^{1}

^{1}

^{1}

^{2}

^{1}

^{2}

^{3}

^{4}

^{*}

^{1}

^{2}

^{3}

^{4}

Edited by: Simone Tassani, Pompeu Fabra University, Spain

Reviewed by: Luca Modenese, Imperial College London, United Kingdom; Raphael Dumas, Université Gustave Eiffel, France

This article was submitted to Biomechanics, a section of the journal Frontiers in Bioengineering and Biotechnology

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

Differences in motion patterns can be attributed to a large number of associated variables: velocity, proprioceptive, vestibular, and visual stimuli as well as neurocognitive and executive functions, body weight, sex, aging effects, and pathological deviations (

Recent advances in computational methodology allow for improved characterization of bone morphometry as well as motion at a population wide level. Statistical shape modeling enables to describe individualized bone geometry more precisely than consensus bone geometry or linearly scaled generic bone models (

Understanding how bone morphometry affects joint function might reveal fundamental insights into how geometrical features contribute to musculoskeletal disorders (

The objectives of this paper are therefore twofold. First, we intend to quantify individual differences for a wide range of activities of daily living (ADL) at a population-wide level. Statistical kinematic models that aim to describe the inter-subject variance in natural joint motion are appropriate for this purpose (

A group of able-bodied males aged between 18 and 25 years was recruited to establish the relationship between morphometric and motion variability. A healthy and homogeneous group was chosen to minimize potential bias from clinical (e.g., neurological and musculoskeletal pathology) origin or age-related differences. Therefore, individuals with musculoskeletal disorders or history of surgery were excluded. A second prerequisite to participate involved the absence of overweight (i.e., BMI less than 25 kg/m^{2}). An overview of population demographics (

General characteristics of the investigated population.

Age (years) | 22.1 (21.5–22.7) | 2.2 |

Length (cm) | 181.4 (179.8–183.0) | 6.3 |

Weight (kg) | 71.5 (69.5–73.5) | 7.8 |

CE angle (°) | 28.2 (26.9–29.4) | 4.8 |

Alpha angle (°) | 64.5 (62.4–66.5) | 8.1 |

CCD angle (°) | 129.2 (128.0–130.3) | 4.6 |

Femoral anteversion (°) | 8.8 (6.8–10.9) | 8.0 |

The gold standard to obtain individualized bone geometry is the segmentation of shapes from high resolution 3D medical imaging (

Twenty-seven skin beads (12 mm) were applied to the bony landmarks of the pelvis-leg apparatus and one at the vertebra prominens. Markers were placed by the same investigator and according to standardized protocols (

Subjects were asked to perform several ADL activities. More specifically, the motion analysis included stationary cycling, squat, lunge, and stair movements. Before motion tracking was initiated, each test subject received a brief training of the intended movements. The purpose of this instruction moment was to secure an adequate and smooth recording of the motor tasks. Then each movement was executed and recorded twice. All experiments were conducted in the same setting and under the same circumstances to limit the influence of external factors.

To mimic the bicycle movements, a bike model was constructed. As provided in

Flow chart of the musculoskeletal simulation in the Anybody Modeling System. On the upper left, there is the bike model that was used by our test panel during the cycling experiments. Motion was recorded frame by frame through skin marker registration. Each subject underwent MRI for determination of the skin tags. Subsequently, Anybody calculated the kinematics for all the recorded frames. An identical workflow was applied for the simulation of squat, lunge, and stair movements.

Subjects were asked to take a seat on the bike model while maximally bending over, to mimic the posture of a professional cyclist. The height of the saddle was adjusted to be equal to the height of the hips of the subject, standing next to the model.

Squatting is one of the most challenging motions for the hip and knee joints as it generates considerably high reaction forces and it approximates the fully functional flexion range of the lower limb (

Like the squat, lunging is a closed-chain movement. The subjects stepped approximately 0.6–1.0 m forward with the right leg onto the other force plate. Consequently, both knees bent at the same time. Ultimately, the subjects stood still in this position for a few seconds and pushed off the right foot to rise. The recording ended when the starting position was reached.

Subjects stood on a step and smoothly stepped up or down to the second step. As for the lunge, the volunteers were asked to use their right leg first. Both ascent and descent staircase motions were modeled individually. Given a previous study showed knee peak flexion angles to be correlated with the step height, the stair level was fixed (

To simulate the ADL activities, the segmented bones, the position of the pelvic, thigh, and shank markers, the motion capture trajectories and force plate data were imported into the AMS (version 7.1.0). For each subject, individualized musculoskeletal models were created using the anybody managed model repository (AMMR) (version 2.0) and the Twente lower extremity model (TLEM) 2.0 (

We evaluated the hip flexion, hip abduction, hip rotation, knee flexion, and ankle flexion of the right leg in a single model for each movement. First, the simulation output from AMS was trimmed based on the knee flexion angles. As such, simulation output denoting subject immobility was rejected (

Once the pre-processing was completed, the registrations were parameterized. Therefore, all data was mean centered to examine the variability. To extract the leading dimensions in the kinematic curves, principal component analysis (PCA) of waveforms was applied. PCA of waveforms has been widely used in the literature for the modeling of gait curves (

The number of significant PC was derived by means of the rank of roots algorithm (

Canonical correlation (CCA) and partial least squares regression (PLSR) are highly related to each other. However, the emphasis is slightly different. In CCA, the aim is to maximize correlation and to allow for a statistical interpretation of this correlation. CCA is a useful tool to understand the relationship between multiple explanatory variables and a set of response variables (

Given the profound dominance of size in statistical shape models (

Partial least squares regression was used to predict the kinematic modes starting from the shape modes or the subject characteristics. To minimize overfitting of the data, only one partial least square regression component was used. To assess the regression fit, reconstruction errors of the shape-specific kinematic predictions were benchmarked again the RMSE when imposing the average curve for all subjects. Again, differences were tested by means of the two-tailed pairwise

Reconstruction errors of the statistical kinematic models vary between 1.09° and 3.54° and are listed in

Root-mean-square errors (RMSE ± standard deviation) from the kinematic parameterization.

City bike | 1.09 ± 0.07 | 1.56 ± 0.08 | 1.62 ± 0.08 | 1.44 ± 0.09 | 2.09 ± 0.13 |

Race bike | 1.12 ± 0.08 | 1.51 ± 0.10 | 1.78 ± 0.10 | 1.22 ± 0.08 | 1.82 ± 0.12 |

Squat | 2.58 ± 0.18 | 1.53 ± 0.08 | 2.19 ± 0.13 | 2.21 ± 0.15 | 1.80 ± 0.09 |

Lunge | 2.82 ± 0.11 | 2.48 ± 0.12 | 2.77 ± 0.13 | 3.54 ± 0.17 | 3.33 ± 0.14 |

Step up | 1.79 ± 0.09 | 1.91 ± 0.08 | 1.87 ± 0.10 | 2.12 ± 0.16 | 2.38 ± 0.10 |

Step down | 1.82 ± 0.09 | 1.77 ± 0.08 | 1.93 ± 0.09 | 2.07 ± 0.09 | 2.16 ± 0.08 |

First principal component of all the parameterized kinematic models based on PCA. Model training curves were first aligned by means of CR. The black line depicts the average motion curve, while the green and blue dashed line represent 2 standard deviations (SD) of the first kinematic mode.

A combined shape model was introduced involving the femur, tibia and fibula bone geometry. Herein, the first 10 modes reproduce 95% of shape variance in the data and all of them are significant according to the rank of roots permutation test. The 3 dominant principal components from the 4 geometry models are shown in

Shape modes from the personalized bone shape models of pelvis, femur, shank (tibia + fibula), and femur and shank combined. Averaged geometry is displayed at the top while variation is represented with a color scheme (average shape ± 3 standard deviations of the shape principal components).

Canonical correlation analysis between the significant shape PC weights or biometric variables and the first kinematic mode.

Age, length, and weight | Pelvis bone shape | Femoral bone shape | Tibia and fibula bones shape | Femur and shank bones combined | |

Correlation measure | r^{2} (p) |
r^{2} (p) |
r^{2} (p) |
r^{2} (p) |
r^{2} (p) |

City bike | 0.1479 (p = 0.036) | 0.7012 (p = 0.060) | 0.2420 (p = 0.249) | 0.2170 (p = 0.355) | 0.3578 (p = 0.245) |

Race bike | 0.0911 (p = 0.178) | 0.6758 (p = 0.157) | 0.2401 (p = 0.291) | 0.2248 (p = 0.357) | 0.4127 (p = 0.128) |

Squat | 0.0433 (p = 0.543) | 0.7399 (p = 0.078) | 0.2769 (p = 0.210) | 0.1502 (p = 0.779) | 0.4156 (p = 0.174) |

Lunge | 0.2559 (p = 0.002) | 0.7787 (p = 0.009) | 0.3364 (p = 0.053) | 0.1731 (p = 0.619) | 0.4902 (p = 0.026) |

Step up | 0.0643 (p = 0.331) | 0.5881 (p = 0.498) | 0.1873 (p = 0.544) | 0.2570 (p = 0.227) | 0.4555 (p = 0.057) |

Step down | 0.1129 (p = 0.093) | 0.6210 (p = 0.276) | 0.1202 (p = 0.850) | 0.1795 (p = 0.549) | 0.2943 (p = 0.509) |

Finally, the performance of the PLS regression is presented in

Partial least squares regression of the demographics and the combined femur and shank model PC weights to predict ADL kinematics.

The canonical correlations between the shape modes and kinematic modes are weak, even for the first mode which is predominantly representing the overall size. The association between the set of subject characteristics and kinematics is also found to be weak with still less explained kinematic variance, lower correlation coefficients and even less prediction ability. As such, our findings are similar to the trial from

While the presented methodology could not demonstrate important shape related variability in motion patterns, future work is mandatory to investigate such in pathological mixed groups, where the impact of bone geometry abnormalities is likely to be of higher importance. The presented methodology seems adequate to investigate these patterns.

The results of our research are affected by some limitations. Although further investigations might extend our findings, our study cohort represents a very selected group, especially with minimal age and height differences among subjects, and accordingly their kinematic variability was limited (

Secondly, our findings are specific to the five model-derived joint angles approach used and the way the bone geometry is taken into account in the multibody kinematics optimization process. As such the study design aimed for the detection of obvious and large scale kinematic features such as walking with toes pointed outward. Our findings can therefore not be entirely generalized. For example, subtle relationships have been previously reported in the literature between joint shape and 6 degrees of freedom tibio-femoral kinematics (

Furthermore, this study relies on the assumption that subject-specific motion can be predicted by a small set of parameters. Reconstruction errors of the statistical kinematic models generally range around 2 degrees, which corresponds to the inter-session error in the gait study from

Lastly, the sparse amount of data remains a major drawback in our investigation. Therefore, the regression analysis should be interpreted cautiously, and one must be aware of the potential risk of overfitting. Moreover, the pelvis shape model is notably less compact than the other models and therefore less suitable to regression analysis, particularly when having restricted numbers of training samples. Thereupon, pelvis bone morphometry was not incorporated into our combined shape model. Even though intra-subject variability was minimized by CR, no obvious patterns could be found here to link between bone morphometry and observable patterns in motion tasks. Alternatively, prediction performance may improve using deep learning methodology, however, such would require sample sizes to be substantially forced up (

In conclusion, motion curves are not prominently related to subject characteristics or personal bone geometry in the present study. Furthermore, when benchmarked against average kinematic reference curves, personalization based on bone geometry appears to lack in clinical relevance.

The datasets generated for this study are available on reasonable request to the corresponding author.

The studies involving human participants were reviewed and approved by Ghent University Hospital Ethics Committee. The patients/participants provided their written informed consent to participate in this study.

JD, KD, and EA designed the algorithms. JD, JV, and EA assisted in the data collection and manipulation. JD and KD carried out the statistical analysis. JD wrote the first draft of the manuscript. All authors contributed to the manuscript revision and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.