Structural brain connectivity predicts acute pain after mild traumatic brain injury

Mild traumatic brain injury, mTBI, is a leading cause of disability worldwide, with acute pain manifesting as one of its most debilitating symptoms. Understanding acute post-injury pain is important since it is a strong predictor of long-term outcomes. In this study, we imaged the brains of 172 patients with mTBI, following a motorized vehicle collision and used a machine learning approach to extract white matter structural and resting state fMRI functional connectivity measures to predict acute pain. Stronger white matter tracts within the sensorimotor, thalamic-cortical, and default-mode systems predicted 20% of the variance in pain severity within 72 hours of the injury. This result generalized in two independent groups: 39 mTBI patients and 13 mTBI patients without whiplash symptoms. White matter measures collected at 6-months after the collision still predicted mTBI pain at that timepoint (n = 36). These white-matter connections were associated with two nociceptive psychophysical outcomes tested at a remote body site – namely conditioned pain modulation and magnitude of suprathreshold pain–, and with pain sensitivity questionnaire scores. Our validated findings demonstrate a stable white-matter network, the properties of which determine a significant amount of pain experienced after acute injury, pinpointing a circuitry engaged in the transformation and amplification of nociceptive inputs to pain perception.

63 and exploring the mechanisms underlying pain at the early acute stage is a promising approach to 64 resolve the relative contribution of psychological, psychophysical, and nociceptive processes in 65 post-injury pain and can shed light on the mechanisms that underlie pain perception immediately 66 after an emotionally charged, pain-inducing incident.

67
In this study, we used MRI to study brain functional and structural properties of mTBI 68 patients after a motorized vehicle accident and probed for properties that can determine or 69 predispose patients to experience pain after injury. Using a machine-learning approach, we 70 examined the functional and structural brain networks associated with early, acute pain. This study 71 sheds light on the properties and emergence of pain, particularly after injury, with implications for 72 the treatment and management of mTBI.

75
Predicting pain after mTBI using functional and structural connectivity 76 We used a machine-learning approach to predict pain from resting-state functional magnetic 77 resonance imaging (rsfMRI) and diffusion tensor imaging (DTI) brain connectivity. Discovery 78 (70% of the data, N rsfMRI = 94 and N DTI = 88 after outlier exclusion) and hold-out (N rsfMRI = 43 79 and N DTI = 37 after outlier exclusion) datasets were used to build the model and assess 80 generalizability (see Fig. 1 and methods for an explanation of the machine-learning pipeline).
81 For structural connectivity, after 10-folds cross-validation (CV), the best p threshold for 82 univariate feature selection was determined to be p = 0.01 (negative features model, highest CV 83 r predvsactual = .33; positive model n.s.), leading to the selection of 26.5 features on average, which 84 mostly appeared inconsistently across CV folds (see Supplementary Fig. 1). We also examined the left thalamus using the FSL thalamic connectivity atlas (13)

155
Given that these clinical parameters have been previously associated with pain ratings in 156 these individuals(8,9), we further wanted to disentangle the relative contribution of these 157 psychological and psychophysical dimensions to that of the brain. To do so, we performed a 158 relative importance analysis (using the R package relaimpo (14)). While the total model with the 159 four parameters (CPM, Pain50, PSQ, and Brain Connectivity) explained 33.6% of the variance in 160 pain ratings (Fig. 4B), brain connectivity alone explained 15% unique variance in pain ratings.  203 and connectivity metrics collected at baseline, as well as the same parameters collected at six 204 months, are able to predict subjects' pain at six months after the initial injury. Together, these 205 findings suggest an a-priori predisposition to pain grounded in brain WM properties, which results 206 in higher pain ratings after acute injury. These findings shed light on brain mechanisms underlying 207 pain sensitivity and inform the current literature of pain perception of a poorly understood patient 208 population -mTBI patients.

209
Our study shows that an important part of the reported post-injury pain can be accounted 210 for by brain imaging features alone. Although previous studies have successfully predicted 211 experimental pain(15) and tonic pain(16) in healthy participants, as well as pain in patients at the 212 chronic stage(16), this is, to the best of our knowledge, the first demonstration that early acute 213 pain, isolated from plastic changes that occur over time after an injury, can be predicted based on 214 brain structure. The strength of the evidence here lies in the clinical sample studied: these patients 215 did not have any pain before injury and were scanned within hours of the accident, making them 216 an ideal group to study injury-related, early acute pain in an ecological manner. The influence of 217 the brain in pain perception is clear in the literature, but reports using brain structure and function 218 to predict pain, particularly post-injury acute pain are scarce. Spisak and colleagues(17) have, for 219 instance, recently showed that brain functional connectivity can predict pain sensitivity in healthy 220 participants, which points to a predisposition to pain grounded on their underlying 221 connectome(17). Here, we extend that idea by showing that injury-related pain can also be 245 suggesting that sensorimotor connectivity is related to pain sensitivity. This is also coherent with 246 previous work showing that higher cortical density in the somatosensory cortex leads to less 247 experienced pain while performing quantitative sensory tasks in healthy subjects(32) and that 248 denser cortical thickness of the somatosensory cortex is associated with higher pain thresholds(33).
249 In fact, it is known that the sensorimotor area is able to modulate pain, namely through direct

280
An important question to discuss is whether these brain features reflect injury-related 281 parameters, short-term plastic changes, or rather an a-priori predisposition to pain hard-wired in 282 the brain. Our results favor the latter. Considering an injury-related explanation to these results, 283 one could argue these parameters are mapping tissue injury as a proxy for pain: it is known that 284 whiplash-like centrifugal forces may cause axonal micro-lesions that are not detectable in 285 conventional scanning protocols(43); and these brain lesions could, in turn, reflect how severe the 286 mTBI/whiplash injury is(44), leading to higher reported pain. We find this possibility unlikely: 287 first, we excluded patients with WAD scores above three and patients with obvious signs of brain 288 damage. Second, we did not find associations between connectivity strength and whiplash severity, 289 and the model was able to predict pain in the group of patients with no whiplash-like symptoms.
290 Lastly, there are accounts that white-matter properties, namely fractional anisotropy and median 291 diffusivity, are unaffected following mTBI(45). Another possibility is that this connectivity is 292 reflecting short-term plastic changes caused by the new pain state. Our data does not support this 293 hypothesis either: participants were scanned within 72h of the injury. It is unlikely that white-294 matter properties changed over a timespan of hours and ceased to change any further over longer 295 timespans. There is evidence of short-term white-matter diffusivity changes in other 296 contexts(46,47), but if the system is so malleable, one would expect to observe diffusivity changes 297 over longer times too, especially as many participants gradually recover from their acute pain. We 298 in fact observe no changes in connectivity strength in these WM networks from the time of the 299 accident to six months and up to a year. Moreover, these same connections measured at six months 300 post-injury also predict the reported pain at the same point in time, further adding strength to the 301 within-subject reliability of this model. In conclusion, we consider it quite unlikely that our results 302 reflect injury-related parameters or plastic changes that occur in the timespan of the data-303 collection; our data instead favors the idea that observed WM properties reflect an a priori brain 304 predisposition for pain sensitivity.

305
While white-matter properties predicted pain quite successfully, fMRI functional 306 connectivity did not. It is thought that functional connectivity is primarily supported by structural 307 connectivity(48), and given the success of using resting-state to study pain sensitivity(17) and 308 phasic pain(16), it is somewhat surprising that functional connectivity was not informative of acute 309 mTBI pain. It is, however, important to point out key differences between DTI and rsfMRI 310 techniques. While white-matter properties are relatively stable and should reflect primarily long-311 term changes in brain structure (i.e., more trait-like), rsfMRI should better reflect the emergence 312 of a state within a given macro-structure (i.e., more state-like(49)). Given that patients were 313 scanned within 72h of a motorized vehicle accident, an often emotionally charged event that will 314 no doubt imprint a state of anxiety and distress on the patient, it may be difficult to identify a 315 reliable and solid pattern that generalizes across subjects, particularly given the heterogeneity of 316 the psychological factors and their influence on functional connectivity. Importantly, the fact that 317 only white matter is predictive of pain post-injury in the patients should be construed as further 318 evidence that they represent stable brain circuitry implicated in the transformation and 319 amplification of nociceptive inputs, rather than simply capturing the perception (state) of pain per 320 se.

321
This study has some limitations. We lack a group of healthy participants, which could be 322 used to directly ascertain if the connectivity profiles are affected by the injury itself. Also, it is 323 necessary to mention that the absolute agreement between predicted pain and observed pain is 324 subpar (i.e. participants with 0 pain are predicted to have an average of 30 pain). Nonetheless, the 325 goal of this study was to infer brain mechanisms from early acute pain, rather than minimizing the 326 error in out-of-sample prediction. Finally, our sample is quite heterogenous in several respects, 327 including acute treatment, type of car accident, and perhaps others. Naturally, there is an vast 328 number confounds that could have been controlled for to improve the predictive ability of the 329 model; here, we favored a less constrained approach as it provides strength to the predictive power 330 of our analyses -it predicts pain regardless of uncontrolled confounds.

331
In conclusion, this study shows that measures of brain diffusivity in white-matter tracts 332 predict an important part of acute pain shortly after an injury. These tracts are implicated on 333 nociceptive and pain-related circuitry, which may reflect preexisting pro-nociceptive or anti-

356
We recruited 249 patients, out of which we excluded 17 patients who did not fulfill mTBI 357 clinical criteria, 54 patients who did not undergo MRI (claustrophobia, or not willing to 358 participate), four patients who did not report baseline pain, and two patients with incidental MRI