Experimental and clinical studies in diverse cohorts of patients (e.g., whiplash, fibromyalgia, osteoarthritis, musculoskeletal disorders, headache, and neuropathic, visceral and post-surgical pain) have shown that these pathologies share common features, which are likely to reflect alterations in central nociceptive processing [28, 29] leading to exaggerated pain sensitivity. It has been previously established that changes in central nociceptive processing can be detected by electrophysiological tests, such as those based on the NWR. In the past, the NWR has been widely used as a biomarker in the assessment of the state of the nociceptive system [5, 6, 30], and it has been proposed as a key tool in the research of central sensitization mechanisms, which are believed to be linked to the development of chronic pain [5, 7, 8, 31, 32].
In this regard, a number of studies showed that several patient groups present lower NWR thresholds compared to control groups of healthy volunteers [7, 8, 33, 34]. Moreover, it was also demonstrated that chronic pain patients (endometriosis, chronic low back and chronic neck pain) present larger RRF compared to pain free subjects [12, 31]. Lower NWR thresholds and enlargement of RRF are objective signs of central hyperexcitability, which could be a consequence of increased number of responsive spinal neurons or an expansion of the receptive fields of spinal neurons as a result of increased synaptic sensitivity [35, 36]. In the light of these facts, there is clear evidence that groups of patients with different chronic pain conditions display on average altered central pain processing. However, the next translational step in this field involves the definition of diagnostic criteria in individual patients, in order to develop treatments that are tailored to detect individual disturbances in central pain processing .
In this study, a set of features derived from the NWR of chronic low back and neck pain patients and healthy volunteers was used as input to a prediction model, in order to test the hypothesis that the NWR contains specific information that would allow individual classification regarding the condition of the test subjects. Several features derived from the NWR have been used in the past for detection or quantification purposes: NWR latencies, raw EMG amplitudes, mean and peak EMG values, EMG probability distribution, EMG root-mean-square (RMS), z-scores and RRF area size, among others [6, 27, 32, 37–41]. Additionally, other EMG features have also been used in classifications tasks in other fields (most notably myoelectric control systems), such as number of zeros crossings, slope sign changes, spectral moments, as well as frequency domain and time-scale features [42–44].
A preliminary analysis showed that, among these variables, EMG probability distributions showed the most promising results in terms of discriminating potential for classifying between patients and healthy volunteers. Thus, they were selected for further development of the prediction model. However, the EMG signals showed a rather large range, thus requiring a high number of bins for their histogram representation. In order to overcome this, a new range was defined, restricting the original range around one standard deviation of the mean, In this way, less bins were required for the representation (as a simple method of feature selection), effectively reducing the number of features to be fed to the prediction model.
The evaluation of stimulation sites for eliciting the NWR revealed that EMG signals recorded after stimulation of electrodes located in the anterolateral (site 3) and heel (sites 9 and 10) regions and of the sole of the foot presented the best prediction rates. This is in accordance with previous research showing that the RRF in chronic low back and neck pain patients are expanded compared to healthy volunteers, precisely towards these regions [31, 45]. On the other hand, EMG signals recorded after stimulation of sites located at or around the arch on the sole of the foot resulted in the worst prediction rates. These locations often have thin skin layer which lead to higher pain sensitivity and large reflexes regardless of whether they are patients or healthy subjects .
The final model evaluation showed an average prediction rate of 80%. For that particular choice of model, there were no differences in the misclassification rates between healthy subjects and patients. A more detailed analysis of the results focusing on the demographics of the two groups, revealed that women and elder subjects are more likely to be misclassified using the selected model. To date, there are no studies describing age or gender differences in EMG signals recorded from chronic pain patients compared healthy volunteers in relation to the NWR, since most of the research is focused on other biomarkers, most notably the NWR thresholds to single and repeated stimulation (temporal summation), and the RRF areas [9, 26, 32]. In this regard, there is still no agreement on the effects of age and gender on the NWR, although most of the evidence seems to point towards generally lower NWR thresholds in women and elderly subjects, most likely due to reduced endogenous analgesic mechanisms [6, 10, 47, 48].
To date, only population differences have been reported between chronic low back and neck pain patients compared to healthy volunteers, showing an enlargement of the RRF in patients . More recently, however, reference ranges for the NWR and RRF have been established for healthy subjects . These ranges establish critical values for several parameters derived from the NWR (e.g. NWR threshold to single and repeated electrical stimulation, RRF area), above which an individual subject can be considered to present widespread central hyperexcitability. Results using this method with RRF areas as the classification parameters showed lower average classification rates (r = 57.5%) and very low classification rates for patients (r
= 20%) compared to the prediction model. This is most likely due to the large inter-individual variation of the RRF areas and the high overlap that exists between the probability distribution of RRF areas in patients and healthy subjects.
Limitations and future work
This work focused on the assessment of central hyperexcitability in individual chronic low back and neck pain patients using the NWR. Although the quantification of the NWR does not rely on subjective self-reports of pain sensation, it is subjected to supraspinal modulation. External factors involving affective and cognitive processes or other ongoing nociceptive processes (e.g. endogenous pain modulatory mechanisms) can affect the NWR characteristics [6, 24], so these factors have to be carefully controlled for in order to provide reliable outcomes. Further tests in other patient groups should be conducted in order to test if this model could also be used to characterize other pain conditions.
Furthermore, this is the first known attempt at individualized classification between healthy subjects and chronic pain patients based on the assessment of central hypersensitivity provided by the NWR. As it is common in classification tasks, there are several variables that require a careful selection, such as the choice of features to be used as input to the prediction model (in this case, the EMG probability distribution), the parameters of the classifier (for kNN, the number of neighbours), the size of the datasets for classification, validation and test, and the number and location of stimulation sites selected. Some of these variables were chosen based on prior knowledge and/or empirical tests, so whereas the proposed statistical model is able to achieve high prediction rates, future research could focus on the application of more advanced signal processing methods, e.g. alternative methods for feature generation and selection, adaptive histograms, adaptive kernel density estimators and optimal parameter selection for the classifier, among others.