Journal of Rehabilitation Research & Development (JRRD)

Quick Links

  • Health Programs
  • Protect your health
  • Learn more: A-Z Health
Veterans Crisis Line Badge
 

Volume 53 Number 6, 2016
   Pages 1007 — 1022

Quantitative assessment of hand motor function in cervical spinal disorder patients using target tracking tests

Sunghoon I. Lee, PhD;1* Alex Huang, BS;2 Bobak Mortazavi, PhD;1 Charles Li, BS;1 Haydn A. Hoffman, BS;2 Jordan Garst, BS;2 Derek S. Lu, BS;2 Ruth Getachew, BS;2 Marie Espinal;2 Mehrdad Razaghy, BS;2 Nima Ghalehsari, BS;2 Brian H. Paak, BS;2 Amir A. Ghavam;2 Marwa Afridi;2 Arsha Ostowari;2 Hassan Ghasemzadeh, PhD;1 Daniel C. Lu, MD, PhD;2–3 Majid Sarrafzadeh, PhD1

1Computer Science Department, University of California Los Angeles (UCLA), Los Angeles, CA; Departments of 2Neurosurgery and 3Orthopedic Surgery, UCLA, Los Angeles, CA

Abstract — Cervical spondylotic myelopathy (CSM) is a chronic spinal disorder in the neck region. Its prevalence is growing rapidly in developed nations, creating a need for an objective assessment tool. This article introduces a system for quantifying hand motor function using a handgrip device and target tracking test. In those with CSM, hand motor impairment often interferes with essential daily activities. The analytic method applied machine learning techniques to investigate the efficacy of the system in (1) detecting the presence of impairments in hand motor function, (2) estimating the perceived motor deficits of patients with CSM using the Oswestry Disability Index (ODI), and (3) detecting changes in physical condition after surgery, all of which were performed while ensuring test-retest reliability. The results based on a pilot data set collected from 30 patients with CSM and 30 nondisabled control subjects produced a c-statistic of 0.89 for the detection of impairments, Pearson r of 0.76 with p < 0.001 for the estimation of ODI, and a c-statistic of 0.82 for responsiveness. These results validate the use of the presented system as a means to provide objective and accurate assessment of hand motor function impairment and surgical outcomes.

Key words: cervical spondylotic myelopathy, classifier, hand impairment, hand movement, machine learning, motor deficit, patient monitoring, quantification, spinal cord disorder, tracking test.

Abbreviations: ADL = activities of daily living, CIDP = chronic inflammatory demyelinating polyneuropathy, CSM = cervical spondylotic myelopathy, ICC = intraclass correlation coefficient, JOA = Japanese Orthopedic Association (scale), LDA = linear discriminant analysis, LOSOCV = leave-one-subject-out cross-validation, MAD = mean absolute difference, MAE = mean absolute error, mJOA = Modified Japanese Orthopedic Association (scale), MLR = multivariate linear regression, MVC = maximum voluntary contraction, NDI = Neck Disability Index, ODI = Oswestry Disability Index, QDA = quadratic discriminant analysis, ROC = receiver operating characteristic, SVM = support vector machine, SVR = support vector regression, TNR = true negative rate, TPR = true positive rate, UCLA = University of California Los Angeles.
*Address all correspondence to Sunghoon I. Lee, PhD; College of Information and Computer Sciences, University of Massachusetts, Amherst, MA, 01003; 413-545-3968; fax: 413-545-1249. Email: silee@cs.umass.edu
http://dx.doi.org/10.1682/JRRD.2014.12.0319
INTRODUCTION

Cervical spondylotic myelopathy (CSM) is a degenerative spinal disorder in the cervical (i.e., neck) region. It is the most common spinal cord dysfunction in adults over 50 yr of age in North America [1–2]. Chronic disc degeneration, inflammatory diseases, or other soft tissue abnormalities caused by CSM often result in significant pressure on the spinal cord or nerve roots [1]. A major complaint of patients with CSM is the impairment of hand motor function [3], including symptoms such as loss of dexterity, numbness, stiffness, weakness, fatigue, and tremor. Specifically, previous studies in patients with CSM observed force overshoot during the initiation phase of gripping followed by an immediate correction response [4–5], which may significantly restrict fine hand motor control. This overshooting response results from exaggerated command signals adopted to compensate for biomechanical changes due to chronic cervical spinal cord injury, and it was further shown that cervical decompression surgery attenuates these overshooting responses [4]. These symptoms may develop into severe weakness or complete paralysis of hand movements [6]. Thus, frequent monitoring of physical conditions in patients with CSM is essential for assessing the impairment level, evaluating the results of medical treatment, and preventing possible onset of impairments.

Unfortunately, frequent radiographical testing (e.g., X-ray or magnetic resonance imaging) is extremely costly. Consequently, current methods for clinically assessing the progress or level of impairment rely on patient-reported outcomes such as the Oswestry Disability Index (ODI) [7] or Japanese Orthopedic Association (JOA) [8] measurements. However, these methods suffer from variability among responders and, most importantly, are known to carry response shift. Response shift refers to changes in an individual's internal standard of perceived health status, which often occurs after treatment such as a surgical intervention [9]. This reduces the reliability of using these methods for longitudinal tracking of patient progress [10–11].

Consequently, a simple, inexpensive, objective, and reliable assessment method for quantifying the physical condition of patients with CSM is needed [12]. Handgrip motor function has received attention as an area of focus for such an assessment method [13] since functional impairment of the hand closely relates to the quality of life of spinal disorder patients and to their ability to perform activities of daily living (ADL), such as eating, writing, or picking up small objects [12]. For instance, Sisto and Dyson-Hudson investigated various methods for using handgrip strength to determine mobility and self-care ability [12]. However, their work only considered the patients' abilities to exert a certain level of grip force rather than their abilities to control their fine hand movement [12], which has a greater correlation with ADL [14–16].

A target tracking test using handgrip force investigates patients' abilities to control their fine hand movement. Its clinical effectiveness has been well studied in other conditions, such as stroke [17–19], Parkinson disease [18,20], brain injury [14,21], and chronic inflammatory demyelinating polyneuropathy (CIDP) [19]. The target tracking test visualizes a predefined waveform that a subject must track by adjusting handgrip force in order to minimize the error between the waveform and the subject's response. Kurillo et al. showed that the tracking error was significantly larger in patients with neuromuscular diseases (e.g., stroke) than control subjects [18]. Lee et al. showed that patients' responses from the target tracking test contained motor characteristics that were specific to conditions such as stroke and CIDP [19]. Getachew et al. investigated the use of target tracking to quantify the level of hand impairment in patients with chronic spinal cord disorder, but their method employed a rather simple, single-dimensional metric (i.e., tracking error), which had considerable limitations in examining various aspects of hand impairments and their correlations to comprehensive quantification of motor deficits [22].

The goal of this work is to thoroughly investigate performance characteristics of the target tracking test in quantifying hand motor deficits in patients with CSM using machine learning techniques. Unlike previous works that used a simple and comprehensive metric, the method used in this work incorporates machine learning algorithms that mathematically combine multiple metrics (features) that are designed to represent known symptoms of CSM. We hypothesized that the quantification enabled by incorporating the target tracking test and machine learning techniques has the potential to serve as an effective screening and monitoring tool for hand motor function in patients with CSM. This work specifically aimed to investigate a number of important criteria for validating the medical efficacy of the target tracking test [23], including (1) detecting the presence of hand motor impairment and quantifying its severity [24], (2) estimating perceived motor deficits in performing daily activities using the ODI [7,25–26], and (3) detecting changes in physical conditions of patients after receiving surgical intervention (i.e., responsiveness [27–28]), all while ensuring the test-retest reliability of the quantification.

METHODS
Participants

A total of 30 patients with CSM (18 males), mean age 59.5 ± 16.0 yr, were recruited from the University of California Los Angeles (UCLA) Spine Center, excluding younger patients (<45 yr) and those with comorbidities affecting their hand motor function. All patients' evidence of CSM (e.g., location or level of the cervical spinal injury) was verified using conventional X-ray imaging. The average duration of overall back pain (both cervical and lumbar) of the participating patients was 49.9 mo, and the average duration of arm pain was 9.61 mo (until the date they received the surgical intervention). Radiculopathy (i.e., pinched nerve root) was found in 72.4 percent of the patients, while 68.9 percent had severe neck pain. All patients received spinal cord decompression to alleviate pressure on their impacted nerve roots, improving associated pain and motor function. The surgical intervention was performed by a single neurosurgeon, Daniel C. Lu (one of the coauthors of this article). Of the 30 patients, 17 returned to the clinic for follow-up within 3 mo of surgery. A total of 30 age-matched control subjects (14 male), mean age 57.5 ± 9.2 yr, were recruited from the general population.

Examination Protocol

This work used a handgrip device that has been described previously (Figure 1) [5]. The major components of the handgrip device include the springs, the handle, and the displacement sensor embedded in the body. The handle of the device was connected to the main frame by three springs, which provided physical resistance for grasping performance. The length of the handle could be customized using the adjustable pins to accommodate subjects with varying hand sizes. The springs could also be replaced to accommodate participants with varying grip strengths. This study used five springs with different tension forces: 0.38 lbs/in., 0.88 lbs/in., 1.94 lbs/in., 5.10 lbs/in., and 10.7 lbs/in. The displacement sensor was embedded in the bottom of the frame, and it captured the absolute position of the handle at a sampling rate of 32 Hz.

Participants started the test by measuring their maximum voluntary contraction (MVC), which represented the maximum grip force that participants could voluntarily exert. The measured MVC was used to normalize the maximum amplitude of the target waveform such that participants' fine motor function (rather than their absolute grip strength) could be investigated. The unit of the waveform amplitude was percent MVC. The waveform within the screen moved to the left while the horizontal position of the circle was fixed in the middle of the x-axis as shown in Figure 1(a). The vertical position of the circle changed according to the grip force generated by the participants. The screen also displayed a trace history of the patient's response for visual feedback. The length of the test was 45 s.


Figure 1. (a) Handgrip device and the two tracking waveforms used in this study (left: step; right: sine). (b) Patient with cervical spondylotic myelopathy performing the test before her surgical operation.

Figure 1.

(a) Handgrip device and the two tracking waveforms used in this study (left: step; right: sine). (b) Patient with cervical spondylotic myelopathy performing the test before her surgical operation.

Click Image to Enlarge. View as PowerPoint Slide

Participants were tested using two different targets: sine and step waveforms. The sine waveform had a period of 6.17 s (0.16 Hz), which resulted in approximately seven sine cycles per test. The amplitude of the waveform changed from 0 to 100 percent of the subject's MVC, as illustrated in Figure 2. The sine waveform investigated participants' ability to predict and control the muscle movements required for grasping performance to be repeated at a constant rate [29]. The step waveform had a period of 3 s (0.33 Hz), with 50 percent duty cycle, which resulted in 15 cycles per test (Figure 2). The higher amplitude was equal to 80 percent MVC, and the lower amplitude was equal to 20 percent MVC. The step waveform investigated predictive tracking and the ability to produce handgrip force at a constant velocity [29]. Participants repeated each waveform three times per clinical visit, generating a total of six test results. The UCLA Institutional Review Board approved the examination procedure, and all participants provided consent after an explanation of the study protocol and the associated risks.


Figure 2. Sample test results collected from (a) patients with cervical spondylotic myelopathy before surgery and (b) age-matched control subjects. MVC = maximum voluntary contraction.

Figure 2.

Sample test results collected from (a) patients with cervical spondylotic myelopathy before surgery and (b) age-matched control subjects. MVC = maximum voluntary contraction.

Click Image to Enlarge. View as PowerPoint Slide

Patient-Reported Functional Outcomes

Patients with CSM reported their perceived level of motor impairment in performing ADL using the ODI [7] and the modified JOA (mJOA) at their preoperative and postoperative visits. ODI is one of the well-known measures of perceived motor function and quality of life for patients with spinal injuries [7,30]. It contains 10 multiple-choice items assessing the degree of interference from pain in performing various daily activities such as personal care, lifting, walking, sitting, standing, sleeping, sex life, social life, and traveling. The accumulated score ranges from 0 (no dysfunction) to 50 (completely disabled). In this work, the accumulated score was linearly scaled in reverse from 0 (completely disabled) to 1 (no dysfunction) to comply with the general systems performance theory that all dimensions of human performance should be in a form for which a higher numerical value represented superior performance. The mJOA is another well-known measure of motor deficits in patients with CSM [31]. This survey contains four multiple-choice items relating to upper-limb and hand motor function. The accumulated score of the mJOA ranges from 0 (completely disabled) to 18 (no dysfunction), which was again linearly scaled to 0 to 1.

Other Clinical Variables

Ten clinical variables that may have close correlation to hand motor impairment were collected. These variables included age, sex, overall back pain duration (how long had the current episode of back pain, both in cervical and lumbar areas, been present?), arm pain duration (how long had the current episode of arm pain been present?), presence of severe neck pain, presence of radiculopathy, herniated disks (i.e., C1 to C7), smoking or nonsmoking, packs of cigarettes smoked per year, and alcohol consumption (drinks per week). For nondisabled control subjects, age and sex information were collected.

Data Analysis

Figure 3 shows a schematic representation of the data analytics that were used to address the aforementioned objectives in a test-retest reliability manner. Impairment was detected by investigating the ability to differentiate the handgrip data collected from 30 preoperative patients with CSM (with impairments) from the data collected from 30 age-matched control subjects (without impairments). The estimation of the perceived motor deficits investigated the ability to estimate the ODI scores using handgrip data based on the data collected from 30 preoperative and 17 postoperative patients (i.e., a total of 47 data points from 30 patients). Finally, responsiveness investigated the ability to differentiate the patients whose perceived motor functions improved after the surgical intervention from those who did not improve. These objectives were addressed using the same data analytic platform, except that detection of impairments employed a binary classifier, estimating the perceived motor deficits employed a regressor, and responsiveness employed a binary classifier. For all objectives, employing a preprocessing step that eliminated those features that were unreliable ensured the test-retest reliability. All the components of the data analytics summarized in Figure 3 are discussed in detail in the following subsections. Note that this work employed a leave-one-subject-out cross-validation (LOSOCV) to provide a fair evaluation of the quantification results; the data belonging to a subject was left out as a testing set, which was evaluated based on the classification/regression model constructed using the training data set belonging to the rest of the subjects. This approach avoids problems of overfitting and provides fair estimations of the expected diagnostic and estimation accuracy for the binary classifier and the regressor, respectively [32].


Figure 3. Schematic representation of the data analyses that were used to address the aforementioned objectives: (1) ensuring the test-retest reliable of the quantification, (2) detecting the presence of hand motor impairment and quantifying the severity, (3) estimating perceived motor deficits in performing activities of daily living, and (4) responsiveness. Pre-op = preoperative, Post-op = postoperative.

Figure 3.

Schematic representation of the data analyses that were used to address the aforementioned objectives: (1) ensuring the test-retest reliable of the quantification, (2) detecting the presence of hand motor impairment and quantifying the severity, (3) estimating perceived motor deficits in performing activities of daily living, and (4) responsiveness. Pre-op = preoperative, Post-op = postoperative.

Click Image to Enlarge. View as PowerPoint Slide

Hand Motor Features
Sine waveform

The following features were extracted from the sine waveform: Mean absolute error (MAE) (MAE-SINE) computed the average error between the target sine waveform and the patient's response over the length of the signals: ∑k | wt [k] – wr [k] |, where wt and wr represented the target waveform and the patient's response, respectively. The level of the overshooting response during the initiation phase of gripping was uniquely observed in patients with CSM [4–5]. We further extended this finding by also considering the overshooting response during the release phase, as illustrated in Figure 4. The overshooting response from initiating gripping was quantified using two features: FirstQMeanErr and FirstQMaxErr. FirstQMeanErr and FirstQMaxErr computed the MAE and the maximum error between the target waveform and the patient's response during the first quarter of a sine cycle, respectively. The first quarter of the sine cycle, as annotated in Figure 4, was used to estimate the time period of initiating the gripping action. ThirdQMeanErr and ThirdQMaxErr were derived in a similar manner as FirstQMeanErr and FirstQMaxErr to quantify the overshooting response when releasing the grip during the third quarter. PhaseShift was designed to estimate how quickly patients reacted to correct the error caused by the overshooting response. PhaseShift computed the time to reach the grip force that was equal to the amplitude of the target waveform at the end of each quarter (i.e., 50% MVC). Since a single test contained seven sine cycles, features related to the overshooting response were averaged over the seven cycles. Muscle fatigue, which is defined as the temporary inability of muscles to perform optimally [33], was assessed using two measures: LAST-PK-SINE and ΔPK-SINE. LAST-PK-SINE computed the peak amplitude of the patient's response during the last sine cycle. ΔPK-SINE computed the difference in the peak amplitudes between the first and the last sine cycles. Tremor was quantified using the following three features, which were all computed in the frequency domain: 2ndFreq, ΔFreq, and ΔGain. 2ndFreq computed the frequency with the second largest gain after the fundamental frequency of the patient-generated waveform; the fundamental frequency should be very close to that of the target waveform (i.e., 0.16 Hz). ΔGain was the gain difference between the two frequencies. ΔFreq computed the difference between the fundamental frequencies of the target waveform and the patient-generated waveform.


Figure 4. Example of over-reaction responses that are illustrated with a sudden peak at the initiation and a sudden drop at the releasing of grip forces. MVC = maximum voluntary contraction.

Figure 4.

Example of over-reaction responses that are illustrated with a sudden peak at the initiation and a sudden drop at the releasing of grip forces. MVC = maximum voluntary contraction.

Click Image to Enlarge. View as PowerPoint Slide

Step waveform

A total of 12 features were extracted from the step waveform. MAE (MAE-STEP) was computed to represent the comprehensive motor capacity under step waveform. VEL-INC, AMP-INC, VEL-DEC, and AMP-DEC were used to investigate how fast a patient can exert and release the submaximal grip force. VEL-INC calculated the velocity of the grip force (% MVC per second) when switching from the lowest (20% MVC) to the highest grip force (80% MVC). AMP-INC represented the maximum difference in the amplitudes during this initiation phase of gripping. VEL-DEC and AMP-DEC were computed in a similar manner when releasing the grip force. Muscle endurance, defined as the ability to sustain repeated contractions against a resistance for an extended period of time [33], was quantified using AVG-HIGHEST, STD-HIGHEST, AVG-LOWEST, and STD-LOWEST. AVG-HIGHEST and STD-HIGHEST computed the mean and the standard deviation of the amplitude of the patient's waveform while maintaining the highest grip force (80% MVC). Similarly, AVG-LOWEST and STD-LOWEST computed the mean and the standard deviation when the subject was maintaining the lowest grip force (20% MVC). Muscle fatigue was assessed similarly to that of the sine waveform. LAST-PK-STEP computed the mean amplitude of the patient's response during the highest grip force (80% MVC) of the last step cycle. ΔPK-STEP computed the difference in the mean amplitudes of the highest grip force between the first and last step cycles.

Eliminating Unreliable Features

Employing a preprocessing step that examined the test-retest reliability eliminated unreliable features, as introduced in Palmerini et al. [34]. The features extracted from the last two (out of three) tests were used to compute the intraclass correlation coefficient (ICC); the first test was considered a practice trial and was not included in the reliability test. The value of the ICC ranged from 0 to 1, where an ICC < 0.4 indicated poor test-retest reliability, 0.4 ≤ ICC < 0.75 indicated fair to good test-retest reliability, and an ICC ≥ 0.75 indicated excellent test-rest reliability [35]. The ICC values were computed separately for the control, the preoperative CSM, and the postoperative CSM data sets. The features that produced ICC < 0.75 for any of these three data sets were removed from further analyses. This ensured a reliable quantification of motor impairment, since the decision function (i.e., kernel function) of a classifier or a regressor was constructed by mathematically combining features that were test-retest reliable. Palmerini et al. provides more detailed information on this method [34].

Feature Selection

We utilized a wrapper approach for feature selection that evaluated all feature subsets within its feature searching space for their classification/regression performance and selected the subset that produced the best performance [36]. First, the maximum cardinality of a feature subset was constrained based on the data set-to-feature ratio, as suggested by Prichep et al. [37]. For a linear classification model, the minimum data set-to-feature ratio was limited to 10:1, and for a quadratic model, the cardinality of the selected feature set (denoted as F) was limited such that F × (F + 3) / 4 did not exceed the number of subjects of the smallest class [37]. A forward selection was used to construct the feature searching space. The forward selection approach started with an empty feature set and progressively added a feature that produced the best classification/regression performance until its size reached the defined maximum cardinality. Note that the same classifier/regressor, which was followed to address the objective, was used to compute the classification/regression performance. Another layer of a LOSOCV was employed within the training set to evaluate each feature subset throughout the forward selection. The selected feature sets were then used to construct a classification/regression model to address each of the three objectives, which are discussed in the following subsections.

Detection of Impairments in Hand Motor Function

In order to detect the presence of hand motor impairments in patients with CSM, we formulated the impairments problem as a binary classification problem between the handgrip data of 30 preoperative patients with CSM and 30 age-matched control subjects. This allowed us to construct an equation using the selected features to maximize the probability of distinguishing the two groups and thus produce the binary prediction of the presence of impairment (i.e., have impairment or not?) and the severity of impairment (i.e., the posterior probability of having impairment). We investigated three different classification algorithms: support vector machine (SVM) with a linear kernel, linear discriminant analysis (LDA), and quadratic discriminant analysis (QDA). The c-statistic, which is an effective technique for assessing diagnostic and predictive accuracy in disease management [38], was used to evaluate the classification performance [39]. The c-statistic is also known as the area under the receiver operating characteristic (ROC) curve [32] and represents the probability of a randomly selected subject being correctly predicted in his/her class (e.g., has impairment or not?). The ROC curve is a graph of the true positive rate (TPR) (sensitivity) against the false positive rate (1 – specificity), which visualizes the classification performance. The c-statistic ranges from 0.5 (unable to discriminate) to 1.0 (able to perfectly discriminate), where 0.8 is known to represent fairly good discriminatory ability.

The prediction results were compared to the clinical variables introduced in the "Methods: Other Clinical Variables" section in an attempt to find possible relationships between the hand motor patterns detected by the algorithms and existing clinical knowledge. More specifically, the binary predictions made on the patient (positive) data, which included true positives and false negatives, were compared to the 10 clinical variables using a t-test. The posterior probabilities of the predictions were also compared to the clinical information using Pearson linear correlation. The binary predictions made on the control (negative) data and their posterior probabilities were similarly compared to the nondisabled subjects' age and sex using the t-test and Pearson linear correlation, respectively.

Estimation of Perceived Motor Deficits

Estimating the perceived motor deficits was formulated as a regression problem between the handgrip data and the ODI scores of 30 preoperative and 17 postoperative patients. Note that estimating mJOA scores was not investigated in this work, since the values did not show much variability and were highly unbalanced toward a single score; the minimum and the maximum scores were 15 and 18, respectively, and a vast majority of the scores were 16. Furthermore, ODI has a larger number of questions that are closely related to motor functions for ADL. The regression models tested in this work include (1) support vector regression (SVR) with a linear kernel; (2) SVR with a nonlinear, radial kernel; and (3) multivariate linear regression (MLR). The maximum cardinality of a feature set was limited to F = (30/10) = 3 since the data involved 30 patients. The nonlinear SVR also employed this rule for convenience. The estimation accuracy was evaluated using the mean absolute difference (MAD) between the estimated and the actual ODI scores, and the p-value of Pearson linear correlation was used to test the null hypothesis of zero correlation. Then, the actual and estimated ODI scores were compared with the clinical variables using Pearson correlation in order to provide clinical interpretations.

Responsiveness

Responsiveness was formulated as a binary classification problem between the patients whose perceived motor functions improved and those whose perceived motor functions did not improve after the surgical intervention, in a similar manner to that used by Deyo and Centor [28]. The improved patients were defined as those whose postoperative ODI and mJOA scores were both improved compared to their preoperative values. In our study, 12 patients were categorized as improved and 5 patients as not improved. The maximum cardinality of a feature subset was limited to F = (17/10) = 2 for a linear model and F = 3 for a quadratic model. The difference between postoperative and preoperative values was computed for all motor features and was used as the input data. Three different classification algorithms (i.e., SVM, LDA, and QDA) were used as the models, and their classification accuracies were evaluated using the c-statistic. The binary prediction results and their posterior probabilities were compared with the clinical variables using the t-test and Pearson linear correlation in order to find clinical justification.

RESULTS
Test-Retest Reliability of System

Table 1 summarizes the features used in this work. The first column represents the symptoms of CSM that the features were designed to quantify. The second and third columns represent the name and the associated waveform, respectively. The rest of the columns summarize the mean, standard deviation, and ICC values for test-retest reliability for data sets collected from control subjects, preoperative CSM patients, and postoperative CSM patients. LAST-PK-SINE and LAST-PK-STEP showed an ICC < 0.75 for the data set collected from nondisabled subjects and were removed from any further analyses. These two unreliable features were indicated with shading in Table 1.


Table 1. 
1.42 (0.21)
1.46 (0.42)
1.27 (0.16)
66.41 (3.11)
65.71 (10.41)
64.03 (5.77)
1.23 (12.12)
*Full descriptions of the features listed in this table can be found in the "Methods: Data Analysis, Hand Motor Features" section.
Detection of Impairments in Hand Motor Function

The ability of the system to detect the presence of hand motor impairments is summarized in Table 2. QDA outperformed LDA and SVM in classifying patients with CSM with hand motor impairments from control subjects, with a c-statistic of 0.89; the anticipated TPR and true negative rate (TNR) were 0.83 and 0.87, respectively. This implies that the system can discriminate between individuals with hand motor impairments and those without hand motor impairments with an average accuracy of 89 percent. The ROC curve used to compute the c-statistic is illustrated in Figure 5. The detection results were compared with the clinical variables, but no statistical significance was found.


Table 2. 
The c-statistics, expected true positive rates (TPRs), and expected true negative rates (TNRs) for classifying cervical spondylotic myelopathy patients with hand motor impairments from nondisabled control subjects. Quadratic discriminant analysis (QDA) (in bold) produced superior classification performance compared to support vector machine (SVM) and linear discriminant analysis (LDA). The most frequently selected features from the cross-validation included MAE-SINE, FirstQMeanErr, 2ndFreq, PhaseShift, ΔGain, and AVG-HIGHEST.
c-Statistic
*Full descriptions of the features listed in this table can be found in the "Methods: Data Analysis, Hand Motor Features" section.

The most frequently selected feature subset for QDA contained MAE-SINE, FirstQMeanErr, 2ndFreq, PhaseShift, ΔGain, and AVG-HIGHEST. These are the features that significantly contributed in achieving the expected detection accuracy of 89 percent. MAE-SINE, FirstQMeanErr, and 2ndFreq were the most significant features; each of these features showed statistical significance in differentiating the two groups: MAE-SINE showed p < 0.001, FirstQMeanErr showed p < 0.003, and 2ndFreq showed p < 0.007 using a t-test.


Figure 5. Receiver operating characteristic curve for detecting the presence of impairment, generated from the quadratic discriminant analysis-based model.

Figure 5.

Receiver operating characteristic curve for detecting the presence of impairment, generated from the quadratic discriminant analysis-based model.

Click Image to Enlarge. View as PowerPoint Slide

Estimation of Perceived Motor Deficits

SVR with a radial kernel produced the most accurate results in estimating the ODI score (Table 3). The estimated ODI scores were compared to the actual ODI scores and produced MAD = 0.12, which represents the expected estimation error rate. Furthermore, the estimated ODI scores showed a statistically significant correlation to the actual ODI scores with Pearson r = 0.76 and p < 0.001. Figure 6(a) illustrates the scatter plot between the estimated and actual ODI scores, and Figure 6(b) illustrates its Bland-Altman plot, where the bias of the difference was –0.016 and the magnitude of the limit of agreement was 0.27. Clinical variables were compared with the estimated ODI scores as well as with the actual ODI scores. The arm pain duration showed statistically significant correlations to the actual and estimated ODI scores collected postoperatively, with p < 0.006 and p < 0.007, respectively.


Table 3.
Expected estimation performance based on mean absolute difference (MAD), coefficient of determination (R2), and Pearson correlation coefficient (r). Support vector regression (SVR) with a radial kernel (in bold) produced the best estimation results compared to multivariate linear regression (MLR) and SVR with a linear kernel. The most frequently selected features from the cross-validation included FirstQMeanErr, ΔGain, and VEL-INC.
R2

Figure 6. (a) Scatter plot of the estimated and actual Oswestry Disability Index (ODI) scores, which achieved mean absolute difference = 0.12,Pearson r = 0.76, and p < 0.001. (b) Bland-Altman plot of the estimated and actual ODI scores, which achieved a bias of ·0.016 and a limit of agreement of 0.27. SD = standard deviation.

Figure 6.

(a) Scatter plot of the estimated and actual Oswestry Disability Index (ODI) scores, which achieved mean absolute difference = 0.12,Pearson r = 0.76, and p < 0.001. (b) Bland-Altman plot of the estimated and actual ODI scores, which achieved a bias of ·0.016 and a limit of agreement of 0.27. SD = standard deviation.

Click Image to Enlarge. View as PowerPoint Slide

The most frequently selected feature subset for this nonlinear SVR contained FirstQMeanErr, ΔGain, and VEL-INC. Among the selected features, VEL-INC was the most significant feature, with r = 0.42 and p < 0.003 when compared to the ODI scores.

Responsiveness

The ability of the system to detect patients whose perceived motor function improved (or did not improve) after the surgical intervention is summarized in Table 4. Linear SVM produced the best classification accuracy, with a c-statistic of 0.82, expected TPR of 0.92, and expected TNR of 0.80. This result shows that the handgrip system can be used to monitor the changes in perceived motor function with an average accuracy of 82 percent. The ROC curve produced by SVM is illustrated in Figure 7.


Table 4. 
The c-statistics, expected true positive rates (TPRs), and expected true negative rates (TNRs) for classifying improved and nonimproved patients after surgical operation. Support vector machine (SVM) with a linear kernel (in bold) produced the best performance compared to quadratic discriminant analysis (QDA) and linear discriminant analysis (LDA). The most frequently selected features from the cross-validation included ThirdQMeanErr and AVG-LOWEST.
c-Statistic

Figure 7. Receiver operating characteristic curve produced by support vector machine for detecting the changes in physical conditions of patients after surgical operation.

Figure 7.

Receiver operating characteristic curve produced by support vector machine for detecting the changes in physical conditions of patients after surgical operation.

Click Image to Enlarge. View as PowerPoint Slide

The classification results and their associated posterior probabilities were compared to the collected clinical variables. The overall back pain duration showed statistically significant correlation to the prediction results and posterior probabilities with p < 0.004 (t-test) and p < 0.008 (Pearson linear correlation), respectively. This shows that the surgical outcome predicted by the handgrip device has a significant correlation to back pain duration.

The most frequently selected feature set included ThirdQMeanErr and AVG-LOWEST. The changes in ThirdQMeanErr scores for the improved and the not-improved groups were –4.41 and 5.24 (p < 0.004 [t-test]), respectively, which implies that ThirdQMeanErr values for the improved group significantly improved after the surgery compared with the not-improved group. In a similar manner, the improvement in AVG-LOWEST was –3.03 for the improved and –0.77 for the not-improved group (p < 0.08 [t-test]).

DISCUSSION

The aim of this pilot study was to validate the use of the target tracking test in (1) detecting the presence of hand motor impairment, (2) estimating the ODI scores reported by patients, and (3) detecting changes in perceived motor deficits in performing ADL (i.e., responsiveness). These objectives were performed in a test-retest reliable manner by employing a preprocessing step that eliminated unreliable features. The reported results showed acceptable accuracy in detecting the presence of hand motor impairment and in the responsiveness; results demonstrated that the system might be considered for use as a screening tool prior to surgical treatment and as a tool to monitor ailment ­progress over time. The system is easy-to-use and inexpensive, and it takes no more than 5 min to complete a test, which also supports the system's potential to remotely monitor patients in their home and community settings. Estimation of the ODI scores did not show outstanding performance, but the results were comparable to other works. The following subsections will provide detailed discussion regarding each study objective as well as their limitations and planned future work.

Test-Retest Reliability

The results summarized in Table 1 support the test-retest reliability of the motor features extracted from the handgrip device. A total of 22 features were extracted, and 20 of them showed excellent test-retest reliability in both control and patient groups. This indicates that the target tracking test provides reliable representation of one's hand motor function. Two features in the control group did not show good test-retest reliability: LAST-PK-SINE and LAST-PK-STEP, which computed the peak amplitude of the patient's response during the last cycles of the sine and step waveforms, respectively. The most likely reason for this unreliability is that the peak value may be too sensitive to a single value that overrepresents the patient's response during the last cycle. Two patients produced outlying peak values for each of the two features and significantly reduced the ICC. Consequently, these two features were removed from further analyses to ensure the reliability of the classification and regression models for the three objectives.

Detection of Impairments in Hand Motor Function

Results summarized in Table 2 show that the handgrip system can detect the presence of motor impairments in hand movement with an average accuracy of 89 percent (with a TPR of 83% and a TNR of 87%). This indicates that the system has the potential to be used as a diagnostic tool for the CSM population. The reported detection accuracy could be achieved by employing QDA with the following features: MAE-SINE, FirstQMeanErr, 2ndFreq, PhaseShift, ΔGain, and AVG-HIGHEST. Among these features, MAE-SINE, FirstQMeanErr, and 2ndFreq were the most significant features; each individual feature showed statistically significant diagnostic ability in differentiating the CSM and control groups. This suggests that if a clinic has a limited access to advanced algorithms such as QDA, investigating these three features can also provide insights regarding the presence of hand motor impairments.

The detection results showed that five patients were incorrectly classified as having no impairments in hand motor function and four control subjects were incorrectly classified as having hand impairments. These results were compared with various clinical variables ("Methods: Other Clinical Variables" section), but no statistical significance was found. This implies that the objective quantification of the hand motor impairment, obtained by employing the handgrip system and the algorithms, provides unique information that cannot be found in other clinical variables. Thus, we propose that the system can be used as part of diagnostic and screening processes to quantify the level of hand motor function.

Estimation of Perceived Motor Deficits

The handgrip system could estimate the ODI scores in patients with CSM with moderate accuracy, namely MAD = 0.12 and r = 0.76 (Table 3). This is not surprising since the handgrip device quantifies patients' hand motor functions and ODI quantifies the degree of interference of motor impairments in performing various ADL. Nevertheless, the reported correlation results (r = 0.76) were comparable to the findings in other studies. In Grönblad et al., the ODI was compared to the Pain Disability Index and the Visual Analog Scales for pain and achieved r = 0.83 and r = 0.62, respectively [40]. Furthermore, the ODI was compared to the McGill Pain Questionnaire [26], Short Form-36 [41], and Roland-Morris questionnaire [25], and achieved r = 0.62, r = 0.77, and r = 0.66, respectively. Specifically, in Fairbank and Pynsent [7], the Bland-Altman plot between ODI and Roland-Morris showed that the magnitude of the limit of agreement was approximately equal to 0.32, which is comparable to the limit of agreement reported in this work (i.e., 0.27 as shown in Figure 6). Note that the correlation results reported in the aforementioned works [25–26,40–41] used the original ODI scores, which used the scale between 0 and 50, where 0 represented completely nondisabled and 50 represented completely disabled conditions. On the other hand, this work used a reversed scale that ranged from 0 (completed disabled) to 1 (completely nondisabled) as discussed in the "Methods: Patient-Reported Functional Outcomes" section. Nonetheless, the reported results demonstrate that the handgrip device and algorithms can together quantify the hand motor function with close correlation to the level of perceived deficits in performing ADL with acceptable accuracy. The estimated ODI scores, as well as the actual ODI scores that were collected postoperatively, showed statistically significant correlation to arm pain duration (p < 0.007 and p < 0.006, respectively), which agrees with findings in prior work [42].

Table 3 shows that SVR with a radial kernel, which is a nonlinear regression model, performs better than the two linear models (MLR and SVR with a linear kernel). This demonstrates that the relationships between the predictors and the ODI scores can be more accurately described using nonlinear functions, which partially agrees with the findings in Hoffman et al. [42]. The reported estimation accuracy (MAD = 0.12 and r = 0.76) can be achieved when FirstQMeanErr, ΔGain, and VEL-INC are used to construct an SVM-based model. However, if a clinic has limited access to such an advanced regression algorithm, it can also utilize VEL-INC as an estimation factor since it produced a statistically significant correlation to ODI scores with r = 0.42 and p < 0.003.

Responsiveness

Results shown in Table 4 demonstrate that the handgrip device can classify patients whose perceived motor function has improved after their surgical intervention with an average accuracy of 82 percent (the TPR was 92% and the TNR was 80%). This supports that the handgrip device and the target tracking examination have the potential to be used as a monitoring tool that tracks longitudinal changes in perceived motor function. The reported classification accuracy can be achieved by computing ThirdQMeanErr and AVG-LOWEST and constructing a classification model using a SVM with a linear kernel. ThirdQMeanErr showed statistical significance in differentiating the improved and not improved groups (p < 0.004 [t-test]). Thus, if a clinic has limited access to sophisticated algorithms such as SVM, healthcare professionals can investigate the values of ThirdQMeanErr, which can be easily computed from the handgrip device.

The classification results and the associated posterior probabilities showed statistical significance to the overall back pain duration with p < 0.004 (t-test) and p < 0.008 (Pearson linear correlation), respectively. It is especially interesting that the results showed significant correlation to the overall back pain duration rather than the arm or neck pain. This may be because major symptoms of CSM include not only deterioration of hand use but also difficulty in gait; a previous study shows that approximately 75 percent of patients with CSM experience deterioration of hand motor function and 80 percent experience difficulty in gait [43]. Therefore, the overall back pain duration, which is a major prognostic factor for surgical outcomes [44], shows significant correlation to the predicted surgical outcomes.

Limitations and Future Works

This work is the first study to thoroughly evaluate the use of a handgrip device for detecting the presence of ailments, estimating the perceived motor function represented by ODI scores, and measuring the responsiveness to the surgical intervention in patients with CSM. Some limitations deserve discussion. The small sample size makes it difficult to generalize our findings to the general CSM population. However, all the classification and regression performances reported in this article were computed using the LOSOCV technique, which produced a fair estimate rather than an optimistic estimate [32]. Given this, the results reported in this article are promising.

The ODI was employed to represent the perceived motor deficits in performing ADL; ODI scores were estimated using the features extracted from the target tracking tests. As noted previously, both the ODI and mJOA were collected from the participating patients, but only the ODI was used in this work because the mJOA scores of the patients were highly unbalanced. The ODI investigated the degree of interference from pain in performing a number of ADL involving both upper and lower limbs. The ODI has been widely used to assess functionality of patients with CSM in a number of previous studies [7,22,42,45–46], because a vast majority of patients with CSM have complaints about the use of their hands (upper limb) as well as difficulties in gait (lower limb) [43]. However, other patient-reported outcomes such as the Neck Disability Index (NDI) [47] may provide better correlation to hand motor skills that can be captured by the handgrip device. The current study has modified the protocol to collect ODI as well as NDI to find possible correlations, which remains as future work.

CONCLUSIONS

This article introduced a method for quantifying hand motor function using a handgrip device and target tracking tests. Data analytic methods based on machine learning techniques were designed to validate the use of the target tracking test in (1) detecting the presence of impairments in hand motor function and quantifying their severity, (2) estimating the perceived motor deficits of CSM patients that are measured by ODI using the ­features extracted from hand motor function, and (3) detecting the changes in physical condition (improved versus not improved) after surgical decompression by investigating the changes in hand motor function. The estimation results, which were produced from a LOSOCV-based technique, showed a c-statistic of 0.89 for detection of impairments, Pearson r of 0.76 with p < 0.001 for the estimation of ODI, and a c-statistic of 0.82 for responsiveness. This pilot study reports promising results validating the use of a handgrip device and target tracking tests to provide objective quantification of various hand motor functions. This enables new research and development opportunities for more objective, easy-to-use, and inexpensive methods to assess the level of impairment and the surgical outcomes in patients with CSM.

ACKNOWLEDGMENTS
Author Contributions:
Study concept and design: S. I. Lee, D. C. Lu, M. Sarrafzadeh.
Acquisition of data: S. I. Lee, C. Li, H. A. Hoffman, D. S. Lu, R. Getachew, J. Garst, M. Espinal, M. Razaghy, N. Ghalehsari, B. H. Paak, A. A. Ghavam, M. Afridi, A. Ostowari.
Analysis and interpretation of data: S. I. Lee, R. Getachew, B. ­Mortazavi, M. Sarrafzadeh.
Drafting of manuscript: S. I. Lee, H. A. Hoffman, B. Mortazavi.
Critical revision of manuscript for important intellectual content:
B. Mortazavi, A. Huang, H. Ghasemzadeh, D. C. Lu, M. Sarrafzadeh.
Statistical analysis: S. I. Lee.
Study supervision: D. C. Lu, M. Sarrafzadeh.
Financial Disclosures: The authors have declared that no competing interests exist.
Funding/Support: This material was unfunded at the time of manuscript preparation.
Additional Contributions: We acknowledge the assistance of Ms. Naomi Gonzalez for providing logistical support in patient scheduling.
Institutional Review: The study protocol was approved by the UCLA Institutional Review Board (IRB# 12-000009), and all participants signed informed consent forms.
Participant Follow-Up: The authors do not plan to inform participants of the publication of this study.
REFERENCES
1.
Young WF. Cervical spondylotic myelopathy: a common cause of spinal cord dysfunction in older persons. American Fam Phys. 2000;62(5):1064–70,73.
3.
Doita M, Sakai H, Harada T, Nishida K, Miyamoto H, Kaneko T, Kurosaka M. Evaluation of impairment of hand function in patients with cervical myelopathy. J Spinal Disord Tech. 2006;19(4):276–80. [PMID:16778663]
http://dx.doi.org/10.1097/01.bsd.0000203275.97627.9f
4.
Au Yong N, Lee SI, Getachew R, Kimball J, Sarrafzadeh M, Lu DC. Characterizing transient activation hypertonia during a handgrip force tracking task in cervical myelopathy. In: Society for Neuroscience Annual Meeting—Neuroscience 2013; 2013 Nov 9–13; San Diego, California. Washington (DC): Society for Neuroscience; 2013. Available from: http://www.abstractsonline.com/Plan/ViewAbstract.aspx?sKey=761b20a3-e8a4-4854-a62f-1afe9fb6ef6e&cKey=8070609e-db7d-45bf-b64c-4952bc1801f3&mKey=%7b8D2A5BEC-4825-4CD6-9439-B42BB151D1CF%7d
5.
Lee SI, Ghasemzadeh H, Mortazavi BJ, Yew A, Getachew R, Razaghy M, Ghalehsari N, Paak BH, Garst JH, Espinal M, Kimball J, Lu DC, Sarrafzadeh M. Objective assessment of overexcited hand movements using a lightweight sensory device. In: 2013 IEEE International Conference on Body Sensor Networks (BSN); 2013 May 6–9; Cambridge, Massachusetts. Piscataway (NJ): IEEE; 2013.
6.
Fehlings MG, Jha NK, Hewson SM, Massicotte EM, Kopjar B, Kalsi-Ryan S. Is surgery for cervical spondylotic myelopathy cost-effective? A cost-utility analysis based on data from the AOSpine North America prospective CSM study. J Neurosurg Spine. 2012;17(1 Suppl):89–93.
[PMID:22985375]
http://dx.doi.org/10.3171/2012.6.AOSPINE111069
8.
Fukui M, Chiba K, Kawakami M, Kikuchi S, Konno S, Miyamoto M, Seichi A, Shimamura T, Shirado O, Taguchi T, Takahashi K, Takeshita K, Tani T, Toyama Y, Yonenobu K, Wada E, Tanaka T, Hirota Y. Japanese Orthopaedic Association Back Pain Evaluation Questionnaire. Part 2. Verification of its reliability: The Subcommittee on Low Back Pain and Cervical Myelopathy Evaluation of the Clinical Outcome Committee of the Japanese Orthopaedic Association. J Orthop Sci. 2007;12(6):526–32.
[PMID:18040634]
http://dx.doi.org/10.1007/s00776-007-1168-4
9.
Finkelstein JA. Response shift following surgery of the lumbar spine [dissertation]. [Toronto (Canada)]: University of Toronto; 2010.
10.
Singh A, Gnanalingham KK, Casey AT, Crockard A. Use of quantitative assessment scales in cervical spondylotic myelopathy—survey of clinician's attitudes. Acta Neurochir (Wien). 2005;147(12):1235–38, discussion 1238.
[PMID:16205864]
http://dx.doi.org/10.1007/s00701-005-0639-7
11.
Singh A, Tetreault L, Casey A, Laing R, Statham P, Fehlings MG. A summary of assessment tools for patients suffering from cervical spondylotic myelopathy: A systematic review on validity, reliability and responsiveness. Eur Spine J. 2015;24(Suppl 2):209–28. [PMID:24005994]
http://dx.doi.org/10.1007/s00586-013-2935-x
13.
Jakobsen LH, Rask IK, Kondrup J. Validation of handgrip strength and endurance as a measure of physical function and quality of life in healthy subjects and patients. Nutrition. 2010;26(5):542–50. [PMID:19804953]
http://dx.doi.org/10.1016/j.nut.2009.06.015
14.
Kurillo G, Gregoric M, Goljar N, Bajd T. Grip force tracking system for assessment and rehabilitation of hand function. Technol Health Care. 2005;13(3):137–49.
[PMID:15990417]
15.
MacKenzie CL, Iberall T. The grasping hand. Amsterdam; New York: North-Holland; 1994. xvii, 482 p.
16.
17.
Naik SK, Patten C, Lodha N, Coombes SA, Cauraugh JH. Force control deficits in chronic stroke: Grip formation and release phases. Exp Brain Res. 2011;211(1):1–15.
[PMID:21448576]
http://dx.doi.org/10.1007/s00221-011-2637-8
18.
Kurillo G, Zupan A, Bajd T. Force tracking system for the assessment of grip force control in patients with neuromuscular diseases. Clin Biomech (Bristol, Avon). 2004;19(10): 1014–21. [PMID:15531051]
http://dx.doi.org/10.1016/j.clinbiomech.2004.07.003
19.
Lee SI, Ghasemzadeh H, Mortazavi BJ, Sarrafzadeh M. A pervasive assessment of motor function: A lightweight grip strength tracking system. IEEE J Biomed Health Inform. 2013;17(6):1023–30. [PMID:24240720]
http://dx.doi.org/10.1109/JBHI.2013.2262833
20.
Pradhan SD, Brewer BR, Carvell GE, Sparto PJ, Delitto A, Matsuoka Y. Assessment of fine motor control in individuals with Parkinson's disease using force tracking with a secondary cognitive task. J Neurol Phys Ther. 2010;34(1): 32–40. [PMID:20212366]
http://dx.doi.org/10.1097/NPT.0b013e3181d055a6
21.
Kriz G, Hermsdörfer J, Marquardt C, Mai N. Feedback-based training of grip force control in patients with brain damage. Arch Phys Med Rehabil. 1995;76(7):653–59.
[PMID:7605185]
http://dx.doi.org/10.1016/S0003-9993(95)80635-0
22.
Getachew R, Lee SI, Kimball JA, Yew AY, Lu DS, Li CH, Garst JH, Ghalehsari N, Paak BH, Razaghy M, Espinal M, Ostowari A, Ghavamrezaii AA, Pourtaheri S, Wu I, Sarrafzadeh M, Lu DC. Utilization of a novel digital measurement tool for quantitative assessment of upper extremity motor dexterity: A controlled pilot study. J Neuroeng Rehabil. 2014;11:121. [PMID:25117936]
http://dx.doi.org/10.1186/1743-0003-11-121
23.
Kopp B, Kunkel A, Flor H, Platz T, Rose U, Mauritz KH, Gresser K, McCulloch KL, Taub E. The Arm Motor Ability Test: Reliability, validity, and sensitivity to change of an instrument for assessing disabilities in activities of daily living. Arch Phys Med Rehabil. 1997;78(6):615–20.
[PMID:9196469]
http://dx.doi.org/10.1016/S0003-9993(97)90427-5
24.
McCombe Waller S, Whitall J. Fine motor control in adults with and without chronic hemiparesis: Baseline comparison to nondisabled adults and effects of bilateral arm training. Arch Phys Med Rehabil. 2004;85(7):1076–83.
[PMID:15241753]
http://dx.doi.org/10.1016/j.apmr.2003.10.020
25.
Leclaire R, Blier F, Fortin L, Proulx R. A cross-sectional study comparing the Oswestry and Roland-Morris Functional Disability scales in two populations of patients with low back pain of different levels of severity. Spine. 1997; 22(1):68–71. [PMID:9122784]
http://dx.doi.org/10.1097/00007632-199701010-00011
26.
Haas M, Nyiendo J. Diagnostic utility of the McGill Pain Questionnaire and the Oswestry Disability Questionnaire for classification of low back pain syndromes. J Manipulative Physiol Ther. 1992;15(2):90–98. [PMID:1532978]
27.
Hsieh YW, Wu CY, Lin KC, Chang YF, Chen CL, Liu JS. Responsiveness and validity of three outcome measures of motor function after stroke rehabilitation. Stroke. 2009; 40(4):1386–91. [PMID:19228851]
http://dx.doi.org/10.1161/STROKEAHA.108.530584
28.
Deyo RA, Centor RM. Assessing the responsiveness of functional scales to clinical change: An analogy to diagnostic test performance. J Chronic Dis. 1986;39(11):897–906.
[PMID:2947907]
http://dx.doi.org/10.1016/0021-9681(86)90038-X
29.
Jones RD. The biomedical engineering handbook. 2nd ed (2 vol set). Boca Raton (FL): CRC Press; 1999. Chapter 149. Measurement of sensory-motor control performance capabilities.
31.
Benzel EC, Lancon J, Kesterson L, Hadden T. Cervical laminectomy and dentate ligament section for cervical spondylotic myelopathy. J Spinal Disord. 1991;4(3):286–95.
[PMID:1802159]
http://dx.doi.org/10.1097/00002517-199109000-00005
32.
Hastie T, Tibshirani R, Friedman JH. The elements of statistical learning: data mining, inference, and prediction. 2nd ed. New York (NY): Springer; 2009. xxii, 745 p.
34.
Palmerini L, Mellone S, Avanzolini G, Valzania F, Chiari L. Quantification of motor impairment in Parkinson's disease using an instrumented timed up and go test. IEEE Trans Neural Syst Rehabil Eng. 2013;21(4):664–73.
[PMID:23292821]
http://dx.doi.org/10.1109/TNSRE.2012.2236577
35.
Rosner B. Fundamentals of biostatistics. 7th ed. Boston (MA): Brooks/Cole, Cengage Learning; 2011. xvii, 859 p.
36.
Guyon I, Elisseeff A. An introduction to variable and feature selection. J Mach Learn Res. 2003;3:1157–82.
37.
Prichep LS, Jacquin A, Filipenko J, Dastidar SG, Zabele S, Vodencarević A, Rothman NS. Classification of traumatic brain injury severity using informed data reduction in a series of binary classifier algorithms. IEEE Trans Neural Syst Rehabil Eng. 2012;20(6):806–22. [PMID:22855231]
http://dx.doi.org/10.1109/TNSRE.2012.2206609
38.
Linden A. Measuring diagnostic and predictive accuracy in disease management: An introduction to receiver operating characteristic (ROC) analysis. J Eval Clin Pract. 2006; 12(2):132–39. [PMID:16579821]
http://dx.doi.org/10.1111/j.1365-2753.2005.00598.x
39.
Beurskens AJ, de Vet HC, Köke AJ. Responsiveness of functional status in low back pain: A comparison of different instruments. Pain. 1996;65(1):71–76. [PMID:8826492]
http://dx.doi.org/10.1016/0304-3959(95)00149-2
40.
Grönblad M, Hupli M, Wennerstrand P, Järvinen E, Lukinmaa A, Kouri JP, Karaharju EO. Intercorrelation and test-retest reliability of the Pain Disability Index (PDI) and the Oswestry Disability Questionnaire (ODQ) and their correlation with pain intensity in low back pain patients. Clin J Pain. 1993;9(3):189–95. [PMID:8219519]
http://dx.doi.org/10.1097/00002508-199309000-00006
41.
Grevitt M, Khazim R, Webb J, Mulholland R, Shepperd J. The short form-36 health survey questionnaire in spine surgery. J Bone Joint Surg Br. 1997;79(1):48–52.
[PMID:9020444]
http://dx.doi.org/10.1302/0301-620X.79B1.1269
42.
Hoffman H, Lee SI, Garst JH, Lu DS, Li CH, Nagasawa DT, Ghalehsari N, Jahanforouz N, Razaghy M, Espinal M, Ghavamrezaii A, Paak BH, Wu I, Sarrafzadeh M, Lu DC. Use of multivariate linear regression and support vector regression to predict functional outcome after surgery for cervical spondylotic myelopathy. J Clin Neurosci. 2015;22(9):1444–49.
[PMID:26115898]
http://dx.doi.org/10.1016/j.jocn.2015.04.002
43.
Chiles BW 3rd, Leonard MA, Choudhri HF, Cooper PR. Cervical spondylotic myelopathy: Patterns of neurological deficit and recovery after anterior cervical decompression. Neurosurgery. 1999;44(4):762–69, discussion 769–70.
[PMID:10201301]
http://dx.doi.org/10.1097/00006123-199904000-00041
44.
Radcliff KE, Rihn J, Hilibrand A, DiIorio T, Tosteson T, Lurie JD, Zhao W, Vaccaro AR, Albert TJ, Weinstein JN. Does the duration of symptoms in patients with spinal stenosis and degenerative spondylolisthesis affect outcomes?: Analysis of the Spine Outcomes Research Trial. Spine. 2011;36(25):2197–210. [PMID:21912308]
http://dx.doi.org/10.1097/BRS.0b013e3182341edf
45.
Lee SI, Mortazavi B, Hoffman HA, Lu DS, Li C, Paak BH, Garst JH, Razaghy M, Espinal M, Park E, Lu DC, Sarrafzadeh M. A prediction model for functional outcomes in spinal cord disorder patients using gaussian process regression. IEEE J Biomed Health Inform. 2016;20(1):91–99.
[PMID: 25423659]
46.
Daffner SD, Hilibrand AS, Hanscom BS, Brislin BT, Vaccaro AR, Albert TJ. Impact of neck and arm pain on overall health status. Spine. 2003;28(17):2030–35. [PMID:12973155]
http://dx.doi.org/10.1097/01.BRS.0000083325.27357.39
47.
Vernon H, Mior S. The Neck Disability Index: A study of reliability and validity. J Manipulative Physiol Ther. 1991; 14(7):409–15. [PMID:1834753]
This article and any supplementary material should be cited as follows:
Lee SI, Huang A, Mortazavi B, Li C, Hoffman HA, Garst J, Lu DS, Getachew R, Espinal M, Razaghy M, Ghalehsari N, Paak BH, Ghavam AA, Afridi M, Ostowari A, Ghasemzadeh H, Lu DC, Sarrafzadeh M. Quantitative assessment of hand motor function in cervical spinal disorder patients using target tracking tests. J Rehabil Res Dev. 2016;53(6):1007–22.
http://dx.doi.org/10.1682/JRRD.2014.12.0319
iThenticateCrossref

Go to TOP

VHA

Last Reviewed or Updated Wednesday, February 22, 2017 12:14 PM

Valid HTML 4.01 Transitional