Effect of level of Sensitization on gait in Chronic low back pain: insights from a machine learning approach

Background: Despite central sensitization (CS) often presents in patients with chronic low back pain (CLBP), there is a lack of quantitative and qualitative analysis of the effect of CS on gait performance. The study aimed to investigate the daily-living gait performance of patients with CLBP with low and high CS levels (CLBP- and CLBP+, respectively). was Sensitization vertical/mediolateral and maximal Lyapunov exponent per stride-vertical (gait (cid:5)(cid:6)(cid:1) stability), and sample entropy-anteroposterior (gait predictability). (cid:5)(cid:7)(cid:1) Conclusions (cid:12) CLBP- and CLBP+ presented different motor control strategies. CLBP- (cid:5)(cid:8)(cid:1) presented a more “loose” control, including higher gait smoothness and stability. (cid:5)(cid:9)(cid:1) CLBP+ presented a more “tight” control, including a more regular, less variable and (cid:5)(cid:10)(cid:1) more predictable gait pattern. (cid:5)(cid:11)(cid:1) (cid:6)(cid:2)(cid:1) (cid:6)(cid:4)(cid:1)


Background
Chronic low back pain (CLBP) is one of the most prevalent chronic musculoskeletal pains in the world [1]. It is responsible for high treatment costs, sick leave and individual suffering and it represents a significant socioeconomic burden [2]. For 85% to 90% of patients with CLBP, the relation between pathoanatomical and clinical presentations is weak [3] and, therefore, it is classified as nonspecific CLBP [4]. In CLBP, and other chronic musculoskeletal disorders, central sensitization (CS) might be present [5]. CS is defined as "increased responsiveness of nociceptive neurons in the central nervous system to their normal or subthreshold afferent input" [6] and manifests as mechanical hypersensitivity, allodynia and hyperalgesia [7]. A considerable number of people need treatment for CLBP. Although the overall efficacy of CLBP rehabilitation programs is positive, the effect sizes are modest [8].
Correctly recognizing the physical and psychosocial factors perpetuating pain and physical disability of patients with CLBP remains a challenge [9]. Altered motor control of patients with CLBP could possibly contribute to the persistence of CLBP [10]. Altered motor control could affect daily-living activities, as patients with CLBP often exhibit altered movement patterns and motor control strategies; probably in order to avoid painful movement, such as walking [11]. Walking is one of the abilities most affected by CLBP. Many clinicians may intuitively identify "abnormal" gait patterns in patients with CLBP, but identification and objectifying of specific "abnormal" gait features is challenging. During walking, it is suggested that patients often adopt a "protective guarding" or "splinting" strategy [12] to avoid painful movements of the spine. These adaptations may lead to a slower and less flexible gait pattern [13].
A possible explanation for these inconsistencies might be an unknown heterogeneity within the samples, such as the presence of CS. CS could plausibly be related to the inconsistent results, because the presence of high CS levels is associated with longlasting chronic pain [19] and movement may be changed due to pain. Also, general gait features such as walking speed and stride length, might not be sensitive enough to detect small differences between patients with high and low levels of CS. In addition to stride related parameters, gait outcomes that reflect gait quality in terms of regularity, synchronization, smoothness, local stability and predictability, are sensitive to detect differences in gait performance. These gait outcomes were successfully used to detect the differences between age groups [20], older adults with and without fall risk [21], and patients with and without Parkinson's disease [22]. Despite the fact that the effects of CLBP on gait have been frequently investigated in controlled laboratory studies, there are no studies about the relationship between CS levels and gait performance under daily-living environment circumstances. Advances in wearable technology and machine learning methods offer new opportunities in gait data collection and analysis. Wearable technology allows researchers to record patients' physical activities in unobserved, daily-living environments over extended periods of time. This data can reflect the real gait performance of the patients since the controlled laboratory environment, while being observed, may change the performance of patients [23]. The successful employment of machine learning methods in gait analysis makes it possible to extract the most informative gait outcomes from the accelerometer sensor data [20]. If patients with lower and higher levels of CS walk differently, the machine learning methods will be able to successfully recognize these differences by their gait outcomes. Many gait outcomes are not independent and interact with each other, such as gait speed and step regularity. Machine learning methods such as Random Forest approach, are able to process high dimensional and non-linear data structures and take the interrelation and interaction of the gait outcomes into consideration [20]. Therefore, the aim of this study was to analyze whether and how the presence of CS is related to differences in gait performance of patients with CLBP during daily life by using a machine learning method. It was hypothesized that patients with CLBP and higher levels of CS show differences in daily life gait performance, compared with those with lower levels.

Patients
This study included patients with primary CLBP who were recruited from the outpatient Pain Rehabilitation Department of the Center for Rehabilitation of the University Medical Center Groningen (CvR-UMCG). Primary CLBP is defined as low back pain persistent for more than three months, with pain not being the result of any other diagnosis. The patients were selected according to the following inclusion criteria: (a) age between 18 and 65 years old at the time of recruitment; (b) admitted to the interdisciplinary pain rehabilitation program; (c) could follow instructions; (d) signed informed consent. Additionally, patients were excluded if they: (a) had a specific diagnosis that would better account for the symptoms (e.g. cancer, inflammatory diseases and/or spinal fractures); (b) had neuralgia and/or radicular pain in the legs; (c) were pregnant.
The study was approved by the Medical Research Ethics Committee of the University Medical Center Groningen (METc 2016/702) and conducted according to the principles expressed in the Declaration of Helsinki. The data used in this paper was derived from a larger study, of which protocol details were described elsewhere [19].

Data collection
Demographics were collected and standard clinical test were applied as part of the usual care of CLBP patients that are referred to the outpatient Pain Rehabilitation Department of the Center for Rehabilitation. Assessments included: Visual Analogue Scale for pain intensity (VAS Pain; 0-10), the Dictionary of Occupational Titles (DOT, the Pain Disability Index (PDI; 0-70), the physical functioning subscale of the Rand36 questionnaire (Rand36-PF; 0-100), the Pain Catastrophizing Scale (PCS, 0-52), the Injustice Experience Questionnaire (IEQ, 0-48), and the Brief Symptom Inventory (BSI global severity index t-score (GSIT))(see Table 3).

Central sensitization (CS). The presence of CS-related manifestations was assessed
with section A of the Central Sensitization Inventory (CSI) [24]. Section A has 25items to assess the presence of common CS-related symptoms. Scores can range from 0-100 where a higher scoring represents a higher level of CS. A score lower 40 indicates low CS level (CLBP-group) and a score of 40-100 is interpreted as high CS level (CLBP+ group) [25].
Mann-Whitney U test was used to statistically test the differences between CLBPand CLBP+ groups for demographics and CSI scores.
Accelerometer data. The accelerometer data was collected between 2017 and 2019.
Patients were instructed to wear a tri-axial accelerometer (ActiGraph GT3X, Actigraph Corporation, Pensacola, FL) at all times for about one week, excluding sleeping or bathing times. The accelerometer was worn at the front right hip of the patient (at the anterior superior iliac spine). Assuming a standing and upright position, the Y-axis pointed to the ground (vertical direction, V), Z-axis faced the walking direction (anteroposterior direction, AP), and the X-axis was perpendicular to the walking direction, pointing from a patient's right to left (mediolateral direction, ML). The sampling frequency of the accelerometer was set to 100 Hz and the dynamic range was ±6 gravity.

Raw data segmentation
Each patient's accelerometer data was segmented into 24 hours span data segments (from 12:00 P.M. to next day 11:59 A.M) to represent the activities during the days.
Data which was not completely covering this 24-hour span was discarded from the analysis. Due to technical or personal reasons of the patient, not all patients were collected a full week of data. In order to compare the data between different patients fairly, 4 segments (representing 4 days) of each patient were included. Therefore, 7 patients who had less than 4 segments, were excluded and from patients with more than 4 segments, 4 segments were randomly sampled. Figure 1a graphically shows the process of the raw data segmentation.

Walking bouts extraction
The accelerometer data of the 4 segments was first smoothed by a low-pass filter with a 2nd order Butterworth, filtered with a 20 Hz cut-off frequency, such that only frequencies lower than 20 Hz remained. Subsequently, potential walking events were detected by the Fast Fourier Transform (FFT) based method [26], which identified periods with 0.5-3.0 Hz power spectrum values. To remove false walking events from the potential walking periods, the zero-cross method [27] was employed. If the time interval between any two adjacent walking events was shorter than 2 seconds, these two walking events were merged into one walking bout. Finally, the walking bouts in Testing data 5-fold cross-validation each segment were extracted and their gait outcomes were calculated. Figure 1b presents the walking bouts as the yellow vertical bars in the rectangle.

Gait outcomes
All walking bouts in one 24-hour segment were used to determine the total duration of walking, the total number of steps, the maximum duration of a walking bout and the maximum number of steps of a walking bout. And then, all walking bouts exceeding 10 seconds were selected and cut into non-overlapping 10-second windows [28].
From the segment, each 10-second window was used to calculate different gait outcomes, and these values were averaged over all 10-second windows in the segment representing the patient's gait performance on that day.
The gait outcomes were divided into two categories, quantitative and qualitative gait outcomes. From one segment, we obtained one gait outcome vector, including 36 gait outcomes, based on the walking bouts (see Figure 1c). The detailed descriptions of the quantitative and qualitative gait outcomes are presented in Table 1 and Table 2 -for extended explanation of variables see reference [29]. SR is computed by using the unbiased autocorrelation coefficient: , where "//(0) is the sample acceleration signal, 5 the number of samples, and % the number of time lag. The first peak of "#(%) is "# ) and it represents the stride regularity. Higher values (maximum 1.0) reflect repeatable patterns between strides.

Gait symmetry index (GSI)
GSI quantifies the ratio of the first and second peak of the "#(%), as "# ) /"# 7 . It is a measure of the degree of symmetry of the left and right lower limbs during walking.

Smoothness
Index of harmonicity (IH; V, ML, AP, All) . It is the ratio of the power spectral density of the fundamental frequency ? @ and the sum of the power spectral density of the first six frequency

Random Forest Classifier
To separate CLBP-and CLBP+ groups by gait outcomes, a Random Forest classifier was used. The Random Forest classifier is considered as the optimal machine learning classification method for the present data, because it performs well with (a) nonlinear and linear data; (b) high dimensional data, obsoleting dimensionality reduction; and (c) unbalanced and small data sets [30].
The input data of this method was < G, h >. G represents the gait outcome vectors of The Random Forest is constructed in four steps.
Step one: Randomly sample O gait outcome vectors from G and O corresponding labels from h, with replacement. These new set of gait outcome vectors and labels are called G E and h E . In G E , k 3 may appear more than one time or not appear.
Step two: Step three: Training a decision tree P E on G E v , h E .
Step four: Repeat steps one to three 1000 times and combine the decision trees into an ensemble, called random forest, that predicts by voting (see Figure 2). Output: Random Forest training data and the hyperparameters were used to fit the final Random Forest classifier and the testing data set was used to evaluate the generalizability of the model.

Accuracy evaluation
Accuracy, sensitivity, specificity, precision, and F1-score were calculated to evaluate the performance of the classification (see Figure 1f). In this study, CLBP+ was Specificity refers to the rate of correctly predicted negative cases in all negative cases (true negative rate).
Precision is the ratio of the correctly predicted positive cases in all predicted positive cases.
The receiver operating characteristic (ROC) curve was calculated to evaluate the performance of the Random Forest classifier. The Y-axis of this curve represents the true positive rate (sensitivity) and the X-axis means false positive rate (1-specificity).
Therefore, the overall classification performance of the Random Forest classifier was evaluated by the area under the ROC curve (AUC). A classification model with a larger AUC value has a higher correct rate, and AUC = 1 means perfect performance. Since the random forest can also output the percentage of trees that vote CLBP+, AUC represents the probability that the random forest outputs a lower percentage for a negative sample than for a positive sample.

Feature importance
SHapley Additive exPlanations (SHAP) [31] was used to explain the gait features importance to the classification model. SHAP connects optimal credit allocation with local explanations using the classic Shapley values from game theory [32]. Shapley values, ∅ 3 , explains the importance of gait outcome 0 for Random Forest classifier and is defined as: where 5 is the size of the full set of gait outcomes, s is the subset that includes 0 in computing the exact Shapley values is computationally expensive, SHAP method uses a tree explainer to exploit the information stored in the tree structure to calculate the SHAP values which are highly approximate Shapley values. Therefore, higher SHAP values represent higher impact to classify CLBP-and CLBP+ groups.

Results
Demographic characteristics are provided in Table 3. Out of a total of 60 patients, 11 were excluded because essential parts of their dataset were incomplete (CSI scores or/and accelerometry data), 7 were excluded since they had less than 4 segments data (3 had 1 segment, 2 had 2 segments, and 2 had 3 segments). Therefore, 42 patients were included for the data analysis. Differences between CLBP+ and CLBP-group characteristics ( Table 3) were not statistically significant (p > 0.05), with exception of CSI score (p < 0.0001) and BSI (p = 0.01).
Testing data was used to evaluate the generalizability of the Random Forest classifier and the confusion matrix is shown in in Figure 3. From the confusion matrix, accuracy, sensitivity, specificity, precision, and the F1-score were calculated to evaluate the performance metric of the model. The Random Forest classifier achieved an accurate classification-result (84.4% accuracy) and the sensitivity and specificity were 75.0% and 93% respectively. The precision is 92% and the F1-score is 82.6%. The ROC curve is presented in Figure 4 showing that the Random Forest classifier achieved a 0.83 AUC.   The importance of the gait outcomes for the Random Forest classifier is shown in Figure 5. In this study, only the top 10 gait outcomes (above the red line in Figure    Similarly, in SR-ML, CLBP+ has a concentrating distribution while the peak of CLBPis lower. In Max LyE-V and Max LyE per stride -V, CLBP-shows a log-normal distribution while CLBP+ shows a wider distribution. In HR-ML and SL Variability, the distributions are similar but CLBP+ has more outliers.

Discussion
The aim of this study was to analyze whether and how the presence of CS is related to differences in gait performance of patients with CLBP during daily life by using a machine learning method. Based on quantitative and qualitative gait outcomes, using a Random Forest method, the two groups (CLBP-and CLBP+) could be classified with an accuracy of 84.4%. The classification results indicated that CLBP-patients walk differently from CLBP+. Furthermore, the SHAP values showed that the differences between CLBP-and CLBP+ groups were present in gait outcomes that represented smoothness, stability, predictability, regularity and variability.
In the present study, we addressed the walking of patients with CLBP in a daily-living environment. Walking in a controlled laboratory or during a clinical assessment is different from self-initiated gait, during activities of daily living. Walking in daily life, might be subject to environmental perturbations, quick changes while performing a task, and often involves the performance of several actions at the same time [33], e.g.
walking when carrying a cup of coffee. These influences on gait are not present in controlled studies, and are not captured by averaged based conventional gait outcomes that average outcomes over stride cycles, such as mean, step length, stride or step time, and number of steps. Therefore, in the present study we included gait outcomes that take into account how gait cycles evolve over time, e.g. the interdependency of gait cycles, using sample entropy as a measure of predictability of the gait pattern, the maximal Lyapunov exponent as quantification of local stability and correlation-based measures that take into account how gait cycles evolve over time [34].
The Additionally, it can help to reduce the chance of overfitting to training data. Therefore, the Random Forest improves predictive accuracy and it can provide a generalized model of the difference between CLBP-and CLBP+ groups.
The SHAP tree explainer can provide good explanations for the Random Forest [31]. It fastens the calculation of the SHAP values by exploiting the information stored in the tree structure which already captured the gait outcomes interactions. The top 10 SHAP values suggest that the differences between CLBP-and CLBP+ groups are gait outcomes which represent smoothness, stability, predictability, regularity and variability in gait. Compared with CLBP-group, CLBP+ group exhibited worse smoothness and local stability in gait. Moreover, the CLBP+ group exhibited a more regular, less variable and more predictable gait pattern.
Most studies on walking in patients with CLBP are compared with control participants with no back pain. To the best of our knowledge, this is the first study in patients with CLBP that addresses the difference in gait pattern between two CLBP groups based on low and high CS level, which makes a direct comparison with other studies intricate.
However, the results of different gait patterns between low and high CS level support the notion that within the heterogenous CLBP group, different motor control strategies are adopted. Two motor control strategies on a continuum have been suggested with "tight" control and "loose" control at each end, and normal trunk control in the middle [35].
There is considerable evidence on the observed changes in muscle activation of patients with CLBP [36]. Tight control which involves increased trunk muscle activation and enhanced muscle co-contraction, might enhance control over trunk posture and movement [35]. Increased muscle activation and enhanced co-contraction would help individuals to maintain the stability of lumbar spine [37], which is an unstable structure.
However, this strategy might impair patients' ability to maintain balance in a complex daily-living environment where unstable surfaces and environmental perturbations occur [38]. Increased co-contraction would reduce the demand for the intricate control of the sequences of muscle activation. It might avoid the potential error raised by inaccurate sensory feedback of CLBP [35]. This might allow patients to control their trunks' movement precisely and result in lower variability of movement patterns [39].
Our results might infer that CLBP+ group exhibited a more "tight" control. Tight control allows them to control their movement more strictly with results in a decreased motor variability, higher regularity and predictability of the gait pattern compared with CLBP-group.
The CLBP-group on the other hand might use another 'loose control' strategy. The "Loose control", which involves reduced muscle excitability, might reduce the control over trunk movements [35]. The spine of which each spinal unit has 6 degrees of freedom, is controlled by its surrounding musculature. Reduced muscular excitability, leads to a reduced control over the spinal muscle, to larger amplitude movements, and to more movement variability during repeated tasks [35]. Increased motor variability might probably prevent muscle fatigue [40] since it allows sharing the load between different structures or tissues. Moreover, motor variability makes it possible to explore new pain-free motor control solutions [41]. The results of the present study hint at a more 'loose control' in the CLBP-group, which increased motor variability, which might allow them to flexibly adapt to the complex daily-living environment using different movement solutions. Additionally, irrespective of the larger variability local stability and smoothness of the gait pattern was higher in this group than in the CLBP+ group.
Although both motor control adapted strategies might have beneficial effects for the short term, they might lead to negative long-term consequences. Increased muscle activation and co-contraction in tight control would increase the forces acting on the spine and it would lead to higher spinal loading. Moreover, even when patients are at rest, the co-contraction of muscles is continuous [42]. These might result in accumulation of waste products, muscle fatigue, and intervertebral disc degeneration [40] [43]. Regarding loose control, the reduced control of lumbar spine may eventually increase the tissue strains, with subsequent increases in spinal loading, and pain [41].
Thus, eventually both strategies affect mechanical loading on lumbar tissues. The loading might be the source of nociceptive input and might contribute to the CS since the load may sufficient to excite sensitized afferents [35].
Clinically, the important gait outcomes of this study may assist clinicians in providing a more accurate understanding of the gait performance of patients with CLBP, with low or high CS levels, and a more explicit operationalization of the observed "abnormal" gait pattern of patients with chronic pain. Whether "abnormal" should be interpreted as a functional or a dysfunctional motor control strategy in the short or long term, remains to be studied. The approaches used in this study have presented a novel way to identify interacting feature, and therefore, can be used for further studies. Clinically, the present accurate subclassification could become meaningful if this would lead to effective treatment approaches. While this cross-sectional study has objectified a relation between CS and gait features, the direction of this relation is unknown. Follow-up studies would benefit from a longitudinal design with multiple measurements to help further unraveling of this relation, as well as the relation to disability.
In line with most studies on walking and CLBP, we used cross-sectional data, thus we are not allowed to infer causality between motor control changes, CS and CLBP.
Moreover, we labeled the groups based on CSI score and the cut off values from a previous study [25]. It should also be noted that a gold standard measure to diagnose CS is unavailable. The CSI is regarded as an indirect measure of CS, because higher scores are associated with the presence of CS syndromes [25]. In addition to gait assessment, it would be interesting to explore differences in physical activities between CLBP-and CLBP+, because several studies reported that relationship between CLBP and physical activity levels is heterogeneous [44].

Conclusion
The present study analyzed gait data during daily living of CLBP patients with low and high CS levels. A Random Forest method and the SHAP method were applied for classification and identification the contribution of gait outcomes to the model. This analytic approach demonstrated that Random Forest method has the ability to accurately classify subgroups of patients with CLBP and low or high CS levels based on differences in gait outcomes. The results of SHAP method showed the differences between low and high CS levels were in gait regularity, variability, predictability, smoothness and stability. The differences in gait outcomes may infer that patients with low and high CS levels adopted different motor control strategies. Patients with CLBP and low CS level (CLBP-) may use a more loose control and, therefore, exhibited more smoothness and stability in gait patterns. Patients with CLBP and high CS level (CLBP+) may adopt a more tight control and showed a more regular, less variable and more predictable gait pattern.
The results of this study may contribute to a better understanding of gait characteristics in patients with CLBP, its association with CS, and may in the future assist in betterpersonalized rehabilitation interventions [45].