Multivariate classification of pain-evoked brain activity in temporomandibular disorder

Abstract

Introduction: Central nervous system factors are now understood to be important in the etiology of temporomandibular disorders (TMD), but knowledge concerning objective markers of central pathophysiology in TMD is lacking. Multivariate analysis techniques like support vector machines (SVMs) could generate important discoveries regarding the expression of pain centralization in TMD. Support vector machines can recognize patterns in “training” data and subsequently classify or predict new “test” data.

Objectives: We set out to detect the presence and location of experimental pressure pain and determine clinical status by applying SVMs to pain-evoked brain activity.

Methods: Functional magnetic resonance imaging was used to record brain activity evoked by subjectively equated noxious temporalis pressures in patients with TMD and controls. First, we trained an SVM to recognize when the evoked pain stimulus was on or off based on each individual's pain-evoked blood–oxygen–level–dependent (BOLD) signals. Next, an SVM was trained to distinguish between the BOLD response to temporalis-evoked pain vs thumb-evoked pain. Finally, an SVM attempted to determine clinical status based on temporalis-evoked BOLD.

Results: The on-versus-off accuracy in controls and patients was 83.3% and 85.1%, respectively, both significantly better than chance (ie, 50%). Accurate determination of experimental pain location was possible in patients with TMD (75%), but not in healthy subjects (55%). The determination of clinical status with temporalis-evoked BOLD (60%) failed to reach statistical significance.

Conclusion: The SVM accurately detected the presence of noxious temporalis pressure in patients with TMD despite the stimulus being colocalized with their ongoing clinical pain. The SVM's ability to determine the location of noxious pressure only in patients with TMD reveals somatotopic-dependent differences in central pain processing that could reflect regional variations in pain valuation.

Keywords: Artificial intelligence; Brain function; Magnetic resonance imaging (MRI); Neuroscience/Neurobiology; Orofacial pain/TMD; Support vector machines (SVM).

Figure 1.

Classification of pain vs off. (A) The mean prediction plot for temporalis-evoked pain vs rest for all subjects. Ideal prediction in blue, actual prediction in red, and decision plane at 0 in green. (B) Mean weight vector maps for face pain vs rest in all subjects. Pain > rest in orange, rest > pain in blue. Brighter colors indicate higher predictive value. (C) Negative correlation in patients with TMD between classification accuracy and SF-MPQ Total Score. Confidence intervals of 95% are shown in blue. Two patients have identical coordinates (x = 92.98, y = 3). SF-MPQ, Short form McGill Pain Questionnaire; SVM, support vector machine; TMD, temporomandibular disorder.

Figure 2.

Classification of evoked pain location in patients with TMD. (A) Classification accuracy for temporalis vs thumb pain. Predictions for the thumb are in red, predictions for the face are in blue, and the decision plane at 0 is in green. (B) Mean weight vector maps for temporalis pain vs thumb pain. Regions within the anterior cingulate cortex and operculum depict thumb >face. Brighter colors indicate higher predictive value. (C) Correlation between mean vector weights and SF-MPQ total scores. Confidence intervals of 95% are shown in blue. L, left; R, right; SF-MPQ, Short Form McGill Pain Questionnaire; SVM, support vector machine; TMD, temporomandibular disorder.