You are on page 1of 9

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/315630237

A Machine Learning Based Surface Electromyography Topography Evaluation


for Prognostic Prediction of Functional Restoration Rehabilitation in Chronic
Low Back Pain

Article in Spine · March 2017


DOI: 10.1097/BRS.0000000000002159

CITATIONS READS

19 355

3 authors, including:

Naifu Jiang Yong Hu


Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences Fudan University
15 PUBLICATIONS 50 CITATIONS 625 PUBLICATIONS 17,860 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Nanoparticulate Formulations for Paclitaxel Delivery Across MDCK Cell Monolayer View project

depression EEG View project

All content following this page was uploaded by Naifu Jiang on 28 November 2019.

The user has requested enhancement of the downloaded file.


SPINE Volume 42, Number 21, pp 1635–1642
ß 2017 Wolters Kluwer Health, Inc. All rights reserved.

DIAGNOSTICS

A Machine Learning-based Surface


Electromyography Topography Evaluation for
Prognostic Prediction of Functional Restoration
Rehabilitation in Chronic Low Back Pain
Naifu Jiang, MSc,,y Keith Dip-Kei Luk, FRCSE, FRCSG, FRACS, MCh(Orth), FHKAM(Orth),
and Yong Hu, PhD ,y

were applied to the prediction based on the optimized features


Study Design. A retrospective study.
selected by the sequential floating forward selection (SFFS)
Objective. The aim of this study was to investigate the
algorithm.
feasibility and applicability of support vector machine (SVM)
Results. RMSD feature parameters following rehabilitation in
algorithm in classifying patients with LBP who would obtain
the ‘‘responding’’ group showed a significant difference
satisfactory or unsatisfactory progress after the functional restor-
(P < 0.05) with the one in the ‘‘nonresponding’’ group. The SVM
ation rehabilitation program.
classifier with Quadratic kernel based on SFFS-selected features
Summary of Background Data. Dynamic surface electro-
showed the best prediction performance (accuracy: 96.67%,
myography (SEMG) topography has demonstrated the potential
sensitivity: 100%, specificity: 93.75%, average area under curve
use in predicting the prognosis of functional restoration rehabili-
[AUC]: 0.8925) comparing with linear kernel (accuracy:
tation for patients with low back pain (LBP). However, proces-
80.00%, sensitivity: 85.71%, specificity: 75.00%, average AUC:
sing from raw SEMG topography to make prediction is not easy
0.7825), polynomial kernel (accuracy: 93.33%, sensitivity:
to clinicians.
Methods. A total of 30 patients with nonspecific LBP were 92.86%, specificity: 93.75%, average AUC: 0.9675), and radial
recruited and divided into ‘‘responding’’ and ‘‘non-responding’’ basis function (RBF) kernel (accuracy: 86.67%, sensitivity:
group according to the change of Visual analog pain rating scale 85.71%, specificity: 87.50%, average AUC: 0.7900).
Conclusion. The use of SVM-based classifier of SEMG topogra-
and Oswestry Disability Index. Each patient received a 12-week
phy can be applied to identify the patient responding to
functional restoration rehabilitation program. A normal database
functional restoration rehabilitation, which will help the health-
was calculated from a control group from 48 healthy partici-
care worker to improve the efficiency of LBP rehabilitation.
pants. Root-mean-square difference (RMSD) was extracted from
Key words: dynamic surface electromyography topography,
the recorded dynamic SEMG topography during symmetrical
functional restoration rehabilitation, low back pain, machine
and asymmetrical trunk-movement. SVM and cross-validation learning.
Level of Evidence: 3
Spine 2017;42:1635–1642
From the Department of Orthopaedics and Traumatology, Li Ka Shing
Faculty of Medicine, The University of Hong Kong, Pokfulam, Hong Kong;
and yDepartment of Orthopaedics and Traumatology, Shenzhen Key
Laboratory for Innovative Technology in Orthopaedic Trauma, The Uni-

L
versity of Hong Kong-Shenzhen Hospital, Shenzhen, 518053, China. ow back pain (LBP) is a common problem faced by a
Acknowledgment date: October 17, 2016. First revision date: January 17, vast population.1,2 It significantly affects people’s
2017. Second revision date: February 17, 2017. Acceptance date: February daily life. To relieve the LBP, various treatments have
24, 2017.
been proposed and applied.3–7
The device(s)/drug(s) is/are FDA-approved or approved by corresponding
national agency for this indication. Functional restoration rehabilitation program has been
No relevant financial activities outside the submitted work. proved to be a common and effective treatment for non-
Shenzhen Knowledge Innovation Program of Basic Research specific LBP patients.6,7 Nevertheless, not all nonspecific
(JCYJ20150331142757393) funds were received in support of this work. LBP patients would have a good recovery result after reha-
Address correspondence and reprint requests to Dr. Yong Hu, PhD, Depart- bilitation program.8,9 Thus, the patient who would respond
ment of Orthopaedics and Traumatology, The University of Hong Kong, to rehabilitation program needed to be triaged.
Address: 12 Sandy Bay Road, Pokfulam, Hong Kong; E-mail: yhud@hku.hk
The current method of triaging LBP patient was to seek
DOI: 10.1097/BRS.0000000000002159 the direct correlation between the outcome of rehabilitation
Spine www.spinejournal.com 1635
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.
DIAGNOSTICS Prognostic Prediction of Functional Restoration Rehabilitation in Chronic LBP  Jiang et al

and clinical variables.10–15 However, those variables (e.g., All LBP patients have no spinal deficits, or any other back
age, pain intensity, days of sick leave, fear of physical surgeries. The duration of the LBP for each patient is over 3
activity, and frequency of previous episodes) were quite months without related organic findings warranted for
subjective and imprecise, resulting in lead inappropriate surgical intervention. All LBP patients were asked to com-
therapies to some patients. It is very disappointing for a plete a standard questionnaire including the age, sex,
patient to join a long-term rehabilitation without expected medical history, and nature of their pain symptoms (i.e.,
outcome. If the patient is referred to a wrong type of treat- location, frequency and duration of the pain, visual analog
ment, it will waste not only the patient’s rehabilitation time, scale [VAS]). The Oswestry Disability Index (ODI) of each
but also the valuable healthcare resources. Therefore, it is patient was calculated from results of the Oswestry Low
crucial to accurately predict the prognosis of functional Back Pain Questionnaire. Those patients with a high level of
restoration for each LBP patient before he/she attends this pain-related disability (ODI >60) or a disability in perform-
program with an appropriate expectation. ing lumbar flexion-extension and lateral-bending/turning
Surface electromyography (SEMG) is a noninvasive were excluded. To avoid the effect of the ‘‘fear of re-injury’’
medical diagnosis technique to record the electrical activity and kinesiophobia,23,24 those subjects not willing to per-
produced by skeletal muscles.16 Through visualizing the form trunk-movement were excluded. To establish the nor-
lumbar muscle myoelectric activity, dynamic SEMG top- mal value of SEMG topography, a control group with 48
ography was proposed to be a new objective tool to assess healthy subjects (mean age 36  7.3 years, 33 males and 15
the LBP and its rehabilitation progress.17–21 A recent report females) was recruited. Before testing, all subjects and
proposed to use some feature parameters from SEMG top- healthy volunteers signed written informed consent forms
ography during symmetrical flexion-extension trunk-move- approved by the ethics committee.
ment to predict the prognosis of functional restoration A 12-week functional restoration rehabilitation program
rehabilitation for chronic LBP.17 The root-mean-square was applied to all LBP patients. It consisted of physical
difference (RMSD) of relative area (RA) and relative width conditioning (5 weeks), working conditioning (4 weeks),
(RW) within both flexion and extension phases showed and work readiness (3 weeks).6,7 To avoid potential bias, a
potential to predict the LBP patient who would respond senior orthopedic surgeon recruited the patient based on his
to the conservative care program.17 clinical assessment (interrogation, physical examination,
However, there was a lack of an optimization for the and radiographic assessment), and referred to a standard
prediction technique on SEMG topography in the previous exercise rehabilitation program conducted by a
study for two reasons.17 First, the information of lumbar senior physiotherapist.
neuromuscular function for LBP rehabilitation assessment
was not comprehensive because it only included SEMG Clinical Assessment
topography during symmetrical flexion-extension trunk- Before and after rehabilitation program, patients were
movement without asymmetrical movement. Second, the required to accomplish an 11-point VAS and an ODI ques-
previous prediction was often based on threshold decision tionnaire for LBP.17,25 The patients with good progress
method, which had clinicians to make classification with the (VAS decreased by two-points and/or ODI decreased by
complicated SEMG topography. During the prediction pro- 10-points) were classified as ‘‘responding’’ group, whereas
cedure, the feature parameters were analyzed as separate the others were ‘‘nonresponding’’ group.26,27
parameters while they were actually inter-related and there-
fore should have been analyzed in an integrated approach. Data Collection
Comparatively, support vector machine (SVM)—a common An array of 3X7 surface Ag/AgCl electrodes with three rows
machine learning algorithm—could provide a more accurate and seven columns was attached evenly to the region of
prediction as it integrates feature parameters from SEMG lumbar muscles from the spinal level of L2 to L5
topography. SVM’s fundamental principle is to find the (Figure 1).17,19 The diameter of each electrode is 1.5 cm.
maximum-margin hyperplane, which could best classify Before electrodes were attached, the skin had been cleaned
the data.22 with alcohol, whereas the impedance was tested under
Therefore, this study is to apply a SVM classification 10 kV. The SEMG signal was acquired at a sample rate
method to automatically predict and identify the LBP of 2000 Hz and filtered between 15 and 950 Hz (YRKJ-
patients who respond to the functional restoration rehabil- A2004, Zhuhai Yiruikeji Co, Ltd, China).
itation program. The method uses time-varying dynamic To comprehensively assess the lumbar muscle’s behav-
SEMG topography during both symmetrical and asymmet- iors, the subjects were asked to accomplish several sym-
rical trunk-movement. metrical and asymmetrical trunk-movements. SEMG data
were acquired during these movements. The types of trunk-
MATERIALS AND METHODS movements included flexion-extension trunk-movement,
left-/right- lateral-bending trunk-movement, and left-/right-
Subjects turning trunk-movement shown in the Figure 2. The detailed
A total of 30 patients with nonspecific LBP (mean age procedure was at first they were required to straightly stand
43.8  9.3 years, 19 males and 11 females) were recruited. on the ground. Then they spent about 1 second to forward-
1636 www.spinejournal.com November 2017

Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.
DIAGNOSTICS Prognostic Prediction of Functional Restoration Rehabilitation in Chronic LBP  Jiang et al

Figure 1. Placement of 3X7 electrodes and a typical surface electromyography topography during flexion-extension trunk-movement with the
definition of the quantitative feature parameters. REF indicates reference electrode; GND, ground electrode; 1-16, 16 channels of target
surface electromyography electrode.

bend (308), lateral-bend (458), or turn (608) their trunk (the Feature Extraction
degree was measured by protractor). Subsequently, they After selecting three channels of SEMG signal as the reference
maintained this posture for about 2 seconds. Finally, they and two channels as the ground from 3X7 array of surface
extended or turned back their trunk and returned to the electrodes, the remaining 60 channels of SEMG signals were
original standing posture within 2 seconds. All subjects were recorded and amplified by 2000 times (Figure 1). Through
asked to perform each movement twice and had a rest for calculating the root mean square (RMS) value of the ampli-
about 2 minutes between each performance. More rest was fied data and the linear cubic spline interpolation value of the
allowed if the subject required. All SEMG measurements RMS among electrodes with a 400 ms of moving window, the
were carried out in one room with a constant temperature. dynamic time-varying SEMG topography was formed.17–19
RMS parameters including RA, RH, and relative width (RW),
which appeared in previous study, could reflect SEMG top-
ography.18,19 The Figure 1 showed the method to calculate
RMS parameter during flexion-extension trunk-movement,
which was same with trunk-lateral-bending movement and
trunk-turning movement.
Based on RMS parameter, we used RMSD for further
prediction. The RMSD was a feature parameter to calculate
the difference of RMS parameter between the normal and
the LBP patient. It had been defined in previous study17 and
expressed in Appendix A, http://links.lww.com/BRS/B267.
In this study, above RMS parameters were applied to
measure the features of the highest 60% RMS value region
in SEMG topography. To predict the ‘‘responding’’ and
‘‘nonresponding’’ groups, RMSD parameters within the
flexion/turning phase and extension/returning phase were
selected and compared between the ‘‘responding’’ and
‘‘nonresponding’’ groups.

Feature Selection
In previous study, researchers found the performance of the
machine learning model was often not perfect owing to the
redundant features.28 Thus, in this study, it was necessary to
test whether the features were redundant and then to opti-
Figure 2. View of the defined trunk-movement for subject. mize the prediction result based on the selected optimal
Spine www.spinejournal.com 1637
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.
DIAGNOSTICS Prognostic Prediction of Functional Restoration Rehabilitation in Chronic LBP  Jiang et al

In Figure 3, the original dataset was divided using cross-


validation method. The cross-validation method evaluates
the performance of SVM model by dividing the original
dataset into different groups of training dataset and testing
dataset and calculating the average performance from test-
ing datasets.33,34 In this study, because of the small sample
Figure 3. Flow chart of SVM-based prediction procedure. SVM
size, the leave-one-out cross-validation (LOOCV) was
indicates support vector machine. applied to calculate the average accuracy, sensitivity, and
specificity of the SVM model, whereas the 10-fold cross-
validation (10 times) was applied to obtain the average area-
features. Sequential floating forward selection (SFFS), a under-curve (AUC) from receiver-operating-characteristic
feature selection algorithm, was applied in this study.29–32 (ROC) curves.22,33–36
Its basic principle is to perform an iterative process: to
calculate the best performance of model by adding the feature Statistical Analysis
one by one from one feature to all features and simultaneously All presented data were analyzed using MATLAB 7.14.0
recalculate it by removing the feature one by one till the (Mathworks, Natick, MA). RMSD parameters from SEMG
performance does not turn worse. It is guaranteed to obtain topography between the ‘‘responding’’ and ‘‘nonrespond-
the optimal features and corresponding feature’s number. ing’’ group were compared by one-way analysis of variance
Appendix B, http://links.lww.com/BRS/B267 describes SFFS (ANOVA). P value <0.05 was considered as statistically
in detail. significant and <0.01 was statistically highly significant.

Support Vector Machine Model and Evaluation of RESULTS


Model Performance
After extracting and selecting the optimal features from Classification of LBP Patients by Clinical Assessment
SEMG topography, it is significant to form a proper pre- The demographic data of healthy subjects and LBP patients
diction/classification model and evaluate its generalized were summarized in Table 1. In this study, there were
performance. 14 patients in ‘‘responding’’ group and 16 patients in ‘‘non-
In this study, SVM was used to form the model and cross- responding’’ group. After rehabilitation program, changes
validation method was applied to evaluate the performance of pain intensity (VAS/ODI) of 30 LBP patients were
of the model.22 The procedure was shown in Figure 3. The listed in Table 2. In the ‘‘responding’’ group, the VAS
original dataset of SEMG topography was firstly divided decreased from 3.46  1.37 to 1.92  1.32, whereas the
into two parts: training dataset and testing dataset. Then the ODI decreased from 46.21  7.49 to 34.79  8.95. In the
features (RMSD parameters) were respectively extracted ‘‘nonresponding’’ group, the VAS changed from 3.98  1.69
and selected from these two datasets. Applying SVM to 3.96  1.56 and the ODI changed from 47.13  10.24 to
method to the feature data from the training dataset, the 45.50  10.19.
prediction/classification model could be obtained. Finally,
the performance of the model was evaluated by the feature Comparison of RMSD Parameters Between the
data extracted from the testing dataset. Responding and Nonresponding Groups
The SVM algorithm shown in Figure 3 is a supervised The RMSD RA and RMSD RW during asymmetrical trunk-
machine learning algorithm.33,34 The performance of SVM movement in the ‘‘responding’’ group were consistently
model is strongly dependent on the type of kernel function of lower than the one in the ‘‘nonresponding’’ group
SVM.33 Thus, in this study, we used and compared the SVM (Figure 4A-D). Especially during left lateral-bending move-
model in different kernels such as linear, quadratic, poly- ment, there was significant difference (P < 0.05) or even
nomial and radial basis function (RBF) to identify the highly significant difference (P < 0.01) between these two
‘‘responding’’ and ‘‘nonresponding’’ group. groups (Figure 4A).

TABLE 1. Demographic of Patients With Low Back Pain and Healthy Subjects
Responding Group (n ¼ 14) Nonresponding Group (n ¼ 16) Healthy Subjects (n ¼ 48)
n Percent (%) n Percent (%) N Percent (%)
Age, y
20–40 4 28.6 6 37.5 32 66.7
41–60 10 71.4 10 62.5 16 33.3
Sex
Male 10 71.4 9 56.2 33 68.8
Female 4 28.6 7 43.8 15 31.2

1638 www.spinejournal.com November 2017

Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.
DIAGNOSTICS Prognostic Prediction of Functional Restoration Rehabilitation in Chronic LBP  Jiang et al

TABLE 2. Pain Intensity Result of Patients With Low Back Pain


Responding Group (n ¼ 14) VAS/ODI Nonresponding Group (n ¼ 16) VAS/ODI
Case No. Before Rehabilitation After Rehabilitation Before Rehabilitation After Rehabilitation
1 4/48 2/48 2.5/36 1.8/34
2 4.6/48 0.9/38 5/36 5 /31
3 3/46 4/30 3.2/44 3.1/38
4 3/48 2/33 4/54 3/52
5 3.5/50 1.5/46 2.4/48 3.3/42
6 5.2/52 5.2/42 3/38 3.9/44
7 5.5/30 0.9/34 2.1/56 2.1/54
8 2.1/57 1.4/47 5/52 5/56
9 4.7/38 2.1/28 7/60 6/56
10 2.4/42 1/26 1.2/28 1.2/30
11 5/40 3/37 4/42 4/49
12 1/38 0/22 2.4/60 3.3/58
13 3/56 2/38 3.8/42 3.7/38
14 1.5/54 0.9/18 5/40 5/32
15 6/60 6/58
16 7/58 7/56
ODI indicates Oswestry Disability Index; VAS, Visual Analog Pain Rating Scale.

Comparison Among Model Performance in Different Owing to different kernels in SVM model, we needed to
SVM Models compare the model performance in different kernels. The
Although the 20 RMSD feature parameters including changing curves of accuracy with varying number of SFFS-
RMSD RA and RMSD RW during symmetrical flexion- selected features in different kernel’s SVM model were
extension trunk-movement had been extracted, it did not shown in Figure 5. The highest prediction accuracy and
mean the performance of SVM model based on all 20 the corresponding sensitivity/specificity in different kernels
features was optimal. Some features might be redundant. were shown in Table 3.
Therefore, the SFFS algorithm and LOOCV algorithm were From Figure 5, it was obvious that the optimal number
applied to select the optimal feature parameters. of features in Quadratic kernel was 8 with the highest

Figure 4. Comparison of root-mean-square differ-


ence parameter among the ‘‘responding’’ and
‘‘nonresponding’’ groups during different asym-
metrical trunk-movements (P < 0.05, P < 0.01).
Spine www.spinejournal.com 1639
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.
DIAGNOSTICS Prognostic Prediction of Functional Restoration Rehabilitation in Chronic LBP  Jiang et al

In previous studies, the association between standardized


sociodemographic, physical or psychological variables,
and treatment response status was often investi-
gated.10,12,14,15,37,38 Generally, it was found that some
variables (e.g., age, straight-leg raise, prone instability test,
aberrant motions, lumbar hypermobility, early pain inten-
sity, fear-avoidance beliefs, and kinesiophobia) had corre-
lation with the change of treatment response status.
However, no consistent evidence proved that some psycho-
logical variables (e.g., somatization, hypochondriasis,
anxiety, and depression) assessed by Middlesex Hospital
Figure 5. Prediction performance of SVM model with different ker- Questionnaire (MHQ)39 and Multiphasic Personality Inven-
nels using different numbers of SFFS-selected suboptimal features. tory (MMPI and MMPI-2) scale40 are correlated with the
Linear indicates linear kernel; Quadratic, quadratic kernel; Poly- treatment outcome.37,38,41,42 To obtain the prediction per-
nomial, polynomial kernel (order of the polynomial kernel is 3); Rbf,
formance, the ROC analysis and/or regression analysis was
RBF kernel (the scaling factor, sigma, in the RBF kernel is 1).
often applied.14,15 Hicks et al14 showed the cutoff score
(sensitivity/specificity) of the prediction model from ROC
prediction accuracy (accuracy: 96.67%) and the smallest curve, but unfortunately the values of sensitivity and speci-
number of features. Meanwhile, the corresponding sensi- ficity were often not simultaneously high. Berglund et al15
tivity and specificity were, respectively, 100.00% and showed the adjusted R2 as a measurement of the correlation,
93.75%, as shown in Table 3. The eight optimal feature but the value was often even <0.5. This study showed high
parameters were RMSD RA (flexion) during symmetrical prediction accuracy (96.67%) with high corresponding sen-
flexion-extension trunk-movement, RMSD RW (extension) sitivity (100%) and specificity (93.75%).Thus, the variables
during left lateral-bending, RMSD RA (extension)/RMSD from dynamic SEMG topography based on SVM algorithm
RW (flexion) during right lateral-bending, RMSD RA (turn- provided a better prediction with higher accuracy as shown
ing)/RMSD RW (returning) during left turning and RMSD in this study.
RW (turning/returning) during right turning. Therefore, the In our previous study, the features from SEMG topog-
SVM model in Quadratic kernel with eight optimal features raphy during symmetrical trunk-movement had been pre-
showed the best prediction performance. liminarily investigated for prediction.17 The prediction
Meanwhile, the average AUC from ROC curves in differ- performance was obtained from the ROC analysis and
ent SVM models based on these optimal features could be the largest AUC was 0.747. In addition to SEMG topog-
obtained. Owing to the small sample size, we applied the 10- raphy in symmetrical trunk-movement, the SEMG features
fold cross-validation (10 times) to form the ROC curves. from asymmetrical trunk-movement were included into
The result was shown in the Table 4. SVM algorithm to increase the prediction accuracy. Then,
we obtained the average AUC of >0.8 with a prediction
DISCUSSION accuracy of >90%. Therefore, this study showed a better
This study proposed SVM method to identify patients who choice to identify ‘‘responding’’ and ‘‘nonresponding’’
would respond to the 12-week rehabilitation program via group.
dynamic SEMG topography. The classification of features Another issue was the feature selection. SFFS was used to
was from the dynamic SEMG topography during both reduce the number of features and optimize the prediction
symmetrical and asymmetrical trunk-movements. Through performance. As a method to select optimal features, SFFS
SFFS and cross-validation, the number of features was was successfully used in other medical application.31,43 Tan
optimized and the Quadratic kernel was confirmed as the et al31 used SFFS and SVM to improve radiologists’ per-
best kernel of SVM with the highest prediction accuracy. formance in classification between malignant and benign
Thus, this study is helpful for therapists/doctors to recom- breast lesions. Lin et al43 developed a two-stage model for
mend the appropriate treatment for each LBP patient. the detection and diagnosis of injury of the supraspinatus

TABLE 3. Model Performances (Highest Accuracy and the Corresponding Sensitivity/Specificity)


Based on LOOCV
SVM Kernel Highest Accuracy (%) Sensitivity (%) Specificity (%)
Linear 80.00 85.71 75.00
Quadratic 96.67 100.00 93.75
Polynomial (order is 3) 93.33 92.86 93.75
RBF (sigma is 1) 86.67 85.71 87.50
LOOCV indicates leave-one-out cross-validation; RBF, radial basis function; SVM, support vector machine.

1640 www.spinejournal.com November 2017

Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.
DIAGNOSTICS Prognostic Prediction of Functional Restoration Rehabilitation in Chronic LBP  Jiang et al

TABLE 4. Model performances (average AUC) Supplemental digital content is available for this article.
with SFFS-selected optimal features Direct URL citations appearing in the printed text are
provided in the HTML and PDF version of this article on
based on 10-fold cross-validation (10
the journal’s Web site (www.spinejournal.com).
times)
SVM Kernel Average AUC References
1. Hoy D, Bain C, Williams G, et al. A systematic review of the global
Linear 0.7825
prevalence of low back pain. Arthritis Rheum 2012;64:2028–37.
Quadratic 0.8925 2. Walker BF. The prevalence of low back pain: a systematic review
Polynomial (order is 3) 0.9675 of the literature from 1966 to 1998. Clin Spine Surg 2000;13:
RBF (sigma is 1) 0.7900 205–217.
3. Lehmann TR, Russell DW, Spratt KF, et al. Efficacy of electro-
AUC indicates area under the curve; RBF, radial basis function; SFFS, acupuncture and tens in the rehabilitation of chronic low back pain
sequential floating forward selection; SVM, support vector machine.
patients. Pain 1986;26:277–90.
4. Middelkoop M, Rubinstein SM, Kuijpers T, et al. A systematic
review on the effectiveness of physical and rehabilitation inter-
ventions for chronic non-specific low back pain. Eur Spine J
tendon by SFFS-selected features. The classifier’s perform- 2010;20:19–39.
ance constructed by SFFS-selected features was often 5. Mirza SK, Deyo RA. Systematic review of randomized trials
comparing lumbar fusion surgery to nonoperative care for treatment
superior to those using all features in above two appli-
of chronic back pain. Spine (Phila Pa 1976) 2007;32:816–23.
cations. Meanwhile, compared with other feature selection 6. Mayer TG, Barnes D, Kishino ND, et al. Progressive isoinertial
methods such as Branch and Bound (BB), Sequential Feature lifting evaluation. I. A standardized protocol and normative data-
Selection (SFS), the computational time, and optimized base. Spine (Phila Pa 1976) 1988;13:993–7.
7. Mayer TG, Barnes D, Nichols G, et al. Progressive isoinertial
performance were better.44–46 In this study, the result based lifting evaluation. II. A comparison with isokinetic lifting in a
on SFFS showed the best prediction performance when the disabled chronic low-back pain industrial population. Spine (Phila
number of features was about half. Therefore, it coincided Pa 1976) 1988;13:998–1002.
with previous study. 8. Richards MC, Ford JJ, Slater SL, et al. The effectiveness of
physiotherapy functional restoration for post-acute low back pain:
There are still some limitations in this study. The average a systematic review. Manual Ther 2013;18:4–25.
age of the healthy subjects was younger than the LBP 9. Caby I, Olivier N, Janik F, et al. A controlled and retrospective
patients because it was difficult to recruit patients older study of 144 chronic low back pain patients to evaluate the
than 45 years. Besides, the sample size of the LBP patient effectiveness of an intensive functional restoration program in
France. Healthcare (Basel) 2016;4:E23.
was a bit small. More patients need to be recruited to form a 10. Tong HC, Geisser ME, Ignaczak AP. Ability of early response to
large scale of sample in the future. Finally, it needs further predict discharge outcomes with physical therapy for chronic low
development of a user-friendly prediction tool for clinical back pain. Pain Pract 2006;6:166–70.
application. 11. Bendix AF, Bendix T, Hæstrup C. Can It be predicted which
patients with chronic low back pain should be offered tertiary
rehabilitation in a functional restoration program?: A search for
CONCLUSION demographic, socioeconomic, and physical predictors. Spine (Phila
This study provides a new tool with high accuracy to predict Pa 1976) 1998;23:1775–83.
12. Wong AYL, Parent EC, Funabashi M, et al. Do various baseline
the prognosis of functional restoration rehabilitation in LBP characteristics of transversus abdominis and lumbar multifidus
patients, by using the topography’s feature parameters in predict clinical outcomes in nonspecific low back pain? A system-
both symmetrical and asymmetrical trunk-movement. It can atic review. Pain 2013;154:2589–602.
help clinicians to recommend the appropriate person with 13. van Hooff ML, Spruit M, O’Dowd JK, et al. Predictive factors for
successful clinical outcome 1 year after an intensive combined
LBP to receive 12-week rehabilitation program. physical and psychological programme for chronic low back pain.
Eur Spine J 2014;23:102–12.
14. Hicks GE, Fritz JM, Delitto A, et al. Preliminary development of a
Key Points clinical prediction rule for determining which patients with low
back pain will respond to a stabilization exercise program. Arch
SVM algorithm can be used to predict the Phys Med Rehabil 2005;86:1753–62.
prognosis of functional restoration rehabilitation 15. Berglund L, Aasa B, Hellqvist J, et al. Which patients with low back
pain benefit from deadlift training? J Strength Cond Res 2015;
in LBP patients.
29:1803–11.
Through dealing with the dynamic SEMG 16. Kamen G. Electromyographic Kinesiology. In: Robertson E,
topography, SVM algorithm can be used to editor. Research Methods in Biomechanics. Champaign, IL:
predict the prognosis of functional restoration Human Kinetics Publication; 2004.
rehabilitation in LBP patients. 17. Hu Y, Kwok JW, Yuk-Hang Tse J, et al. Time-varying surface
electromyography topography as a prognostic tool for chronic low
Through dealing with the dynamic SEMG back pain rehabilitation. Spine J 2014;14:1049–56.
topography during symmetrical and 18. Hu Y, Mak JNF, Luk KDK. Effect of electrocardiographic con-
asymmetrical trunk-movement, SVM algorithm tamination on surface electromyography assessment of back
can be used to predict the prognosis of muscles. J Electromyogr Kinesiol 2009;19:145–56.
19. Hu Y, Siu SHF, Mak JNF, et al. Lumbar muscle electromyographic
functional restoration rehabilitation in LBP dynamic topography during flexion-extension. J Electromyogr
patients. Kinesiol 2010;20:246–55.
Spine www.spinejournal.com 1641
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.
DIAGNOSTICS Prognostic Prediction of Functional Restoration Rehabilitation in Chronic LBP  Jiang et al

20. Mak JNF, Hu Y, Cheng ACS, et al. Flexion-relaxation ratio in 35. Wong T-T. Performance evaluation of classification algorithms
sitting: application in low back pain rehabilitation. Spine (Phila Pa by k-fold and leave-one-out cross validation. Pattern Recog
1976) 2010; 35:1532–8. 2015;48:2839–46.
21. Mak JNF, Hu Y, Luk KDK. An automated ECG-artifact removal 36. Bradley AP. The use of the area under the ROC curve in the
method for trunk muscle surface EMG recordings. Med Eng Phys evaluation of machine learning algorithms. Pattern Recog 1997;
2010;32:840–8. 30:1145–59.
22. Cortes C, Vapnik V. Support-vector networks. Machine Learning 37. Ramond A, Bouton C, Richard I, et al. Psychosocial risk factors for
1995;20:273–97. chronic low back pain in primary care—a systematic review. Fam
23. Vlaeyen JWS, Seelen HAM, Peters M, et al. Fear of movement/ Practic 2011;28:12–21.
(re)injury and muscular reactivity in chronic low back pain 38. van der Hulst M, Vollenbroek-Hutten MMR, Ijzerman MJ. A
patients: an experimental investigation. Pain 1999;82:297–304. systematic review of sociodemographic, physical, and psychologi-
24. Sullivan MJL, Thibault P, Andrikonyte J, et al. Psychological cal predictors of multidisciplinary rehabilitation—or, back school
influences on repetition-induced summation of activity-related pain treatment outcome in patients with chronic low back pain. Spine
in patients with chronic low back pain. PAIN 2009;141:70–8. (Phila Pa 1976) 2005;30:813–25.
25. Williamson A, Hoggart B. Pain: a review of three commonly used 39. Julkunen J, Hurri H, Kankainen J. Psychological factors in
pain rating scales. J Clin Nurs 2005;14:798–804. the treatment of chronic low back pain. Follow-up study of
26. Childs JD, Piva SR, Fritz JM. Responsiveness of the numeric pain a back school intervention. Psychother Psychosom 1988;50:
rating scale in patients with low back pain. Spine (Phila Pa 1976) 173–81.
2005;30:1331–4. 40. Love AW, Peck CL. The MMPI and psychological factors in
27. Hägg O, Fritzell P, Nordwall A. The clinical importance of changes chronic low back pain: a review. Pain 1987;28:1–12.
in outcome scores after treatment for chronic low back pain. Eur 41. Haazen IWCJ, Vlaeyen JWS, Kole-Snijders AMJ, et al. Behavioral
Spine J 2003;12:12–20. rehabilitation of chronic low back pain: searching for predictors of
28. Powell W. Approximate dynamic programming: solving the curses treatment outcome. J Rehabil Sci 1994;7:34–43.
of dimensionality, 2nd ed New York: Wiley-Interscience; 2007. 42. Vendrig AA, Derksen JJL, de Mey HR. Utility of selected MMPI-2
29. Pudil P, Novovičová J, Kittler J. Floating search methods in feature scales in the outcome prediction for patients with chronic back
selection. Pattern Recognition Lett 1994;15:1119–25. pain. Psychol Assess 1999;11:381–5.
30. Somol P, Pudil P, Novovičová J, et al. Adaptive floating search 43. Lin C-C, Wang C-N, Ou Y-K, et al. Combined image enhance-
methods in feature selection. Pattern Recognition Letters 1999; ment, feature extraction, and classification protocol to improve
20:1157–63. detection and diagnosis of rotator-cuff tears on MR imaging. Magn
31. Tan M, Pu J, Zheng B. Optimization of breast mass classification Reson Med Sci 2014;13:155–66.
using sequential forward floating selection (SFFS) and a support 44. Jain A, Zongker D. Feature selection: evaluation, application, and
vector machine (SVM) model. Int J Comput Assist Radiol Surg small sample performance. IEEE Trans Pattern Anal Mach Intell
2014;9:1005–20. 1997;19:153–8.
32. Qiu Z, Jin J, Lam H-K, et al. Improved SFFS method for channel 45. Mitra P, Murthy CA, Pal SK. Unsupervised feature selection using
selection in motor imagery based BCI. Neurocomputing. feature similarity. IEEE Trans Pattern Anal Mach Intell 2002;
33. Bishop C. Pattern Recognition and Machine Learning. New York, 24:301–12.
NY: Springer; 2007; 325–358. 46. Sengupta D, Aich I, Bandyopadhyay S. Feature selection using
34. Mitchell TM. Machine Learning. New York, NY: McGraw-Hill feature dissimilarity measure and density-based clustering: Appli-
Education; 1997; 1–19. cation to biological data. J Biosci 2015;40:721–30.

1642 www.spinejournal.com November 2017

Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.
View publication stats

You might also like