Machine learning predicts spinal cord stimulation surgery outcomes and reveals novel neural markers for chronic pain

Abstract

Spinal cord stimulation (SCS) is a well-accepted therapy for refractory chronic pain. However, predicting responders remain a challenge due to a lack of objective pain biomarkers. The present study applies machine learning to predict which patients will respond to SCS based on intraoperative electroencephalogram (EEG) data and recognized outcome measures. The study included 20 chronic pain patients who were undergoing SCS surgery. During intraoperative monitoring, EEG signals were recorded under SCS OFF (baseline) and ON conditions, including tonic and high density (HD) stimulation. Once spectral EEG features were extracted during offline analysis, principal component analysis (PCA) and a recursive feature elimination approach were used for feature selection. A subset of EEG features, clinical characteristics of the patients and preoperative patient reported outcome measures (PROMs) were used to build a predictive model. Responders and nonresponders were grouped based on 50% reduction in 3-month postoperative Numeric Rating Scale (NRS) scores. The two groups had no statistically significant differences with respect to demographics (including age, diagnosis, and pain location) or PROMs, except for the postoperative NRS (worst pain: p = 0.028; average pain: p < 0.001) and Oswestry Disability Index scores (ODI, p = 0.030). Alpha-theta peak power ratio differed significantly between CP3-CP4 and T3-T4 (p = 0.019), with the lowest activity in CP3-CP4 during tonic stimulation. The decision tree model performed best, achieving 88.2% accuracy, an F1 score of 0.857, and an area under the curve (AUC) of the receiver operating characteristic (ROC) of 0.879. Our findings suggest that combination of subjective self-reports, intraoperatively obtained EEGs, and well-designed machine learning algorithms might be potentially used to distinguish responders and nonresponders. Machine and deep learning hold enormous potential to predict patient responses to SCS therapy resulting in refined patient selection and improved patient outcomes.

Keywords: Chronic pain; EEG; Machine learning; Responders; Spinal cord stimulation.

Fig. 1

Global activity. (A) Power spectrum density (PSD) estimates. PSDs were averaged across EEG regions and patients per group during stimulation OFF, tonic and HD. Blue: Responders. Red: Nonresponders. Shaded area indicating ± standard deviation. HD: high-density. (B) Global alpha band activity. Boxplots indicating global alpha activity (8–12 Hz) during baseline (top row), tonic stimulation (middle row), and HD stimulation (bottom row). Blue: Responders. Red: Nonresponders. Black lines indicating the groups’ mean.

Fig. 2

Comparison of alpha/theta peak power ratio between responders and nonresponders. (A) Topographical maps of alpha-theta peak power ratio in responders and nonresponders during stimulation OFF (top-row), tonic (middle-row) and HD (bottom-row). Color bar indicating alpha-theta peak power ratio in decibel (dB) scale. (B) Correlation analysis between alpha/theta peak power ratio in FP1-FP2 during baseline and preoperative Numeric Rating Scale (NRS) worst pain scores. Blue circle: Responders. Red diamond: Nonresponders. (C) Correlation analysis between alpha/theta peak power ratio in C3-C4 during baseline and preoperative Beck Depression Inventory (BDI) scores.

Fig. 3

Comparison of relative power between responders and nonresponders. (A) Topographical distribution of relative global (1–50 Hz) power in tonic (top-row) and HD (bottom-row). Color bar indicating relative power in decibel (dB) scale. (B) Correlation analysis between relative global power in CP3-CP4 during HD stimulation and preoperative Numeric Rating Scale (NRS) average pain scores. Blue circle: Responders. Red diamond: Nonresponders. (C) Correlation analysis between relative theta (4–7 Hz) power in CP3-CP4 during tonic stimulation and preoperative Oswestry Disability Index (ODI) scores.

Fig. 4

Subband peak frequencies between responders and nonresponders. (A) Heatmaps showing the subband peak frequency per SCS condition in each EEG region in responders (top-row) and nonresponders (bottom-row). Color bar indicating the frequency in Hz. (B) Correlation analysis between theta peak frequency in CP3-CP4 during HD stimulation and preoperative McGill Pain Questionnaire (MPQ) and Oswestry Disability Index (ODI) scores. (C) Correlation analysis between alpha peak frequency in CP3-CP4 and FP1-FP2 regions during HD stimulation and preoperative Numeric Rating Scale (NRS) scores. Blue circle: Responders. Red diamond: Nonresponders.

Fig. 5

Machine learning pipeline. (A) Diagram of full pipeline. EEG data was collected during SCS implant surgery and was subsequently exported to MATLAB for offline analysis. The data was preprocessed and the relevant neural features were extracted. Principle component analysis (PCA) was employed to rank these features by explained variance. The top features, along with demographics, clinical features, and pain scores, were used as inputs to a machine learning model that predicted responders. (B) Comparison of different numbers of EEG features as input to the decision tree classifier. Accuracy, F1 score, and area under the receiver operating characteristic curve (AUROC) are shown. The optimal number of EEG features was defined as that with the highest AUROC, indicated by a vertical arrow. (C) ROC for the optimal version of each of the architectures compared. AUROC is noted in the figure legend. (D) Confusion matrix for the optimal decision tree classifier.