Gait function improvements, using Cardiff Classifier, are related to patient‐reported function and pain following hip arthroplasty

Summarizing results of three‐dimensional (3D) gait analysis into a comprehensive measure of overall gait function is valuable to discern to what extent gait function is affected, and later recovered after surgery and rehabilitation. This study aimed to investigate whether preoperative gait function, quantified and summarized using the Cardiff Classifier, can predict improvements in postoperative patient‐reported activities of daily living, and overall gait function 1 year after total hip arthroplasty (THA). Secondly, to explore relationships between pre‐to‐post surgical change in gait function versus changes in patient‐reported and performance‐based function. Thirty‐two patients scheduled for THA and 25 nonpathological individuals were included in this prospective cohort study. Patients were evaluated before THA and 1 year postoperatively using 3D gait analysis, patient‐reported outcomes, and performance‐based tests. Kinematic and kinetic gait parameters, derived from 3D gait analysis, were quantified using the Cardiff Classifier. Linear regressions investigated the predictive value of preoperative gait function on postoperative outcomes of function, and univariate correlations explored relationships between pre‐to‐post surgical changes in outcome measures. Preoperative gait function, by means of Cardiff Classifier, explained 35% and 30% of the total variance in change in patient‐reported activities of daily living, and in gait function, respectively. Moderate‐to‐strong correlations were found between change in gait function and change in patient‐reported function and pain, while no correlations were found between change in gait function and performance‐based function. Clinical significance: Preoperative gait function predicts postsurgical function to a moderate degree, while improvements in gait function after surgery are more closely related to how patients perceive function than their maximal performance of functional tests.


| INTRODUCTION
Osteoarthritis (OA) is one of the leading causes of years lost to disability worldwide, and one of the most common chronic diseases of the musculoskeletal system. 1,2 Beside painful joints and decreased quality of life, patients with hip OA walk more slowly, walk with altered gait pattern characteristics 3,4 that are often reported as limping gait, 5 and have reduced hip muscle strength compared to healthy controls. [6][7][8][9] Total hip arthroplasty (THA) is a well-accepted and frequently used surgical intervention for severe hip OA, and is considered one of the most successful orthopedic procedures. 10,11 The literature on THA typically investigates functional capacity using simple performance-based activities such as short and long distance walking, stair negotiation, 12 and/or patient-reported outcome measures. 13 While such measures provide information about function, they fail to provide insight into the objective biomechanics of movement which can be observed by three-dimensional (3D) gait analysis. The extensive datasets generated by instrumented 3D gait analysis are usually reduced into a substantially smaller set of discrete metrics (maximum, range, integral) calculated from selected waveforms. As an example, discrete metrics from 3D gait analysis have demonstrated reduced hip adduction and extension angles in THA patients, 6,14,15 but are these discrete measures of functional importance to the degree where the overall gait function is affected, (i.e., summarized gait pattern and performance), and if so, to what extent? Moreover, such discoveries on discrete metrics may in part be false-positive findings due to multiplicity of potential endless numbers of variables. 16 The Gait Deviation Index on kinematics and kinetics have been proposed as single scores that summarize overall gait patterns of the patient's kinematics and kinetics, respectively. 17,18 Studies have shown that the Gait Deviation Index is associated with patient reported outcome measures of physical function, pain and quality of life in patients scheduled for THA, 19 and that the preoperative Gait Deviation Index before THA, to some extent, predict the postoperative Gait Deviation Index. 20 However, the responsiveness of the Gait Deviation Index in patients with hip OA has been questioned since no, 21 or only small to moderate improvements in index scores 22 were observed following THA.
The Cardiff Classifier is a novel approach for generating an overall index of gait function, named after the institution where it was developed. Previous applications of the Cardiff Classifier include the differentiation of pathologic gait function seen in individuals with knee OA and healthy controls, and to monitor postoperative recovery following total knee arthroplasty. [23][24][25] Despite not yet being applied to THA patients, the Cardiff Classifier has potential methodological advantages over the Gait Deviation Index. 26 Opposed to Gait Deviation Index, gait kinematics and kinetics (including ground reaction forces) have typically been included in the single measure of Cardiff Classifier, and all frontal and transverse plane measures at the hip, knee and ankle have been considered. In addition, three of the five most discriminatory biomechanical features found by the Cardiff Classifier in individuals with knee OA are not included in either the Gait Deviation Index for kinematic or kinetics. 26 Moreover, in total knee arthroplasty patients, the Cardiff Classifier methodology has been found to predict postoperative outcome, 23 and has shown surprisingly strong correlations with patient reported outcome measures. 26 To date, application of the Cardiff Classifier in patients with hip OA, and whether it has any predictive value for post THA outcomes, remains unknown.
Simplifying 3D gait analysis data into a single metric describing the overall gait pattern would be of great value in clinical practice to discern whether the overall gait function is affected and to what extent, and to inform healthcare providers and patients what can be expected in terms of change in gait patterns. Further, knowledge on whether it is patients with the greatest perceived recovery who also have the best biomechanical outcomes, and vice versa, is limited.
A comprehensive metric, accounting for interdependencies of biomechanical variables, would facilitate interpretation of results of 3D gait analyses among clinicians, and facilitate monitoring over time and following interventions. Thus, this study aims to (1) quantify and summarize overall gait function using the Cardiff Classifier, (2) evaluate whether this comprehensive gait measure can predict improvements in postoperative patient-reported function and overall gait function 1 year after THA. Secondary, to explore potential relationships between change (pre vs. post THA) in overall gait function by means of Cardiff Classifier versus changes in patient-reported and performance-based function.

| MATERIALS AND METHODS
The regional ethical review board in Stockholm, Sweden approved the study (DNR 2010/1014-31/1). All participants provided verbal and written informed consent in accordance with the Declaration of Helsinki.

| Design and reporting
This longitudinal prospective cohort study (level of evidence: II) reports ancillary data on a previously published study on performancebased function, gait and patient-reported function. 27 The study was reported following the "Strengthening the Reporting of Observational Studies in Epidemiology" (STROBE) Statement as a guideline. 28

| Surgical technique and postoperative regimes
Patients with hip OA received a THA with a direct lateral approach as described by Hardinge. 36 The surgeries were performed by five senior orthopedic surgeons from two different hospitals. Patients had no postoperative movement restrictions, allowing full weight-bearing together with use of an appropriate walking aid, and standard postoperative rehabilitation lasted for a median duration of 2 months after surgery, all according to standard practice at each hospital.

| Cardiff Classifier
Classification of gait patterns using Cardiff Classifier was carried out according to multiple steps (Appendix 1).

| Data reduction, raking, and selection of input features
The present study makes improvements on previously published feature-selection methods before the application of the Cardiff Classifier, 23 to reduce the risk of over-fitting. The training data were split into two halves and the classifier was used to rank the input features within both datasets rendering the top 19 most robustly discriminatory input features for classification (Appendix 2). During the feature selection stage, the data was split into two halves and the classification procedure was followed using every variable within the training dataset. For each input feature, all the subjects were classified using the feature, and a classification accuracy was determined.
This was repeated for each half, and the average classification accuracy across the two sets was used to rank the input features. The target number of retained features was 18, based on previous work in a knee OA cohort, 26   This process is repeated n times until each subject has been left out.

| RESULTS
Out of the 40 patients with hip OA, six received a THA with a posterior approach, and two had incomplete 3D gait data, rendering a total study sample of 32 patients with complete pre-and 1-year postoperative assessments that were included in this study (Figure 1).
The excluded individuals with hip OA did not differ from the studied OA group with regards to age, weight, BMI, or years with symptomatic hip OA.

| Classification of preoperative gait function and change after surgery
The trained classifier had an accuracy of 96.4% (using leave one out cross-validation) in distinguishing between gait patterns of individuals with hip OA and nonpathological individuals ( Figure 3A). Overall gait T A B L E 2 Evaluation of function in individuals with hip osteoarthritis before and 12 months after total hip arthroplasty (THA)

| Predictive value of preoperative gait function
Preoperative gait function, quantified and summarized using Cardiff Classifier, in combination with age and BMI at baseline explained 35% of the total variance in the change in patient-reported Activities of Daily Living subscale of HOOS (Table 3). Preoperative gait function, age and BMI at baseline explained 30% of the total variance in change in gait function after THA (i.e., a reduction in Belief of OA) (   (Table 4). On the contrary, no correlations were found between change in gait function and change in any of the three performance-based functional tests (Table 4).

| DISCUSSION
This study investigated whether preoperative gait function, quantified and summarized by means of the novel Cardiff Classifier, could predict improvements in patient-reported function, and gait function 1 year after THA. Results showed that preoperative gait function, together with age and BMI, predicted more than a third of the total variance of improvements in patient-reported activities of daily living (35%), as well as improvements in gait function after surgery (30%).
Thus, preoperative overall gait function is associated with postoperative patient-perceived outcomes, as well as with objectively assessed postoperative gait function. Moreover, improvements in gait function were also strongly associated with improvements in patient-reported hip-related symptoms and pain, moderately associated to patient-reported function, however, not to improvements in performance-based function. These findings suggest that improvements in gait function after surgery are more closely related to how patients self-report function and pain, rather than their maximal performance of functional tasks of various sorts.
To the best of our knowledge, this is the first study reporting results of Cardiff Classifier among patients with hip OA, while in previous literature it has been successfully applied in patients with knee OA. 23,25,26,39 In the present study, the trained classifier had an accuracy of 96.4% in distinguishing between hip OA and nonpathological individuals, demonstrating that this approach is applicable also in this group of patients. In patients with knee OA, accuracy of the classifier ranged between 90% and 94%. 25   test, which was carried out at self-selected speed, i.e., "usual" performance. Correspondingly, the 3D gait analysis was also conducted at self-selected speed, thus, reflecting ability to perform these ac-  The mathematical process to calculate a confidence factor, convert it to a set of belief functions and combine the evidence is described below. An example calculation using the data from a subject in a

Calculation of confidence factor
Each variable is converted into a confidence factor, cf(x), between 0 and 1 using a sigmoid transfer function: The value θ defines the center of the transfer function, where cf (x) = 0.5, calculated as the follows: Where μ NP and μ OA are the group means and σ NP and σ OA are the standard deviations for variable x in the OA and NP groups.
The value k defines the gradient of the transfer function, and is calculated as follows: Where l is a constant, defined below, ρ x y ( , ) is the Pearson correlation coefficient between the variable x, and the categorical class labels y, and σ x is the standard deviation of x across both groups.
Where K is the normalisation factor for conflicting probability masses.