock-like sensations, numbness, or paresthesia, which resemble neuropathic pain symptoms.^10,11 Together, these findings indicate that KOA pain is mechanistically heterogeneous and may involve altered pain processing beyond peripheral nociceptive drive alone.

Neuropathic pain is defined by the International Association for the Study of Pain as pain caused by a lesion or disease of the somatosensory nervous system.¹² In KOA, although some patients present with neuropathic pain-like symptoms, objective evidence of a definite somatosensory lesion or disease is generally absent. For this reason, the term neuropathic-like pain (NLP) is considered more appropriate than true neuropathic pain in this population. Systematic reviews have suggested that approximately 20–40% of patients with KOA exhibit NLP features.^13,14 Importantly, this pain phenotype is not merely a descriptive symptom cluster, but appears to have substantial clinical relevance. Compared with patients without NLP, those with NLP generally report more severe pain,¹⁵ worse functional status,¹⁶ and a markedly increased risk of persistent pain after total knee arthroplasty.¹⁷ Accurate identification of NLP in KOA may therefore be important for risk stratification, treatment planning, and individualized pain management.

One potential mechanism underlying NLP in KOA is central sensitization, a state of enhanced responsiveness of the central nervous system to peripheral input that contributes to pain amplification, hyperalgesia, and allodynia.¹⁸ Consistent with this mechanism, previous electrophysiological studies have shown that chronic pain is accompanied by frequency-specific abnormalities in neural oscillations; however, the direction and frequency-band distribution of these abnormalities vary across studies,¹⁹ possibly reflecting heterogeneity across pain conditions, pain phenotypes, and recording paradigms. Increasing evidence suggests that central sensitization may play an important role in KOA patients with NLP.^20,21 Nevertheless, direct evidence characterizing central nervous system alterations associated with this pain phenotype remains limited. Neuroimaging offers a noninvasive means of probing central pain-related mechanisms. Functional magnetic resonance imaging (fMRI), for example, measures blood oxygen level-dependent signals and indirectly reflects neural activity through hemodynamic responses, enabling the assessment of regional brain activity and functional connectivity.²² Prior fMRI work has shown that, relative to patients without NLP, patients with KOA and NLP exhibit altered activity in regions involved in pain modulation, including reduced stimulus-evoked activity in the rostral anterior cingulate cortex and increased activity in the rostral ventromedial medulla.²³ These findings support the involvement of central pain-processing abnormalities in NLP, but they do not capture the fast electrophysiological dynamics that may underlie this phenotype.

Magnetoencephalography (MEG) directly records magnetic fields generated by synchronized neuronal activity and provides millisecond temporal resolution, making it particularly well suited for investigating neural oscillatory abnormalities across frequency bands.²⁴ Compared with fMRI, MEG offers a more direct measure of neuronal activity and may therefore provide complementary insight into the neurophysiological basis of pain phenotypes.²⁵ A pain-evoked MEG paradigm is particularly relevant in KOA because it can provoke typical KOA pain and allows neural activity to be assessed during a clinically symptomatic state. Using this approach, Quante et al investigated noxious counterirritation in patients with advanced KOA and provided evidence suggesting possible abnormalities in diffuse noxious inhibitory controls and descending pain modulation.²⁶ To date, however, no study has used pain-evoked MEG to characterize oscillatory alterations associated with NLP in KOA or to assess their value for distinguishing KOA pain phenotypes.

Systematic reviews of electroencephalography (EEG) and MEG studies further suggest that oscillatory features may provide candidate neurophysiological markers of chronic and neuropathic pain, although their specificity and clinical utility remain insufficiently established.²⁷ Because current identification of NLP relies largely on questionnaire-based symptom reporting, candidate objective neurophysiological markers are needed as complementary tools to support pain phenotype assessment in KOA. However, for such markers to be clinically informative, predictive performance alone is insufficient; model outputs must also be interpretable and neurophysiologically plausible. In this context, the present study aimed to investigate MEG spectral differences between NLP and NoCP in patients with KOA and to further assess the utility of MEG spectral features for distinguishing these pain phenotypes at the individual level. We first examined sensor-level spectral features and then developed interpretable machine-learning models combined with Shapley additive explanations (SHAP) to improve model interpretability. Based on the thalamocortical dysrhythmia (TCD) framework and prior evidence from other neuropathic pain conditions, we hypothesized that, compared with patients with NoCP, those with NLP would exhibit more pronounced oscillatory abnormalities in sensorimotor-related regions, particularly increased theta- and gamma-band activity together with reduced alpha-band activity.

Materials and Methods

Subjects

This was a single-center, cross-sectional observational study. From December 2024 to December 2025, patients with KOA were consecutively recruited from the Department of Orthopaedics, Qilu Hospital of Shandong University. Eligibility criteria included age 40–80 years, right-handedness, no severe cognitive impairment, no major neurological disorders, and no contraindications to MEG. Anxiety and depressive symptoms were screened using Hospital Anxiety and Depression Scale (HADS). Because HADS subscale scores of 0–7 are generally considered to fall within the normal range, whereas scores ≥8 indicate possible anxiety or depressive symptoms, only patients with both HADS-Anxiety and HADS-Depression subscale scores ≤7 were included to reduce the potential confounding effects of clinically relevant anxiety or depressive symptoms on pain perception and MEG activity.²⁸ All patients met the European League Against Rheumatism evidence-based recommendations for KOA diagnosis,²⁹ had knee pain for more than 6 months, had obtained insufficient pain relief from conservative treatment, including pharmacological treatment with analgesic and/or pain-modulating medications such as nonsteroidal anti-inflammatory drugs, and were scheduled for surgical treatment. Patients were excluded if they had other knee-related disorders, including rheumatoid arthritis, gouty arthritis, infectious arthritis, or knee involvement of ankylosing spondylitis, or a history of previous knee surgery.

All participants underwent demographic data collection, a semi-structured interview, and clinical evaluation. The collected clinical variables included disease duration, Kellgren–Lawrence (KL) grade,³⁰ visual analogue scale (VAS) score,³¹ modified painDETECT questionnaire (mPD-Q) score, Hospital for Special Surgery (HSS) knee score,³² and Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) score.³³ Participants were classified as having questionnaire-defined NLP or NoCP according to their mPD-Q scores, with an mPD-Q score greater than 13 indicating NLP and a score of 13 or lower indicating NoCP.³⁴ Informed consent was obtained from all participants before enrollment. The study protocol was approved by the Ethics Committee of Qilu Hospital of Shandong University (KYLL-2025SL-097).

MEG Recording

All MEG recordings were performed in the Department of Radiology, Qilu Hospital of Shandong University, using an optically pumped magnetometer-based MEG system (LMEG-64A; Lingci Medical Equipment Co., Ltd, Hangzhou, China). The system comprised an optically pumped magnetometer sensor array, a cylindrical magnetic shield, a nonmagnetic bed, a binocular camera, an internal monitoring camera for real-time observation, and a data acquisition unit (Figure 1A). The sensor array contained 64 channels (Figure 1B), operated at a sampling rate of 1000 Hz, and measured magnetic fields along the Z-axis perpendicular to the plane of the array. The binocular camera was used to monitor the relative position between the sensor array and the participant’s head, enabling sensor-to-head co-registration and confirming that head placement remained within an acceptable range throughout the recording.

Figure 1.

Optically pumped magnetometer-based magnetoencephalography system and schematic of the acquisition procedure. (A) Optically pumped magnetometer-based magnetoencephalography system. (B) Sensor layout map. (C) Schematic illustration of the data acquisition procedure.

Before data acquisition, all participants were instructed to remove all metallic objects and to abstain from caffeine and nicotine for at least 2 hours before the examination. During MEG recording, participants lay in a supine position with their eyes closed. The affected lower limb was elevated by supporting the heel, and a 2-kg sandbag was placed over the affected knee to maintain the joint in slight hyperextension (Figure 1C). This procedure, adapted from that reported by Quante et al, was used to evoke sustained and typical KOA pain.²⁶ Each recording lasted 2 minutes.³⁵ Throughout the acquisition, participants were instructed to remain relaxed, keep their eyes gently closed, avoid deliberate mental activity, and minimize unnecessary movements such as blinking or swallowing. They were also asked to remain awake throughout the recording to reduce nonspecific neural fluctuations. Signal quality and participant status were continuously monitored by online signal inspection in combination with the internal monitoring camera.

MEG Data Preprocessing and Feature Extraction

All MEG data preprocessing and feature extraction were performed in a Python 3.11 environment, and the Python libraries used, together with their corresponding versions, are listed in Table S1. The MEG signals were band-pass filtered between 2 and 40 Hz to remove low-frequency drift and high-frequency noise. To further suppress fixed-frequency interference, narrowband zero-phase notch filters were applied at 26.3 Hz and 31.3 Hz. Artifact-contaminated segments were automatically identified using a Z-score-based algorithm. Signal amplitudes were standardized relative to the mean and standard deviation of the corresponding recording, and the resulting Z-scores were used to quantify transient deviations from the overall signal distribution. Time segments with amplitude fluctuations exceeding an absolute Z-score threshold of 4 were marked as bad, and temporally adjacent abnormal samples were merged into continuous artifact-contaminated segments. The automatically marked segments were then manually inspected to verify the accuracy of the automatic annotations and to identify any residual artifacts missed by the algorithm. The final bad segments were excluded from subsequent spectral analyses, and only the remaining artifact-free data were used for power spectral density (PSD) estimation.

Independent component analysis (ICA) was performed using the FastICA fixed-point algorithm, as implemented in MNE-Python, to remove residual artifacts. As a widely used signal decomposition method, ICA separates mixed recordings into independent neural and non-neural sources.³⁶ In accordance with previous studies, the number of components was set to 20 to optimize separation between brain-related activity and artifact components, consistent with prior MEG ICA-based artifact-removal work.³⁷ Using ICA, physiological artifacts related to cardiac activity, eye movements and blinks, respiration, and muscle activity were explicitly corrected. Two raters independently reviewed and labeled the components, and any disagreements were resolved by a third rater. The denoised signals were subsequently re-evaluated using PSD analysis to confirm data stability. The cleaned recordings showed smoother spectral profiles without abnormally high-amplitude peaks. This multistep procedure ensured the integrity and reliability of MEG signal processing and provided a high-quality basis for subsequent analyses.

After preprocessing, spectral features were extracted from the MEG signals. PSD was estimated using the Welch method with a 5-second window length, 50% overlap, and a Hanning window.³⁸ To reduce the influence of inter-individual differences in absolute signal amplitude, normalized PSD values were calculated for the theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–40 Hz) bands. Peak alpha frequency (PAF) was quantified within the 8–13 Hz range using the center-of-gravity approach and was defined as the power-weighted mean frequency derived from the region-averaged Welch PSD.³⁹ MEG channels were then grouped according to anatomical regions (Table 1), including the occipital, parietal, central, centroparietal, frontal, frontal pole, temporal, and frontocentral regions. In total, 45 spectral features were extracted for each participant.

Table 1.

Correspondence Between Brain Regions and Magnetoencephalography Channels

Region	Channel
Occipital	O1, O2, PO3, PO4, PO7, PO8, Oz, Iz, POz
Parietal	P1, P2, P3, P4, P5, P6, P7, P8, P9, P10, Pz
Central	C1, C2, C3, C4, C5, C6, Cz
Centroparietal	CP1, CP2, CP3, CP4, CP5, CP6, CPz
Frontal	F1, F2, F3, F4, F5, F6, F7, F8, Fz
Frontal Pole	FP1, FP2, FPz, AF3, AF4, AF7, AF8, AFz
Temporal	T7, T8, FT7, FT8, TP7, TP8
Frontocentral	FC1, FC2, FC3, FC4, FC5, FC6, FCz

Machine Learning Analysis

Eight machine learning algorithms were used to develop binary classification models: Adaptive Boosting (AdaBoost), Gaussian Naïve Bayes (GaussianNB), K-Nearest Neighbors (KNN), Light Gradient Boosting Machine (LightGBM), Logistic Regression, Random Forest, Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost). These algorithms were selected to cover a range of modeling assumptions and decision boundaries, allowing evaluation of the discriminative performance and robustness of MEG spectral feature-based classification across different algorithmic frameworks. All modeling procedures were performed in a Python 3.11 environment.

To reduce variability introduced by random data partitioning and to improve reproducibility, 10 repeats of stratified 10-fold cross-validation were used as the overall evaluation framework.⁴⁰ Stratified sampling was applied throughout to preserve class proportions across folds. In each outer split, the outer test fold was reserved exclusively for final model evaluation. Given the class imbalance, adaptive synthetic sampling (ADASYN) was applied only to the training data within each training set to improve minority-class representation and mitigate the influence of class imbalance on model training.⁴¹ ADASYN was configured with sampling_strategy=“auto” and random_state=42. The number of nearest neighbors used for synthetic sample generation was set to 3 when feasible and was dynamically reduced when necessary according to the number of minority-class samples available in the corresponding inner training fold. Within each outer training set, feature selection was performed using recursive feature elimination (RFE) to improve model performance and stability. RFE was applied only to the training data to avoid information leakage. Candidate features were ranked according to model-derived feature importance, defined as the absolute values of standardized coefficients for linear models and feature-importance scores for tree-based models.⁴² The number of retained features was not fixed a priori. Candidate feature subsets ranging from 1 to all 45 features were evaluated within an inner stratified 5-fold cross-validation framework, and the subset yielding the highest mean area under the receiver operating characteristic curve (AUC) was selected. Data standardization, RFE-based feature selection, and hyperparameter optimization were all conducted within this inner cross-validation framework. Model-specific hyperparameters were optimized using randomized search, with AUC as the scoring metric. The complete hyperparameter search spaces for all models are provided in Table S2. The optimal feature subset and hyperparameters identified in the inner loop were then used to refit the model on the corresponding outer training set, and predictions were subsequently generated for the held-out outer test set.

Model performance was evaluated using balanced accuracy (BAL-ACC), positive predictive value, negative predictive value, sensitivity, specificity, F1 score, Matthews correlation coefficient, Cohen’s kappa, log loss, and AUC. All metrics were estimated using 1000 bootstrap resamples, from which bias-corrected means and 95% confidence intervals (CI) were calculated. In addition, calibration curves and decision curve analysis were generated to further assess model calibration and potential clinical utility. The optimal model was selected based on an overall comparison of these performance metrics.

To improve model interpretability, SHAP was used to quantify the contribution of each input feature to model predictions.⁴³ SHAP values were computed using model-specific explainers according to the type of classifier. Tree-based models, including AdaBoost, Random Forest, XGBoost, and LightGBM, were interpreted using TreeExplainer; Logistic Regression was interpreted using LinearExplainer; and model-agnostic classifiers, including SVM, KNN, and GaussianNB, were interpreted using KernelExplainer. For each model, the corresponding outer-training data were used as the background reference data, and SHAP values were calculated for the held-out outer-test samples. Based on the SHAP analysis, model outputs were interpreted at both the global and individual levels. At the global level, the most important features contributing to classification were identified and their relative importance was assessed. At the individual level, the direction and magnitude of each feature’s contribution to a given participant’s prediction were examined, thereby illustrating participant-specific MEG characteristics and their influence on model decisions.

Statistical Analysis

Statistical analyses were performed using SPSS version 25.0 (IBM Corp., Armonk, NY, USA) and Python 3.11, and the Python libraries used, together with their corresponding versions, are listed in Table S1. Given the unequal group sizes between the NLP and NoCP groups, a post hoc sensitivity power analysis was performed based on the final sample size. The analysis assumed a two-sided α level of 0.05 and 80% statistical power to estimate the minimum detectable between-group effect size. For baseline characteristics, categorical variables were compared between groups using the chi-square test, whereas continuous variables were compared using the Mann–Whitney U-test. For MEG spectral features, between-group differences in PSD and PAF across brain regions and frequency bands were assessed using the Mann–Whitney U-test. Multivariable linear regression models were then constructed with group, age, sex, VAS, and WOMAC score as independent variables to further examine the association between group status and each spectral feature after adjustment for demographic characteristics, pain intensity, and overall knee-related symptom burden.⁴⁴ Multiple comparisons were corrected using the Benjamini–Hochberg procedure to control the false discovery rate (FDR). All statistical tests were two-sided, and P < 0.05 was considered statistically significant.

Results

Demographic and Clinical Characteristics

A total of 194 patients with KOA were enrolled, of whom 2 were excluded because of excessive signal interference. The final analysis included 192 patients, comprising 55 in the NLP group and 137 in the NoCP group according to mPD-Q scores. Based on these group sizes, the post hoc sensitivity power analysis showed that the study had 80% power to detect a between-group effect of approximately Cohen’s d = 0.45 or larger at a two-sided α level of 0.05. Therefore, smaller between-group effects should be interpreted cautiously. Compared with the NoCP group, the NLP group had significantly higher WOMAC scores (31.78±6.11 vs 28.99±7.45, P=0.003) and VAS scores (6.55±1.70 vs 5.84±1.78, P=0.015). Age, sex, KL grade, disease duration, and HSS score did not differ significantly between groups (P>0.05). The demographic and clinical characteristics of the study population are presented in Table 2.

Table 2.

Demographic and Clinical Characteristics of the NLP and NoCP Groups

Characteristic	NLP (n=55)	NoCP (n=137)	P
Age (years)	65.25±6.68	63.88±7.70	0.227
Sex			0.149
Male	13/55(23.6%)	47/137(34.3%)
Female	42/55(76.4%)	90/137(65.7%)
KL grade			0.069
Grade II	14/55(25.5%)	26/137(19.0%)
Grade III	18/55(32.7%)	70/137(51.1%)
Grade IV	23/55(41.8%)	41/137(29.9%)
Disease duration (years)	7.30±4.93	6.37±5.50	0.092
HSS	67.45±7.47	69.82±7.77	0.070
WOMAC	31.78±6.11	28.99±7.45	0.003
VAS	6.55±1.70	5.84±1.78	0.015
BMI (kg/m2)	25.42±3.20	25.80±3.30	0.331

Abbreviations: HSS, Hospital for Special Surgery; KL, Kellgren–Lawrence; NLP, neuropathic-like pain; NoCP, nociceptive pain; VAS, visual analogue scale; WOMAC, Western Ontario and McMaster Universities Osteoarthritis Index; BMI, body mass index.

MEG Power Spectral Alterations

Across the 45 spectral features, including relative power in four frequency bands and PAF measured over the whole brain and eight anatomical regions, the NLP group showed an overall pattern of higher theta- and gamma-band relative power, lower alpha-band relative power, and slower PAF than the NoCP group, whereas no clear between-group differences were observed in the beta band. After FDR correction, 7 features remained statistically significant. Specifically, theta-band relative power was higher in the NLP group in the central (0.220±0.043 vs 0.193±0.046, Mann–Whitney U=2599, r=0.310, FDR-P=0.024), frontocentral (0.223±0.040 vs 0.198±0.050, Mann–Whitney U=2629, r=0.302, FDR-P=0.024), and frontal (0.219±0.046 vs 0.195±0.056, Mann–Whitney U=2717, r=0.279, FDR-P=0.038) regions. Gamma-band relative power was also higher in the NLP group in the central (0.106±0.026 vs 0.093±0.031, Mann–Whitney U=2811, r=0.254, FDR-P=0.045) and frontocentral (0.118±0.027 vs 0.103±0.034, Mann–Whitney U=2836, r=0.247, FDR-P=0.048) regions. In addition, PAF was lower in the NLP group in the central (10.221±0.247 vs 10.349±0.293, Mann–Whitney U=4757, r=−0.263, FDR-P=0.040) and frontocentral (10.180±0.238 vs 10.294±0.297, Mann–Whitney U=4772, r=−0.267, FDR-P=0.040) regions (Figure 2 and Table S3).

Figure 2.

Regional distribution of relative power across frequency bands in the NLP and NoCP groups. ns: FDR-P>0.05; *FDR-P<0.05.

Abbreviations: FDR, false discovery rate; NLP, neuropathic-like pain; NoCP, nociceptive pain.

Covariate-Adjusted Between-Group Differences in MEG Spectral Features

After adjustment for age, sex, VAS score, and WOMAC in multivariable linear regression models, between-group differences in MEG spectral features remained evident. Relative to the NoCP group, the NLP group continued to show an overall pattern of increased theta- and gamma-band relative power, decreased alpha-band relative power, and slower PAF, whereas no obvious between-group differences were observed in the beta band. After FDR correction, 6 features remained significantly associated with group status. These included higher theta-band relative power in the central (B=0.031, 95% CI=0.016–0.046, FDR-P=0.002), frontocentral (B=0.030, 95% CI=0.015–0.046, FDR-P=0.003), frontal (B=0.029, 95% CI=0.012–0.047, FDR-P=0.017), and centroparietal (B=0.025, 95% CI=0.009–0.041, FDR-P=0.031) regions; higher gamma-band relative power in the frontocentral (B=0.015, 95% CI=0.005–0.026, FDR-P=0.045); and lower PAF in the central (B=−0.126, 95% CI=−0.216 - −0.036, FDR-P=0.046) regions (Table S4).

Machine Learning Performance and Model Interpretability

The eight machine learning classifiers were trained and tested under the same repeated stratified 10-fold cross-validation scheme, with RFE embedded within the training process of each split. In each iteration, the feature subset selected from the corresponding training data was used to fit the model, and performance was then assessed on the held-out test fold. The 10 features with the highest selection frequencies for each model are shown in Figure 3 and Table S5.

Figure 3.

Feature-selection patterns of eight machine-learning models using recursive feature elimination. (A) AdaBoost. (B) GaussianNB. (C) KNN. (D) LightGBM. (E) Logistic Regression. (F) Random Forest. (G) SVM. (H) XGBoost.

Abbreviations: Ada Boost, Adaptive Boosting; Gaussian NB, Gaussian Naïve Bayes; KNN, K-Nearest Neighbors; LightGBM, Light Gradient Boosting Machine; SVM, Support Vector Machine; XGBoost, Extreme Gradient Boosting.

Of the eight models, Logistic Regression achieved the best performance, with an AUC of 0.768 (95% CI: 0.745–0.789) and a BAL-ACC of 0.699 (95% CI: 0.675–0.722) (Figure 4 and Table S6). Separate models were also developed for each individual MEG spectral feature category. Among these single-category models, theta-band features yielded the best classification performance, followed by alpha-band features, whereas models based on PAF, gamma-band, and beta-band features showed comparatively limited discriminative ability (Table 3, Figures S1–S5 and Tables S7–S11).

Figure 4.

Predictive performance of eight machine learning models. (A) Receiver operating characteristic curve. (B) Precision-recall curve. (C) Calibration curve. (D) Decision curve analysis. (E) Multi-metric radar plot.

Abbreviations: Ada Boost, Adaptive Boosting; Gaussian NB, Gaussian Naïve Bayes; AUC, area under the receiver operating characteristic curve; BAL-ACC, balanced accuracy; KNN, K-Nearest Neighbors; LightGBM, Light Gradient Boosting Machine; MCC, Matthews correlation coefficient; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver operating characteristic; SVM, Support Vector Machine; XGBoost, Extreme Gradient Boosting.

Table 3.

Predictive Performance of Machine Learning Models Based on All Features and Individual MEG Spectral Feature Categories

Feature Set	Best Model	AUC (95% CI)	BAL-ACC (95% CI)
All features	Logistic Regression	0.768 (0.745–0.789)	0.699 (0.675–0.722)
Theta	Logistic Regression	0.734 (0.710–0.758)	0.675 (0.652–0.698)
Alpha	Random Forest	0.654 (0.626–0.680)	0.616 (0.592–0.639)
Beta	Logistic Regression	0.610 (0.585–0.638)	0.592 (0.569–0.616)
Gamma	Logistic Regression	0.620 (0.595–0.646)	0.613 (0.588–0.638)
PAF	Logistic Regression	0.633 (0.606–0.662)	0.623 (0.600–0.648)

Abbreviations: AUC, Area Under the Receiver Operating Characteristic Curve; BAL-ACC, balanced accuracy; CI, confidence interval; MEG, magnetoencephalography; PAF, peak alpha frequency.

To improve interpretability of the optimal Logistic Regression model, SHAP analysis was performed to quantify the contribution of individual spectral features to model predictions. At the global level, mean absolute SHAP values identified Theta-Central, Alpha-Frontal Pole, Alpha-Frontal, Alpha-Centroparietal, and PAF-Frontal Pole as the five most important features (Figure 5A). The SHAP summary plot further illustrated the direction and magnitude of each feature’s contribution across participants; positive SHAP values increased the model output toward NLP, whereas negative SHAP values increased the model output toward NoCP (Figure 5B). The dependence plots showed how changes in the values of these key features were related to their SHAP contributions, thereby providing additional information on feature-specific prediction patterns (Figure 5C). At the individual level, the SHAP waterfall plot for a representative sample illustrated how individual feature contributions shifted the model output from the baseline value to the final prediction, suggesting that classification was driven primarily by the combined effects of multiple spectral features rather than by any single feature alone (Figure 5D).

Figure 5.

SHAP-based global and individual interpretation of MEG spectral features for distinguishing NLP from NoCP. (A) SHAP summary bar plot, showing global feature importance ranked by mean absolute SHAP value. (B) SHAP summary dot plot, showing the direction and magnitude of feature contributions; positive SHAP values indicate contributions toward NLP prediction, whereas negative SHAP values indicate contributions toward NoCP prediction. Dot color represents feature value, with red indicating higher values and blue indicating lower values. (C) SHAP dependence plots, showing the relationship between feature values and their SHAP contributions for the top five features; red and blue dots indicate participants with NLP and NoCP, respectively. (D) SHAP waterfall plot, showing how individual feature contributions shift the model output from the baseline value E[f(x)] to the final prediction f(x); red bars increase the model output, whereas blue bars decrease it.

Abbreviations: MEG, magnetoencephalography; NLP, neuropathic-like pain; NoCP, nociceptive pain; SHAP, SHapley Additive exPlanations.

Discussion

This study used optically pumped magnetometer-based MEG to compare neural oscillatory characteristics between NLP and NoCP in patients with KOA and further evaluated the potential value of these spectral features for identifying pain phenotypes in KOA. Overall, patients with NLP showed increased theta- and gamma-band relative power, decreased alpha-band relative power, and slower PAF, with the most prominent alterations observed in sensorimotor-related regions, including the central, frontocentral, frontal, and centroparietal areas. After adjustment for age, sex, VAS score, and WOMAC score, the overall pattern of between-group differences remained largely unchanged, suggesting that the observed spectral abnormalities were not fully explained by demographic characteristics, pain intensity, or overall knee-related symptom burden. In addition, machine-learning analyses showed that models based on these spectral features were able to distinguish NLP from NoCP to a certain extent, indicating that MEG spectral features not only captured neurophysiological differences between the two groups but also had potential value as an objective adjunct for pain phenotype identification in KOA.

The MEG abnormalities observed in the NLP group were mainly characterized by increased theta- and gamma-band relative power, together with decreased alpha-band relative power and slowed PAF. This pattern is broadly consistent with the TCD framework. According to this model, abnormal sensory input or partial deafferentation may lead to slowing of the dominant cortical rhythm, manifested as enhanced low-frequency activity together with abnormal increases in adjacent high-frequency activity.^45,46 Within this framework, the higher theta power observed in the NLP group may reflect more pronounced abnormal low-frequency synchronization, whereas the increase in gamma power may indicate rhythm-slowing-related cortical disinhibition and aberrant enhancement of adjacent high-frequency activity. The slowing of PAF and reduction in alpha power further support the presence of more prominent cortical rhythmic slowing in the NLP group than in the NoCP group. This does not imply that NLP in KOA fully reproduces the oscillatory profile of classical neuropathic pain; rather, it suggests that similar abnormal neural dynamics may be present under pain-evoked conditions. Taken together, these findings indicate that, in addition to persistent peripheral nociceptive input, patients with NLP in KOA may exhibit more pronounced abnormalities in central pain processing.

Patients in the NLP group had more severe pain and poorer knee function, whereas no significant between-group differences were observed in radiographic severity or most general clinical characteristics. The higher VAS and WOMAC scores in the NLP group raised the possibility that the observed MEG differences might partly reflect greater pain intensity and overall symptom burden rather than the NLP phenotype itself. However, after adjustment for age, sex, VAS score, and WOMAC score, the overall pattern of between-group differences remained largely consistent. This finding suggests that the observed spectral abnormalities were not fully explained by differences in pain intensity or knee-related symptom burden and may be associated with the NLP phenotype. This pattern is broadly consistent with previous reports and suggests that the greater symptom burden in the NLP group cannot be explained simply by the extent of structural joint damage,^5,9,10,47 but is more likely to reflect heterogeneity in the underlying pain mechanisms. Compared with NoCP, NLP may involve a more pronounced process of sensitization and pain amplification triggered by persistent peripheral input. Sustained abnormal joint input may first promote peripheral sensitization, leading to increased responsiveness of peripheral nociceptors.⁴⁸ With continued persistence, this abnormal input may further contribute to central sensitization, characterized by enhanced central nervous system responsiveness, pain amplification, and maintenance of abnormal pain states.¹⁸ In turn, central sensitization may further amplify the perception of peripheral input, such that a similar degree of peripheral pathology gives rise to more intense pain and more complex symptom manifestations.⁴⁹ Within this framework, patients with NLP are more likely to exhibit neuropathic-like symptoms extending beyond the scope of purely nociceptive pain, such as burning pain, electric shock-like sensations, numbness, and paresthesia.^10,11 The more pronounced spectral abnormalities observed in the NLP group may therefore reflect greater central pain amplification and aberrant pain processing arising from the interaction between persistent peripheral nociceptive input and maladaptive central sensitization.

From a spatial perspective, the spectral differences between the NLP and NoCP groups were concentrated mainly in the central, frontocentral, and frontal regions. The central and frontocentral regions are closely involved in sensorimotor integration and processing of knee-related somatosensory input,⁵⁰ whereas the frontal region is more strongly associated with cognitive appraisal of pain, attentional allocation, and descending modulation.⁵¹ This spatial distribution suggests that the differences between NLP and NoCP may not be limited to the intensity of peripheral sensory input itself, but may also involve abnormal integration between sensory processing and pain modulatory systems. Such an interpretation is broadly consistent with previous fMRI findings in KOA. Patients with NLP have been reported to show reduced activity in the rostral anterior cingulate cortex and increased activity in the rostral ventromedial medulla relative to those with NoCP, suggesting weaker descending inhibition and greater pronociceptive facilitation.²³ Similarly, fMRI studies of neuropathic pain in conditions other than KOA have shown that related abnormalities are often not confined to a single pain-transmission pathway, but instead involve both sensorimotor networks and frontal-cingulate regulatory systems.^52,53 Taken together, the present findings and previous evidence suggest that the neurophysiological abnormalities associated with NLP in KOA are more likely to reflect conjoint dysfunction of sensory integration and pain-related cortical networks than a localized alteration in a single nociceptive pathway.

NLP in KOA is clinically relevant not only because it is associated with greater pain severity and poorer functional status,^15,16 but also because it may indicate greater treatment complexity and a less favorable prognosis. Previous studies have shown that patients with NLP may respond less well to conventional conservative treatment and are at increased risk of persistent postoperative pain.⁵⁴ This is further supported by recent evidence showing a significantly elevated risk of chronic pain after total knee arthroplasty in patients with preoperative NLP features.¹⁷ In this context, treatment strategies targeting central mechanisms may be particularly important. For example, duloxetine may be beneficial in patients with end-stage KOA who exhibit neuropathic-like symptoms or are at risk of central sensitization,⁵⁵ and neuromodulatory approaches such as transcranial direct current stimulation have also been explored, although the available evidence remains mixed.^56,57 Equally important is the identification of NLP itself. At present, recognition of this phenotype still relies mainly on subjective questionnaires, such as mPD-Q, Douleur Neuropathique 4, and Leeds Assessment of Neuropathic Symptoms and Signs,⁵⁸ which, despite their clinical usefulness, remain susceptible to variation in symptom reporting, emotional state, and assessment context. Against this background, our machine-learning findings are potentially informative. The ability of MEG spectral feature-based models to distinguish NLP from NoCP to a certain extent suggests that the oscillatory abnormalities associated with NLP are not only evident at the group level but also carry some discriminative value at the individual level. Although the current model performance was moderate and independent external validation is still lacking, these findings suggest that MEG combined with machine learning may have potential as an objective complementary approach for identifying KOA patients with NLP features, with possible implications for preoperative risk stratification, treatment selection, and individualized pain management.

Several limitations of this study should be acknowledged. First, classification of NLP was based on mPD-Q scores. Although mPD-Q is widely used to identify neuropathic-like symptoms, it remains a questionnaire-based screening tool rather than an objective mechanistic reference standard. Accordingly, the phenotype identified in this study should be interpreted as a questionnaire-defined NLP subtype, and some degree of misclassification cannot be excluded. Second, the analysis was performed at the sensor level using region-based grouping. Although this approach is useful for characterizing the overall spatial distribution of spectral abnormalities, its spatial resolution is limited and does not allow precise localization of more fine-grained cortical or subcortical abnormalities. Third, although the machine-learning results suggest that MEG spectral features have some individual-level discriminative value, model performance remained moderate and lacked independent external validation. In addition, the moderate sample size and unequal group distribution may have contributed to residual model instability and overfitting risk, despite the use of stratified cross-validation, ADASYN oversampling, and nested feature-selection procedures. Therefore, the potential clinical utility of this approach requires further confirmation in larger, multicenter, prospective, and more balanced cohorts.

In summary, under pain-evoked conditions, patients with NLP in KOA exhibited MEG spectral differences distinct from those of patients with NoCP, mainly characterized by increased theta- and gamma-band relative power, accompanied by decreased alpha-band relative power and slower PAF. These differences were primarily distributed in the central, frontocentral, and frontal regions. Overall, the findings suggest that questionnaire-defined NLP in KOA may be associated with altered sensory-related and pain-related cortical processing, while the specific mechanisms linking these spectral differences to central sensitization or TCD require further investigation. In addition, machine-learning models based on these spectral features were able to distinguish NLP from NoCP to a certain extent, suggesting that the combination of MEG and machine learning may have exploratory value as a complementary research approach for characterizing pain phenotypes in KOA, although its clinical applicability requires further validation. These findings add preliminary neurophysiological evidence related to pain phenotype heterogeneity in KOA and may provide a basis for future work on validating objective markers and clarifying the underlying mechanisms of questionnaire-defined NLP.

Funding Statement

This study was supported by the Central Government Guided Local Science and Technology Development Fund Projects (2024ZY01035), the Key R&D Program of Shandong Province, China (2022ZLGX03), the National Natural Science Foundation of China (82472424), and the horizontal project of Shandong University (6010123041).

Data Sharing Statement

The data that support the findings of this study are not publicly available due to privacy and ethical restrictions. De-identified data are available from the corresponding author upon reasonable request.

Ethics Approval and Informed Consent

The study was conducted in accordance with the Helsinki Declaration, and all participants were provided with, and completed the informed consent approved by the Ethics Committee of Qilu Hospital of Shandong University (KYLL-2025SL-097).

Disclosure

The authors have no conflicts of interest to disclose for this work.