Machine learning-based prediction of clinical pain using multimodal neuroimaging and autonomic metrics

Abstract

Although self-report pain ratings are the gold standard in clinical pain assessment, they are inherently subjective in nature and significantly influenced by multidimensional contextual variables. Although objective biomarkers for pain could substantially aid pain diagnosis and development of novel therapies, reliable markers for clinical pain have been elusive. In this study, individualized physical maneuvers were used to exacerbate clinical pain in patients with chronic low back pain (N = 53), thereby experimentally producing lower and higher pain states. Multivariate machine-learning models were then built from brain imaging (resting-state blood-oxygenation-level-dependent and arterial spin labeling functional imaging) and autonomic activity (heart rate variability) features to predict within-patient clinical pain intensity states (ie, lower vs higher pain) and were then applied to predict between-patient clinical pain ratings with independent training and testing data sets. Within-patient classification between lower and higher clinical pain intensity states showed best performance (accuracy = 92.45%, area under the curve = 0.97) when all 3 multimodal parameters were combined. Between-patient prediction of clinical pain intensity using independent training and testing data sets also demonstrated significant prediction across pain ratings using the combined model (Pearson's r = 0.63). Classification of increased pain was weighted by elevated cerebral blood flow in the thalamus, and prefrontal and posterior cingulate cortices, and increased primary somatosensory connectivity to frontoinsular cortex. Our machine-learning approach introduces a model with putative biomarkers for clinical pain and multiple clinical applications alongside self-report, from pain assessment in noncommunicative patients to identification of objective pain endophenotypes that can be used in future longitudinal research aimed at discovery of new approaches to combat chronic pain.

Conflict of interest statement

Competing interests: No conflict with respect to financial interests.

Figures

Figure 1.. Classification of within-patient clinical pain states using a support vector machine (SVM).

(A) Each patient provided rCBF voxels, S1CONN voxels, and HFHRV values as multimodal input features to the SVM algorithm. (B) The paired classifier discriminated between POST-PRE and PRE-POST in feature space by fitting a decision boundary that maximally separates them, for each modality. This SVM procedure produces (C) decision responses and weights for each modality. The decision responses for each modality were combined for a synergistic classifier of clinical pain states. The weights for each modality were used in the SVR analysis (see Figure 2). N.b. rCBF: regional cerebral blood flow, S1CONN: S1back-connectivity, HFHRV: high frequency heart rate variability power, PRE: pre-maneuver, POST: post-maneuver.

Figure 2.. Prediction of between-subject clinical pain intensity using support vector regression (SVR).

(A) Clinical pain ratings for each patient (N=53) were randomized, and equally allocated into TRAIN and TEST datasets (providing total effective N=106). This randomization into TRAIN and TEST equalized the range of pain ratings, previously found to be discrepant between PRE and POST. Corresponding multimodal features for each clinical pain intensity timepoint were taken, and (B) decision responses were calculated using a dot product between multimodal features and corresponding SVM classification group weights, resulting in decision responses for each modality. (C) The decision responses for TRAIN and corresponding clinical pain ratings were used to build an SVR model. The trained SVR model was then applied to decision responses from TEST, to produce an output of predicted pain ratings. The true (TEST) and predicted pain ratings were plotted and a Pearson’s correlation coefficient was computed to evaluate model performance. N.b. rCBF: regional cerebral blood flow, S1CONN: S1back-connectivity, HFHRV: high frequency heart rate variability power, PRE: pre-maneuver, POST: post-maneuver.

Figure 3.

Clinical pain intensity changes due to physical maneuvers. (A) Individually customized physical maneuvers significantly exacerbated low back pain levels in cLBP patients (N=53). (B) Patients reported a wide range of baseline low back pain levels (pre-maneuvers) and after maneuvers, all included patients reported increased pain levels (post-maneuvers), which were maintained throughout the duration of the post-maneuver scans. N.b. Bar plots in (A) show mean±SD. Each data point in (B) represents an individual patient.

Figure 4.

Brain features significantly contributing to within-patient classification of lower- and higher-clinical pain intensity states. Paired-SVM classification weight maps for rCBF (A) and S1CONN (B) were thresholded (permutation analysis with N=5000, P<0.01) for visualization. N.b. rCBF: regional cerebral blood flow, S1CONN: S1-connectivity, SMA: supplementary motor area, vPCC: ventral posterior cingulate cortex, dlPFC: dorsolateral prefrontal cortex, S1: primary somatosensory cortex, M1: primary motor cortex, mPFC: medial prefrontal cortex, FIC: frontoinsular cortex, R: right hemisphere, L: left hemisphere.

Figure 5.

Receiver operating characteristic (ROC) curves demonstrated superiority of combined parameter classification of relatively higher versus lower clinical pain intensity states. Comparison of individual-parameter versus combined-parameter classification performance using paired-SVM demonstrated most improved performance for the combined 3-parameter model. N.b. for a paired-SVM approach, sensitivity, specificity, and precision are identical to accuracy. rCBF: regional cerebral blood flow, S1CONN: S1-connectivity, HFHRV: high frequency power of heart rate variability, AUC: area under the curve.

Figure 6.

Prediction of clinical low back pain ratings across cLBP patients using all multimodal parameters. True LBP plotted against Predicted LBP from SVR results demonstrated that our model was successfully able to predict between-subject clinical pain intensity ratings for both a training (TRAIN, N=53, r=0.52) and independent testing dataset (TEST, N=53, r=0.63, shown above).