Identifying and characterizing biomechanical gait subpopulations in hip osteoarthritis with explainable machine learning

$\textbf{Summary}$
Understanding specific biomechanical gait adaptations to hip
osteoarthritis (HOA) and joint replacement is highly relevant
in providing tailored rehabilitation. In this study, we identified
three characteristic subpopulations among 108 patients with
unilateral HOA. We then assessed their gait kinematics and
joint moment differences compared to healthy controls (n =
56). Additionally, the subpopulation-specific effects of total
hip replacement (THR) were analyzed based on data from 64
patients.

$\textbf{Introduction}$
Machine learning offers powerful methods to advance HOA
research and treatment options by analyzing high-dimensional
biomechanical datasets, supporting targeted treatment [1,2].
The aims of this study were (1) to identify groups of patients
(i.e., subpopulations) who respond differently to HOA by
exhibiting different adaptations in gait kinematics and joint
moments, and (2) to assess the subpopulation-specific effects
of surgical treatment in a second examination after THR.

$\textbf{Methods}$
Three datasets were analyzed: a dataset of 108 unilateral HOA
patients before THR (PRE), a subset of PRE consisting of 64
... mehr
patients who participated in a second examination (7-25
month) after THR (POST), and a dataset of 56 healthy
controls (HC). All datasets included 3D laboratory-based gait
analysis data. Principal component analysis was used to
reduce the gait cycle waveforms (angles and moments) of 3D
pelvis, ipsilateral 3D hip, ipsilateral 3D knee, and ipsilateral
sagittal ankle joint as well as the ipsilateral foot progression
angle to 14 features. These features were used for k-means
clustering to identify PRE subpopulations. Linear Support
Vector Machines were applied to classify the subpopulations
and the HC, and the POST was re-evaluated using these
models. Classification rates were determined as the
percentage of correctly classified individuals. SHapley
Additive explanations were used to identify the most
important biomechanical waveforms. In addition, a lowdimensional classifier-oriented gait score was evaluated as a
measure of gait quality [3].

$\textbf{Results and Discussion}$
Three PRE subpopulations were identified that exhibited
different adaptations in gait kinematics and joint moments.
Individual classification rates for the subpopulations ranged
from 51.4% to 85.2%. Across all subpopulations, hip flexion
and rotation waveforms (angles and moments) were identified
as most important for classification. Two subpopulations
showed reduced hip moments, while one subpopulation
showed altered hip and pelvic kinematics. The classifieroriented gait score improved uniformly in all subpopulations,
suggesting an overall beneficial effect of THR (Figure 1).
However, subpopulation two showed minor improvements in
gait biomechanics compared to HC, supporting different
functional outcomes after THR [2].

$\textbf{Conclusions}$
This study provides a differentiated view of gait adaptations
in patients with HOA before and after THR. Identifying a
patient’s affiliation with a specific subpopulation allows
clinicians to gain deeper insights into gait adaptations in
HOA, enabling the development of rehabilitation programs
tailored to their individual needs.

$\textbf{Acknowledgments}$
This work was supported by the German Research Foundation
(492686079).

$\textbf{References}$
[1] Halilaj et al. (2018). J Biomech, 81, 1-11.
[2] Van Drongelen et al. (2021). J Clin Med, 10(10), 2167.
[3] Christian et al. (2016). Clin Biomech, 33, 55-60.