Runners with patellofemoral pain demonstrate sub-groups of pelvic acceleration profiles using hierarchical cluster analysis: an exploratory cross-sectional study

Data from 110 physically active individuals with PFP with running as their primary exercise modality for at least 6 months, were analyzed in this cross-sectional study. The presence of PFP was confirmed by a licensed healthcare professional (i.e., athletic therapist, physical therapist or medical doctor) based on specific inclusion and exclusion criteria (Table 1). Subjects experiencing pain in other sites were also included in the study, however the primary complaint had to be PFP. Data was collected either at the University of Calgary or in clinical settings partnered with the Running Injury Clinic.

The data collection methods are described in detail elsewhere [21, 22]. Briefly, 8 high-speed digital video cameras (MX3/Nexus, Vicon, Oxford, UK) were used to film treadmill-running at 200 Hz. Spherical retro-reflective markers (9 mm diameter, Mocap Solutions, Huntington Beach, USA) were attached to the specific lower extremity anatomical landmarks bilaterally along with technical marker clusters on rigid shells placed to represent the pelvis and bilateral foot, shank, and thigh segments. Each participant wore the same shoes (Pegasus, Nike, Beaverton, USA) to standardize the footwear condition.

Following placement of all the anatomical and segment markers, each participant stood on a motorized treadmill (Bertec Corporation, Columbus, OH, USA) for a 1-s static trial. Upon completion of the static trial, the markers on the anatomical landmarks were removed while the technical marker clusters remained. The participants were instructed to warm-up on the treadmill for 2–3 min, and then ran on the treadmill at a comfortable self-selected pace (2.61 ± 0.20 m/s) for 20 s, in which approximately 60–80 consecutive running steps were collected for processing and analysis. All participants were experienced treadmill users and were permitted as much time as they required to familiarize themselves with treadmill running before beginning the data collection.

Ankle, knee and hip joint sagittal plane angular accelerations were used for defining ground contact, using previously published event detection methods [23]. The position of the pelvis was measured using the centroid of the pelvic marker cluster [24] and pelvic acceleration was calculated by double differentiation of pelvis displacement using a modified Savitzky-Golay method [25]. Differentiation was performed at both stages using a time-window of 10 data points, and 4th order polynomial fitting. In order to emulate a wearable device, marker accelerations in the global coordinate frame were then converted to a local coordinate frame on the pelvis, using segment markers and rigid body transformations [26]. The local coordinate frame was aligned with the global frame during the static trial.

Each step cycle was normalized to 100 points, with 80 data points for stance and 20 data points for flight phase, since we are analyzing an axial segment. These normalized phases were then combined to represent 100% of the step cycle, averaged over all extracted steps, and standardized to zero mean and unit variance. The kinematic data (3 planes of motions × 100 time-normalized pelvic accelerations) were combined into one 300-dimensional row vector for each subject, creating a matrix of 110 subjects-by-300 data points.

The HCA method was used to identify homogeneous running gait patterns separately for males and females based on the pelvic acceleration time-series, by creating a cluster tree, or dendrogram for each sex-group. Agglomerative strategy or a “bottom up” approach was used, which consists of three steps: (1) a measure of dissimilarity between sets of subjects using the Euclidean distance, (2) subject linkage using the Ward’s minimum variance method [27], and (3) cluster determination using the variance ratio criterion [28].

Following identification of homogeneous clusters (sub-groups) of PFP runners, differences in demographics, injury characteristics, vertical displacement of the pelvic centroid and peak joint angles were examined using one-way ANOVA (Tukey test for post-hoc analyses) and chi-squared test (α = 0.05), and effect sizes were calculated based on η ² and Cramer’s V indices, respectively. In case the data did not present a normal distribution (Shapiro-Wilk test) or a homogeneous variance between sub-groups (Levene test), the Kruskal-Wallis test was performed (Dunn’s test for post-hoc analyses). Differences in pelvic acceleration patterns were examined after applying a principal component analysis (PCA) to the standardized data matrix, and they were identified based on the interpretation of principal components (PCs) that presented a large effect size (η ²  > 0.14) [12, 29], which were used to reconstruct the acceleration waveforms for a better mechanical interpretation [30]. The squared coefficients of correlations between the PC scores and the raw acceleration data (squared loading) [31] were used to calculate the relative loading of the PCs in the vertical (VT), antero-posterior (AP) and medio-lateral (ML) directions to aid in the interpretation of the PCs [32].

We also selected joint angles that are considered important in PFP pathomechanics and that have been suggested to differ between PFP sub-groups [6, 33], to compare between sub-groups. The analyzed peak joint angles were: ankle eversion; knee flexion, knee abduction and knee external rotation; and hip adduction and internal rotation.

A Pearson’s correlation coefficient was calculated between the significant PCs and demographic, injury characterization and kinematic variables that presented differences between sub-groups to determine whether these latter factors were significantly correlated with the acceleration patterns. All data processing and statistical analysis were performed on MATLAB 9.1 (The MathWorks Inc., Natick, MA,USA).

For the female subjects, the variance ratio criterion determined the optimal number of clusters to be two sub-groups (C1 and C2) (Fig. 1a), whereas for the male subjects, no sub-groups could be identified (Fig. 1b) in the HCA.

Subject clinical and demographic characteristics for each cluster are presented in Table 2. There was a significant difference in height (H = 49.9; df = 2; p < 0.001; η ²  = 0.32), with male subjects being taller than both female clusters (p < 0.001 for males vs C1, and males vs C2); and also in mass (H = 50.2; df = 2; p < 0.001; η ²  = 0.32), with male subjects being heavier than both female clusters (p < 0.001 for males vs C1, and males vs C2). There were no significant differences between clusters with respect to age (F = 2.4; df = 2; p = 0.095; η ²  = 0.04), running speed (H = 1.3; df = 2; p = 0.521; η ²  = 0.01), years running (H = 0.9; df = 2; p = 0.629; η ²  = 0.01); ratio of unilateral to bilateral involvement (χ² = 1.7; p = 0.434; V = 0.12) or subjects with multiple injury sites (χ² = 0.5; p = 0.781; V = 0.07).

Table 3 presents descriptive statistics for each cluster regarding pelvic acceleration components, vertical displacement and lower limb kinematics. There were significant differences with large effect sizes between sub-groups only for the following principal components of pelvic acceleration: PC1 (F = 39.7; df = 2; p < 0.001; η ²  = 0.43), with the distinct group being the females in C2 (p < 0.001 for males vs C2, and C1 vs C2); and PC5 (F = 19.8; df = 2; p < 0.001; η ²  = 0.27), wherein C1 was the sub-group with significant difference (p < 0.001 for males vs C1, and C2 vs C1). However, these PCs explained less than 25% of the variance in the dataset (17.0% and 7.7%, respectively). Height presented a significant correlation with PC1 (p = 0.016), but not with PC5 (p = 0.064), and the correlation coefficients were weak (r < 0.30). Although, body mass was different between subgroups, it had no significant correlation (p > 0.05) with either of the selected PCs.

There was a significant difference in peak ankle eversion (H = 15.1; df = 2; p < 0.001; η ²  = 0.12), wherein male PFP subjects presented greater angles than C1 (p = 0.003) and C2 (p = 0.004), and this joint angle had a low but significant correlation with PC1 (p < 0.036, r = − 0.20), but not with PC5. Peak knee abduction was also significantly different between clusters (H = 12.3; df = 2; p = 0.002; η ²  = 0.09), with males exhibiting lower angles when compared to C1 (p = 0.019) and C2 (p = 0.005), and this joint angle was also only correlated with PC1 (p < 0.001, r = 0.38). There were also differences in hip adduction (F = 6.5; df = 2; p = 0.002; η ²  = 0.11), with lower angles for male subjects in comparison to C1 (p = 0.043) and C2 (p = 0.003). The same tendency occurred for hip internal rotation (F = 5.2; df = 2; p = 0.007; η ²  = 0.09), however the difference was only significant for males compared to women in C1 (p = 0.006). While peak hip adduction was correlated with PC1 (p = 0.005, r = − 0.27), hip internal rotation demonstrated a correlation with PC 5 (p = 0.014, r = − 0.23). There were no significant differences for knee flexion (H = 2.2; df = 2; p = 0.331; η ²  < 0.01) and external rotation (F = 1.4; df = 2; p = 0.251; η ²  = 0.03). Vertical displacement of the pelvis presented differences between sub-groups (F = 11.1; df = 2; p < 0.001; η ²  = 0.17), wherein females from C1 displayed lower magnitudes of displacement when compared to males (p < 0.001) and females in C2 (p = 0.011); and this variable was only correlated with PC5 (p = 0.026, r = 0.21).

PC1 presented a high relative loading in the VT direction (47.9%), representing variations in the peak acceleration and the magnitude at early stance phase (Fig. 2a). There was a lower relative loading of PC1 in the AP direction (28.9%), wherein it represented a phase shift of the posterior acceleration peak in early stance (Fig. 2b). In the ML direction, PC1 also represented phase shifts in ML peak accelerations towards the stance and swing limbs during the first half of stance phase (Fig. 2c), but it was the lowest relative loading (23.2%).

PC5 also had relatively high loadings in the VT axis (43.0%), denoting a difference in the rate of magnitude decrease after the peak acceleration (Fig. 2a), although these differences are subtle. The AP relative loading was the lowest for PC5 (25.0%) and it indicated a magnitude difference in the forward acceleration after weight acceptance (Fig. 2b). In the ML direction, there was a low relative loading (32.1%), representing a difference in magnitude variation during the first half of stance phase (Fig. 2c).

Overall, when comparing the pelvic acceleration patterns, males had similar acceleration patterns to females in C1 in the VT and AP directions, but the latter presented higher and delayed peaks ML accelerations. Females in C2 displayed lower acceleration magnitudes in early stance and a higher peak acceleration in the VT direction; a greater forward peak in early stance; and delayed peak accelerations in the AP and ML directions.

The first purpose of the present study was to determine if running gait patterns in individuals experiencing PFP at the time of testing could be clustered into homogeneous sub-groups based on combinations of pelvic acceleration components. In support of our hypothesis, two distinct and homogenous sub-groups (clusters) were present in females with PFP, and these clusters were different when compared to PFP males. These results are similar to previous studies that also reported two to three different running patterns based on visual inspection of 3D kinematic data [6, 7] and mechanical differences between males and females with PFP [8].

There were no significant differences in running speed between sub-groups, which is a factor that has been shown to affect axial segment acceleration [34], especially in the ML axis [35]. Male subjects were significantly taller and heavier than females and these anthropometric differences are known to influence 3D kinematics during running [36]. However, there was a very weak correlation for height, and no correlation for body mass with the acceleration PCs that presented differences between sub-groups suggesting that the relationship with those factors was minimal.

The advantage of investigating pelvic acceleration as a measure of running mechanics is that it is less influenced by marker placement errors and is a much simpler method than a full 3D gait assessment, as it depends only on the trajectory of a single pelvic marker cluster. Additionally, these factors allow for the use of data from multiple research centres, allowing for the application of ‘big data’ analytics and a better understanding of the interaction between biomechanical factors and musculoskeletal injuries [10, 11]. Furthermore, the results of the present study opens the possibility for the use of wearable devices for data acquisition, such as a single triaxial accelerometer on the pelvis, an approach which is becoming increasingly popular in industry and health care [18, 20]. Therefore, the current work identifying sub-groups of PFP patients is a novel finding that can guide future studies in providing better context that can hopefully improve clinical practice.

A secondary purpose was to analyze peak joint angles between clusters to better understand the practical and clinical implications of clustering subjects with PFP based on 3D pelvic acceleration data. In general, differences in joint kinematics were sex-related, since there were no significant differences between female clusters, except for peak hip internal rotation. Moreover, the magnitude of mean differences were within the threshold for detectable kinematic changes reported by Osis et al. [15] for knee abduction (3.4^o) and hip internal rotation (5.6^o). However, the differences in ankle eversion and hip adduction between males and females are greater than the error margins caused by marker placement errors, confirming the findings of Willy et al. [8] who reported males with PFP to have less hip adduction than their female counterparts.

Phinyomark et al. [12] reported the existence of two different sub-groups of asymptomatic runners based on a HCA of lower limb joint kinematics, and when they compared the peak knee abduction angles of those clusters with a sample of subjects with PFP, group differences were dependent on the cluster of healthy individuals that was used as reference. Interestingly, all PFP sub-groups from the current study presented greater values of knee abduction when compared to the ones reported for healthy runners (healthy C1: 8.0^o; healthy C2: 4.4^o). However, there is a tendency for a progressively greater alteration in knee frontal plane angles when comparing males to females in C1 and C2, although there was no significant difference between the female clusters. This could be related with distinct pathomechanical pathways or differences in response to treatment. For example, in a previous work [37] we found that non-responders to exercise treatment protocol presented greater knee abduction angles during late stance and swing phases of running gait, and the current findings suggest that this could be identified by pelvic acceleration data.

To our knowledge, this is the first study to investigate pelvic acceleration profiles in runners with PFP, and the identification of sub-groups could generate insights about differences in pathomechanics or adaptations to pain. Additionally, the analysis of segmental acceleration profiles minimizes measurement imprecisions originating from marker placement errors that propagate into the calculation of joint angles in 3D kinematics [14, 15]. Furthermore, the results of the current study suggest that accelerations acquired using wearable devices [24] may utilise this method in a clinical setting as an evidence-informed method to improve patient care and rehabilitation decisions.

The pelvic acceleration data can provide some clinical insight that can help clinicians make decisions regarding treatment options. For example, peak resultant pelvic acceleration is related to center of mass acceleration during 10 to 75% of stance phase [38]. Therefore, pelvic accelerations can provide some insights on shock absorption and lower limb stiffness. Nevertheless, this connection must be made with caution, since accelerations based on segmental measures overestimate the behavior of center of mass [38]. Women in C2 presented a higher VT peak acceleration, suggesting a diminished capacity for shock absorption. Since no differences in peak knee flexion angles were detected, this could be an indication of greater leg stiffness in these subjects, which is partially supported by the findings that women present higher leg stiffness during running [39] and drop jump landing tasks [40] when compared to males. In contrast, females in C1 were similar to males regarding VT acceleration patterns, which could be explained by the lower VT displacement.

Women also presented higher and delayed peak accelerations in the ML direction, suggesting differences in the control of side-to-side body movement during the first half of the stance phase, when these oscillations occur. This pattern could be related to the larger hip adduction angles exhibited during running, which led to increases in ML accelerations. In addition, females in C2 displayed a delay in peak AP accelerations in early stance, causing a prolonged period of deceleration. It is possible that this finding is related to strength differences between males and females [41, 42], as stronger individuals may be able to exert shorter impulses to achieve the same net change in momentum, however, strength differences were not quantified in the current study.

Although the identification of sub-groups among the female subjects with PFP did not coincide with significant differences in peak lower limb joint angles, there seems to be a progression of values in knee abduction and hip internal rotation depending on the cluster of female subjects. Specifically, there is a tendency for C1 to have lower knee abduction and higher hip internal rotation than C2. These factors could be related to symptom severity or differences in response to treatment, but would need further investigation.

In addition to the differences in height and weight between males and females that were already discussed, other limitations to the current research study are acknowledged. First, this study included both subjects with uni- or bilateral involvement and with secondary pain symptoms besides PFP, which could have also modified running mechanics. However, there was no significant difference in the distribution of those variables between the two subgroups, leading us to believe that it was not an important factor for this clustering. Additionally, these types of patients are frequently seen in clinical practice, therefore these PFP patients are important to include in research studies.

Second, we did not have access to other clinical variables that could influence running mechanics and explain the differences that were found between sub-groups. For example, Selfe et al. [43] has described 3 clusters of PFP patients that were grouped based on clinical measures of strength, flexibility and joint alignment and mobility. Additionally, experimental pain induction in the knee joint has been shown to cause reductions in peak torque in maximal voluntary contraction of knee flexors and extensors [44] and increased sway displacement during quiet stance [45], indicating that pain level could be a driver of changes in motor control. Therefore, future studies should include the aforementioned clinical variables to investigate whether they are related to the differences in running pattern found between sub-groups to have a better understanding in a clinical context.

Finally, this investigation used an HCA approach, which is an unsupervised machine learning technique suitable for exploratory analyses, to determine whether this type of data could be useful in the identification of subgroups within a cohort of runners with PFP. Overall, our hypothesis was supported by the findings and suggest that a supervised analysis could also be applied to identify specific subgroups with specific clinical relevance. For example, recent work from our laboratory used a supervised machine learning method to classify runners with PFP into responders or non-responders to exercise treatment based on running kinematic data, achieving 78% of classification accuracy [37]. Thus, a similar approach could be applied in this context, using pelvic acceleration data to develop an objective method for the identification of such subgroups with greater accessibility in a clinical setting. Regardless, the present study is an important first step to verify the utility of simple measures, like pelvic accelerations, for the objective assessment of gait biomechanics.