Patients with knee osteoarthritis can be divided into subgroups based on tibiofemoral joint kinematics of gait-an exploratory and dynamic radiostereometric study

y University Clinic for Hand, Hip and Knee Surgery, Holstebro Central Hospital, Lægårdvej 12 Holstebro, 7500, Denmark z Department of Clinical Medicine, Aarhus University, Palle Juul-Jensens Boulevard 99 Aarhus N, 8200, Denmark x AutoRSA Research Group, Orthopaedic Research Unit, Aarhus University Hospital, Palle Juul-Jensens Boulevard 165 Aarhus N, 8200, Denmark k Department of Orthopaedic Surgery, Aarhus University Hospital, Palle Juul-Jensens Boulevard 165 Aarhus N, 8200, Denmark ¶ Department of Radiology, Holstebro Central Hospital, Lægårdvej 12 Holstebro, 7500, Denmark # Department of Radiology, Aarhus University, Palle Juul-Jensens Boulevard 99 Aarhus N, 8200, Denmark yy Department of Materials and Production, Aalborg University, Fibigerstræde 16 Aalborg, 9220, Denmark


Introduction
Knee osteoarthritis (KOA) is commonly associated with pain, stiffness, muscle weakness, and joint instability. Thus gait and movement patterns are affected limiting daily activities 1 . Joint pathomechanics in KOA are complex and may affect the entire gait cycle. Even so, most studies have investigated the kinematics in KOA using observer-selected outcomes, i.e., individual discrete time-points, excursion, maximum, and minimum. The entire kinematic trajectory in patients with KOA has never been studied using non-directed hypothesis testing such as statistical parametric mapping (SPM) preventing selection bias 2,3 . This may potentially contribute with patient-specific characteristics of importance for rehabilitation and surgical results.
Classically, kinematic changes in KOA have been reported as comparisons of group categories such as KOA severity 4e6 , affected knee compartment 7 , anterior cruciate ligament (ACL) deficiency 8 , and walking difficulties 9 . However, such predefined categories may promote unnecessary bias and blur kinematic differences due to other unidentified characteristics. A reverse approach may allocate patients with KOA into subgroups based on homogenous kinematic gait trajectories and identify multiple characteristics that affect the kinematic motion patterns between groups. K-means is a centroidbased clustering algorithm using an unsupervised machinelearning approach that previously has been applied on kinematic trajectories 10 . Without prior knowledge of patient or disease characteristics, the algorithm allocates multidimensional data into homogenous subgroups based on an Euclidean distance measure.
Measurement of small kinematic differences requires precise methods. Dynamic radiostereometry (dRSA) can register the threedimensional bone-pose and accurately measure small kinematic changes in native knees with ligament lesions and reconstructions 7,11e13 .
In the present study, we investigated the full-trajectory knee joint kinematics during level gait in patients with advanced KOA to identify: 1) subgroups of KOA patients based on knee kinematics through clustering, and 2) features of knee kinematics unique to the identified subgroup, linking kinematics to patient characteristics describing the subgroup. We compared the results with data from a group of healthy volunteers with asymptomatic non-arthritic knees.

Methods
This study involved the preoperative data of 81 subjects (Table I) who participated in a randomized controlled study investigating the outcome of knee arthroplasty designs (ClinicalTrials NCT03633201). The inclusion period was from 2017 to 2019. A total of 66 patients with radiographic and symptomatic primary KOA were included. The control group comprised 15 local healthy, agesimilar volunteers with asymptomatic knees and no radiographic KOA. The inclusion and exclusion criteria are presented in Table II. The study was approved by the Committee on Biomedical Research Ethics of the Central Denmark Region (1-10-72-303-16, issued 28 February 2017) and registered with the Danish Data Protection Agency (1-16-02-582-16, issued 31 October 2016). The study was conducted in accordance with the Helsinki Declaration, and written informed consent was obtained from all participants. The total estimated effective dose exposed to each patient was 0.629 mSv.

Experimental protocol
Participants walked barefoot on a leveled treadmill (Sole F63, Jonesboro, AR, USA) to mimic level gait (Fig. 1). They had a habituation period to get familiar with the test environment, with slowly increasing speed reaching a final speed of 0.83 m/s. This is slightly slower than average walking speed (1.25 m/s) 14 and was chosen to avoid exclusion of gait-disabled patients and facilitate sufficient dRSA data generation throughout the entire gait cycle. When the subject felt comfortable, data collection was initiated. Up to seven coherent gait cycles were obtained. For precaution and anticipating loss of balance during testing, subjects had a rail they could hold on to. Only none-rail-supported gait trials were included for further analysis. An overview of the workflow is illustrated in Fig. 2.

Dynamic radiographic imaging
The gait trials were recorded with a dedicated dRSA system (AdoraRSA; NRT X-Ray A/S, Hasselager, Denmark) with a previously described biplanar setup 13 . The radiation exposure parameters were set at 90 kVp, 600 mA, and a pulse-width of 2.5 ms, utilizing the highest frame rate of 15 Hz without compromising the image size. Simultaneously with the dRSA system, we used a minimum of six OptiTrack Prime 13 motion cameras (NaturalPoint, Corvalis, OR, USA) and Motive software (Motive v.2.0.0, NaturalPoint, USA) to assess the trajectories of skin-attached reflective markers (diameter: 10 mm). Prior to usage, the entire recording area was calibrated using an OptiTrack CW-500 Calibration Wand (NaturalPoint, USA). A custom-programmed Raspberry Pi 3 Model B (Linux Mini PC, Broadcom BCM2837 1.2 GHz Quad-Core 64-bit, 1 Gb LPDDR2 RAM) was used to time-synchronize data from the dRSA and the optical marker systems using pulse information from both systems.

Patient group
Healthy control group Inclusion criteria Age above 18 years but no more than 80 years of age. Age above 18 years but no more than 80 years of age. Informed and written consent.
Informed and written consent. Primary knee osteoarthritis in capable men and women.
Asymptomatic knees and no radiographic osteoarthritis. Indication for cruciate-retaining total knee arthroplasty.

Exclusion criteria
Patients with a thigh circumference exceeding 60 cm. Patients with a thigh circumference exceeding 60 cm. Patients with conditions that severely compromise their gait other than knee osteoarthritis in the affected knee.
Patients with conditions that severely compromise their gait. Patients with previous severe fractures at the knee level or severe malalignment at the knee level.
Patients with previous fractures or severe malalignment at the knee level. Surgically implanted metallic parts and pacemaker.
Surgically implanted metallic parts and pacemaker. Patients with need for an augmentation and/or stem extension.
Patients who cannot perform the exercises.
Patients who cannot perform the exercises.

Volumetric imaging
All participants underwent a computed tomography (CT) (Revolution EVO, GE Medical Systems, Milwaukee, WI, US) to construct subject-specific bone models (Fig. 3). A helical scan protocol with the reconstruction kernel 'boneplus' was used covering 15 cm of the most distal and proximal part of the femur and tibia. Knee scans were acquired with axial slices at a peak voltage of 100 kVp and 200 mAs, a slice thickness of 0.625 mm and a pixel spacing of 0.48 Â 0.48 mm. The CT also included imaging with minor changes to reduce dose exposure of the femoral head (slice thickness of 2.5 mm) and ankle (80 kVp and slice thickness of 2.5 mm) to construct an anatomical coordinate system using the mechanical axis. Bones were segmented using an implemented, fully automated graph-cut segmentation method employing the Insight Segmentation and Registration Toolkit (Kitware, Clifton Park, NY, USA) 15,16 . From the segmentations, both three-dimensional volume and surface bone models were created. The volume model comprised the extracted bone containing the greyscale information of the CT. The surface model was extracted using the marching cubes algorithm and the visualization Toolkit (Kitware). An anatomical coordinate system was assigned to all bone models using a modified version of Miranda et al. 17 to implement the mechanical axis 11 .

Analysis of dynamic radiographic imaging
The series of stereoradiographs were analyzed using an automated software system developed at our institution (AutoRSA software, Orthopaedic Research Unit, Aarhus, Denmark). The software utilized a digitally reconstructed radiographic (DRR) registration method to estimate bone pose from virtually generated projections using mathematical optimization algorithms, as previously described in more detail 13,16,18 . The optimal pose was found by minimizing the difference between virtually generated DRR images from a 3D volume to actual stereoradiographs using a normalized gradient correlation approach 13,18,19 . Prior to the DRR registration process, the automated bone pose initialization between frames deviated from previously described studies 13,16,18 by utilizing the skin-attached marker trajectories.

Quantification of knee joint kinematics
We implemented the joint coordinate system [ Fig. 3(e)] initially defined by Grood and Suntay 20 to describe the knee joint kinematics, but using the modified equations proposed by Dabirrahmani and Hogg 21 to account for possible hyper-extension and hyper-flexion. The relative motion in all six degrees-of-freedom Table II Osteoarthritis and Cartilage Group summary and comparison (rotation and translation) of the body-fixed anatomical coordinate systems was computed for each frame. The translation was quantified in mm with medial tibial shift, tibial anterior drawer, and joint distraction as positive directions. Rotations were measured in sequence as presented and quantified in degrees with flexion, adduction, and tibial internal rotation as positive directions. Kinematic values were divided into gait cycles, starting and ending with two successive initial foot contacts of the ipsilateral limb. We used the approach of O'Conner et al. 22 to identify initial contact using the optical markers on foot. To estimate each subjects' most representative gait cycle pattern, we time-normalized the kinematic measures to 21 points representing the gait cycle from 0 to 100% and calculated the median across trials for each subject.

Clinical characteristics
Based on conventional radiographs, KOA was classified according to the Ahlbeck score (grade 1e5). Those without KOA were classified as grade 0. Additionally, we registered the affected tibiofemoral compartment as lateral or/and medial. Knee ligament lesions were assessed with high accuracy and repeatability using detailed magnetic resonance imaging (1.5 T Avanto, SIEMENS, Erlangen, Forchheim, DE) 23 . We applied a modified version of the Osteoarthritis Initiative protocol 24 based on GE scanner recommendations (T2 SAG de3D DESS WE acquisition with 0.7 mm slice thickness and T1 COR fl3D WE with 1.5 mm slice thickness). We evaluated the following knee ligaments: ACL, posterior cruciate ligament (PCL), medial collateral ligament (MCL) and lateral collateral ligament (LCL). The ACL was graded as: 0 (no lesion), 1 (partial lesion), and 2 (full lesion). PCL, MCL, and LCL were registered as 0 (intact) or 1 (with lesion). The clinical outcome measures were assessed by the Oxford Knee Score (OKS), the Forgotten Joint Score (FJS) and the Knee Osteoarthritis Outcome Score (KOOS).

Subgroup allocation
The median kinematic trajectories of each patient were used to construct a 126 by 66 feature matrix (M) containing the 126 trajectory kinematic feature points (6 kinematic parameters Â 21 time-points) for each of the 66 patients. Feature point ordering was consistent across all participants in the matrix. For the purpose of clustering, each feature was standardized to avoid different weightage between features. Each feature was subtracted by the respective mean values and divided by the standard deviation.

Osteoarthritis and Cartilage
Illustration of the treadmill setup. Utilizing the largest recording area possible, we used a biplanar setup where the detectors were placed on individual stands with a relative angle of 40 ; one in front of the subject and one to the side of the limb of interest. The source image distance was approximately 240 cm (frontal view) and 280 cm (side view). The source object distance was approximately 190 cm (frontal view) and 220 (side view). The system was acting in a plane parallel to the floor at a patient-specified height to centralize the knee joint in the radiographic images throughout the entire recording. The setup was mirrored for left/right knees to accommodate having the investigated knee as close as possible to the side view. The six optical cameras were positioned strategically surrounding the subject to avoid occluded markers during analysis of the skin-attached reflective markers. The markers (white dots) were attached at the subjects' pelvis, lower limb, and feet; three cluster markers at the thigh, two at the knee epicondyles, three cluster markers at the shank, two at the ankle malleolus, one at the heel, three at the metatarsales (1,3,5), and one at the first distal phalanx. The mask image was used for a refined optimization on each frame excluding irrelevant pixel values surrounding the DRR overlay image area of interest.
Subgroup allocation was performed with skleans 25 implementation of k-means clustering in Python. K-means clustering is an iterative algorithm randomly allocating every data point to the nearest subgroup, while minimizing the sum of squared Euclidian distance between every data point and its subgroup's centroid. Due to random seeding of the initial allocation, the k-means algorithm may not reach the global optimum, but instead converge to a local optimum. To handle this limitation, each clustering process was based on 50 repetitions with different, randomly allocated seeds. The repetition with the lowest sum of squared Euclidian distance was chosen as the final allocation of one clustering process. Another challenge of the k-means approach is to determine the number of subgroups, k. To do so, we investigated repeatability and quality for k number of subgroups ranging from two to five. Subgroup repeatability was evaluated based on repeating ten clustering processes for each of which we tracked differences in subgroup allocations with increasing k-subgroups. Subgroup quality of the two to five k-subgroups was assessed using the Silhouette value of the clustering process with the lowest sum of squared error of 10 repetitions (considered as the best). The Silhouette value is a measure of similarity describing how well subjects is allocated within a subgroup in relation to all other subgroups. The value ranges from À1 to 1 with a larger value indicating a stronger association with its allocated subgroup, whereas a more negative value indicates a stronger association with other subgroups. Of these four subgroup allocations, the one with the best repeatability and quality was used in the subsequent analysis.

Statistical analysis
Differences in knee joint kinematics across the entire gait cycles were examined using one-dimensional SPM with the open-source code spm1d (spm1d.org, v.0.4.2) for Python (Python Software Foundation, v.3.6). SPM allows for examining the entire onedimensional time series of kinematic trajectories, avoiding selection bias and allowing for non-directed hypothesis testing instead of reducing the dataset to a certain observation or alternatively risking false hypothesis testing due to multiple repeated measurements 2,3 . SPM uses Gaussian random field theory to calculate the threshold that only the significance level of equivalently smooth Gaussian random fields would cross when the null hypothesis is true. QQ-plots revealed that our kinematic data were not normally distributed. Thus, we used statistical non-parametric mapping (SnPM), which deals with smoothness implicitly and estimates the test statistics through permutation 26 . First, a Hotelling test was implemented on the entire vector field. Then, the post-hoc Hotelling test was applied to each vector component if the vector field level reached statistical significance.
We compared demographically and clinically characteristic differences between groups using one-way analysis of variance (ANOVA) for continuous variables, ordinal logistic regression for categorical variables, and logistic regression for binary variables. When significant differences were detected, Bonferroni correction was applied for group differences. Visual inspection of QQ-plots verified normally distributed data.

Subgroup allocation
The quality and repeatability analyses are presented in Fig. 4. . The individual subgroup allocation across the ten consecutive repetitions showed identical subgroup allocation for k ¼ 2. Individual data allocation was more variable for k ¼ 3 and k ¼ 4 with three and ten individuals switching subgroups, respectively. No consistent pattern of subgroup allocation could be identified for the k ¼ 5 solution. Noticeably, the second (k ¼ 3) and third (k ¼ 4) solutions allocated two identical subgroups (G3,G4), whereas the third (k ¼ 4) solution separated the remaining subgroup into two subgroups (G1,G2). All of this indicates that four subgroups may represent the optimal solution for separating the current dataset into the largest number of subgroups with the highest quality and reasonable repeatability. Consequently, the third (k ¼ 4) solution was chosen for further analysis; G1 (n ¼ 20), G2 (n ¼ 17), G3 (n ¼ 10), and G4 (n ¼ 19).

Kinematic and clinical characteristics
The tibiofemoral joint kinematic trajectories for the entire patient cohort showed increased tibial external rotation, tibial lateral shift, and joint narrowing compared to the healthy group (Fig. 5). The four gait-trajectory-based subgroups (G1,G2,G3,G4) are compared to the healthy group in Fig. 6 and Table IIc (color-code highlight the main differences). The in-between subgroup kinematic comparison can be found in the Supplementary material. Clinical differences and a schematic overview of the most relevant differences between subgroups and the healthy control group are presented in Table II.

G1 e The flexion group
This was the only subgroup revealing different knee flexion when compared with the healthy group. Increased knee flexion was identified at initial contact, terminal stance, and terminal swing phase. Additionally, throughout the entire gait cycle, this subgroup showed greater adduction and joint narrowing than the healthy group. The clinical characteristics revealed that this subgroup consisted primarily of cases with medial tibiofemoral osteoarthritis. In relation to the other subgroups, this group displayed a

Osteoarthritis and Cartilage
Illustration of the right leg of a CT (a), bone segmentation along with marked hip and ankle centres (b), anatomical coordinate system (ced), and knee joint coordinate system (e). (a) Volume rendering of a hip-knee-ankle CT. (b) Bone segmentation of the femur (cyan) and tibia (magenta). The hip joint centre (cyan) was defined as the centre of a sphere fitted to the cortical bone of the femoral head using least square. The ankle centre (magenta) was defined as the midpoint between the ankle malleolus that was manually selected at the CT volume-rendered model. (ced) 3D bone models of the femur (c) and tibia (d) bones in a sagittal and frontal view illustration the anatomical coordinate system. Femur: the lateralemedial axis was defined as the centre line of a cylinder placed using a least-square fit to the knee condyles, the proximaledistal axis was defined as an orthogonal projection from the medialelateral axis to the centre of a sphere at the femoral head. The anterior-posterior axis was defined as the cross product of the medialelateral axis and the proximaledistal axis. The origin was defined as the midpoint between the medialelateral axis surface intersections. Tibia: the origin was defined as the centroid of the tibial plateau proximal of the largest cross section. The medialelateral axis was defined as the first principal component axis. The proximaledistal axis was defined as an orthogonal projection from the medialelateral axis to the midpoint between the ankle malleoli. The anterior-posterior axis was defined as the cross product of the medialelateral axis and the proximaledistal axis. (e) Illustration of the knee joint coordinate system used for kinematic pose assessment of the femoral and tibial bones considered in relation to each.
larger flexion angle than the anterior group (loading response and initial swing phase) and the external rotation group (swing phase). Additionally, this subgroup displayed the largest internal rotation of any group.

G2 e The abduction group
This was the only subgroup revealing greater abduction than the healthy group. This was identified throughout the entire gait cycle. In addition, this subgroup showed greater joint narrowing throughout the gait cycle and anterior drawer during the loading response and terminal swing phase. The clinical characteristics revealed that it was the only subgroup that included cases with lateral tibiofemoral osteoarthritis. In relation to the other subgroups, this group displayed the largest abduction. In addition, it revealed a larger anterior drawer than the flexion group (stance, initial swing, and terminal swing) and the external rotation group (initial contact to mid-stance and terminal stance). It was only exceeded in this respect by the anterior drawer group.

Osteoarthritis and Cartilage
Presentation of cluster allocation for k-values ranging from 2 to 5. Top row: Silhouette values of each subject of the repetition that showed the best mean square error of the ten repetitions. Bottom row: Change in cluster allocation across the ten repetitions with respect to each subject's silhouette value.

Osteoarthritis and Cartilage
Kinematic comparison of the entire patient group with the healthy control group. The top row presents the mean trajectories of the two groups with confidence interval as the shaded area. The bottom row presents the post hoc non-parametric scalar field t tests (SnPM{t}), depicting where patients show higher (þ) and lower (À) than healthy subjects. The thin dotted lines indicate the critical thresholds for significance. The greyshaded areas illustrate when critical threshold is exceeded thus determining a significant difference.

Osteoarthritis and Cartilage
Statistical parametric mapping of all kinematic parameters (flexion/extension, adduction/abduction, internal/external tibial rotation, medial/lateral tibial shift, anterior/posterior tibial drawer, joint distraction/narrowing) for each cluster compared with the healthy control group. For each cluster comparison, the top row presents the mean trajectories of the two groups with confidence interval shown as the shaded area. The bottom row presents the post hoc non-parametric scalar field t tests (SnPM{t}), depicting where patients show more (þ) and less (À) than healthy controls. The thin dotted lines indicate the critical thresholds of significance. The grey-shaded areas illustrate when the critical threshold exceeds; thus, a significant difference is present. The superimposed colour-squares highlight trajectories with similar differences when compared with the healthy control group: cyan presents increased flexion angle trajectories; magenta presents increased adduction angle trajectories; yellow presents abduction angle trajectories; green presents increased tibial external rotation and tibial lateral shift trajectories; red presents increased tibial anterior drawer; blue presents increased joint narrowing. The significance level was set to 5%.

G3 e The anterior drawer group
This was the only subgroup revealing severe anterior drawer throughout the gait cycle. Furthermore, this subgroup showed the largest external tibial rotation and lateral tibial shift throughout the motion (similar to G4) and larger adduction and joint narrowing compared with the healthy group. The clinical characteristics revealed that this subgroup was primarily composed of cases with medial tibiofemoral osteoarthritis, with partial and total ACL lesion and the largest KOA score of any group. In relation to the other subgroups, this group displayed the largest anterior drawer and, during the swing phase, the largest joint narrowing. In addition, like the external rotation group, it showed the largest adduction, external rotation, and tibial lateral shift. For this subgroup, increased lateral tibial shift was not observed during mid swing, whereas increased external rotation was not found during mid-swing, but this only applied when compared with the abduction group.

G4 e The external rotation group
This subgroup revealed, similar to G3, more external tibial rotation and lateral tibial shift, but no anterior drawer was observed compared with the healthy group. In addition, this subgroup showed more adduction and joint narrowing throughout the gait cycle. The clinical characteristics of this subgroup included the cases with the largest proportion of MCL and PCL lesions. In relation to the other subgroups, this group displayed, like the anterior drawer group, largest adduction, external rotation, and tibial lateral shift. For this subgroup, increased lateral shift was not observed during the swing phase compared with the abduction group, whereas increased external rotation was observed only when compared with the abduction group. Similar to lateral shift, no difference in external rotation was found during mid-swing between this subgroup and the abduction group.
All subgroups reported larger VAS pain scores during gait than the healthy group, with exception of the abduction group due to large variation in this group. However, no differences between subgroups were identified. Similarly, all subgroups had poorer clinical scores than the healthy group. Although the entire KOA patient group displayed greater Body Mass Index (BMI) than the healthy group and the adduction group included younger patients than the flexion group did, no other differences between groups were found in terms of potentially confounding variables (age, height, weight, side, and gender) (Table II).

Discussion
In patients with advanced KOA, we identified four clusters with homogeneous knee joint kinematics during gait that relate well with clinical characteristics but were clearly different from the knee joint kinematics of healthy volunteers without KOA. Our study differs from previous studies in three ways. First, we recorded the entire gait cycle using the accurate and precise dRSA method to assess knee kinematics (limits of agreement below 0.3 for rotation and 0.4 mm for translation) 13 rather than using traditional motion capture techniques that are often inhibited by soft-tissue artifacts. Second, we applied SPM, which allows examination of an entire kinematic trajectory thereby minimizing the risk of type-II errors and selection bias, which can be a problem with discrete-point comparisons. Third, we separated the patient cohort reversely by homogenous kinematics which present potential KOA phenotypes instead of applying the more often used predetermined clinical characteristics.
Previous gait studies of patients with KOA have compared various KOA group compositions to various control groups. KOA groups have been compared with healthy controls 4,27 and patients with medial tibiofemoral osteoarthritis have been compared with both healthy controls 5 and healthy controls with varus knee alignments 8 . Additionally, various discrete time-points and parameters during different gait phases have been used for group comparisons 4,5,8,27 . Thus, direct comparison with our results is challenging. However, a frequently chosen time point when comparing KOA and healthy groups is initial contact of the foot to the ground. At initial contact, diverse kinematic behaviors have been reported for patients with KOA compared with healthy controls: in patients with KOA, knee flexion has been identified as greater 5,27 , lower 4 , and similar 8 to that of healthy controls. Among these studies, three have investigated adduction and internal rotation. They found greater adduction 4,5,8  When comparing the entire KOA cohort with the healthy control group, we found increased tibial lateral shift, and joint narrowing at initial contact, whereas increased tibial external rotation first occurred later during stance. Thus, we cannot confirm the previous observations 4,5,8,28 which associate patients with KOAs with altered flexion, adduction, or anterior drawer. The likely reason for this is the large variation in kinematic trajectories in the KOA cohort. Thus, when the KOA cohort was clustered, we found statistically significant, clinically relevant differences for all kinematic parameters compared with the healthy control group. This suggests that the cohort of KOA patients comprises different kinematic phenotypes and that divergence in the results of previous investigations of the KOA group as a whole is due to the specific composition of kinematic subgroups included but not controlled for. This statement is substantiated by the fact that the abduction group (G2) was not observed until the number of subgroups increased from three to four. The identified kinematic characteristics of each subgroup could be linked to specific clinical characteristics. For example, the flexion group (G1), the anterior drawer group (G3), and the external rotation group (G4) revealed increased adduction. It displayed the largest portion of patients with medial tibiofemoral OA, whereas the abduction group (G2) revealed the largest proportion of patients with lateral tibiofemoral OA. Mechanically, this can be explained by the fact that reduced medial cartilage thickness increases a varus knee posture and, conversely, reduced lateral cartilage thickness increases a valgus knee posture 29 . Additionally, the abduction group (G2) and the anterior drawer group (G3) were the only subgroups revealing increased anterior drawer and larger ACL lesion grades than the healthy group. Additionally, the anterior drawer group (G3) showed the most prominent increase in anterior drawer and the largest proportion of patients with ACL rupture. Furthermore, the anterior drawer group (G3) also displayed the largest Ahlbeck score, representing a link between the well-established coherence of anterior-posterior knee instability and the development of KOA 30 .
A systematic review 31 found that the course of pain and physical functioning in KOA were diverse and described a high heterogeneity across studies. The authors also suggested that KOA populations consist of subgroups or phenotypes. Our findings support this. However, the cluster scores and clinical outcome variables in our study indicate further subgroup overlap. Other subgroups and phenotypes have previously been identified. Knoop et al. 32 (later confirmed by Holla et al. 33 ) identified three subgroups with distinct trajectories of physical functioning over time (good, moderate, and poor). Esch et al. 34 identified five homogeneous clinical phenotypes (minimal joint disease phenotype, strong muscle strength phenotype, severe radiographic KOA phenotype, obese phenotype, and depressive mood phenotype). Another systematic review proposed six other phenotypes (chronic pain, inflammatory, metabolic syndrome, metabolic bone/cartilage, mechanical overload, and minimal joint disease). These phenotypes may potentially explain some of the variation in clinical outcomes within the identified subgroups in our study. Thus, further investigation is needed, including multiple characteristics, e.g., clinical, biomechanical, psychosocial, and genetic factors, categorizing and identifying phenotypes that embrace the heterogeneous pathology and multifactorial nature of patients with KOA.
Overall, TKA is a successful treatment for pain reduction in patients with KOA. However, up to 20% of treated patients are, dissatisfied with the outcome 35 , and more than 50% have residual knee symptoms 36 . Although considerable effort has been devoted to increase patient satisfaction, this has not yet been accomplished. The natural knee joint is heterogenic in shape, size, and function. Gait kinematic phenotyping within KOA patients may help us better understand the contributory and persistent causes of TKA patients' dissatisfaction with outcomes. Maybe, various patient groups should undergo different interventions or treatment with specific implant designs. Treatment targeted more specifically towards selective phenotypes has also previously been suggested to lead to improved outcomes 37 . To clarify this, new research on the groups identified in this study needs to be conducted.
Some limitations of our study warrant discussion. First, the gait pattern on the treadmill may be different from that over ground. However, some reports have declared that only minimal differences in kinetic and kinematic parameters were found between treadmill and over-ground gait 38,39 . Barefoot walking may influence the load on the lower extremity joints within the subjects, but conditions were similar for the KOA group and the healthy group. Thus, we do not expect this to have influenced either the results or the conclusion. Second, inevitably, crossing leg coursed leg-projection overlap during dRSA acquisition. However, our setup ensured that one of the radiographic views were always free of overlay and we did not see a negative effect of leg-overlay on accuracy in the entire gait cycle. Third, the requirement of a maximum thigh circumference may have excluded obese subjects whom have showed altered kinematics. 40 In conclusion, we found that patients with advanced KOA can be clustered into four subgroups based on homogeneous gait patterns. For these subgroups, we determined a meaningful relationship to different KOA-affected compartments, KOA progression and ligament lesions. A better understanding of knee joint pathomechanics in patients with KOA allows for phenotyping of subgroups, which may inspire improved and more patient-specific treatment strategies in the future.

Contributors
ETP was involved in all aspects of the study and drafted the manuscript. SR, DK, and JD operated the patients, whereas NEL and SR examined and labelled the clinical characteristics. SR, MSA and MS had an essential role in the study design, interpretation, and presentation of data. All authors contributed to data interpretation and manuscript revision.