Subject

The development of knee osteoarthritis (OA) has been attributed to the overloading of articular cartilage. A common, non-invasive treatment strategy to reduce internal knee loads (i.e. Knee Contact Forces (KCFs)) is gait modification. So far, physiotherapists are not able to observe in real time if the KFCs decrease with the retrained gait pattern or how they change in different activities. In fact, the only way to directly measure the internal forces acting on the knee is to insert a sensor inside it because of its invasiveness, this approach is not feasible. The state of the art techniques for a non-invasive estimation of KCFs are therefore based on complex, time-consuming musculoskeletal modelling and simulation. Dedicated software to compute these simulations, such as Opensim and Anybody, are now well established and widely tested on the few available datasets obtained using sensor-based knee prosthesis.

The standard approach to estimate KCFs during gait can be summarized in the following two steps:
- Data acquisition in the laboratory. The trajectories of the markers placed on the subject are acquired with a marker-based mocap system, simultaneously with the Ground Reaction Forces (GRFs) acquired by force plates embedded in the floor.
- Modelling software. Marker trajectories, GRFs and anthropometric data represent the input of the musculoskeletal modelling software. At the end of the simulation, the KCFs are obtained.

But recent studies [1-6] demonstrated that real time estimation of KFCs can be achieved using a Machine Learning (ML) or a Deep Learning (DL) approach instead of musculoskeletal modelling. These studies predict KFCs taking as input various data: joint trajectories, joint angles, EMG signals, GRFs and anthropometric data. [5, 7] acquire joint trajectories using a markerless approach, and nearly all the studies suggest that excluding EMGs and GRFs from the input data does not significantly reduce the model prediction accuracy. Therefore, the literary review suggests that real time KCFs prediction, even outside a highly equipped gait laboratory, can be achieved adopting a ML/DL approach.

Kind of work

The objective of this master thesis is to evaluate different ML/DL approaches, found in the literature, for real time KFCs estimation.

Training datasets obtained using sensor-based knee prosthesis are scarce, and the generation of training data using modelling software is cumbersome. Therefore, we will focus on the following research questions:
- What is the impact of the training dataset size on the model performances?
- How well the model generalizes across different subjects?
- What data (i.e. EMG and GRFs) can be excluded from the input of the models?
- Can a ML/DL approach outperform or achieve the accuracy of modelling software?

From a more practical point of view, the work will consist of 2 parts:
- Using available datasets of sensor-based knee prosthesis ([8-10]), the student will train and evaluate different ML/DL models proposed by related studies ([1-6]). Since the modelling software are tested on the same datasets ([7]), we will be able to compare their results with ours.
- The student will acquire a small dataset in our state of the art movement analysis laboratory. These data (joint trajectories, GRFs...) will be the input for both the pre-trained models and the Opensim modelling software. The output (KCFs) of the 2 approaches will be compared.

Framework of the Thesis

[1]G. Giarmatzis, et al. Real-Time Prediction of Joint Forces by Motion Capture and Machine Learning. Sensors. 2020.
[2]M. M. Ardestani et al. Feed forward artificial neural network to predict contact force at medial knee joint: Application to gait modification. Neurocomput. 2014.
[3] L. Rane et al. Deep Learning for Musculoskeletal Force Prediction. Ann. Biomed. Eng. 2019
[4] T. Dao, From deep learning to transfer learning for the prediction of skeletal mus-cle forces. Medical & biological engineering & computing. 2019
[5] M. A. Boswell et al. A neural network to predict the knee adduction moment in patients with osteoarthritis using anatomical landmarks obtainable from 2D video analysis. Osteoarthritis and Cartilage. 2021
[6] S. D. Uhlrich et Al. OpenCap: 3D human movement dynamics from smartphone videos. BioRxiv (preprint). 2021
[7] Z. I. Nejad, The Capacity of Generic Musculoskeletal Simulations to Predict Knee Joint Loading Using the CAMS-Knee Datasets, Ann. Biomed. Eng. 2020
[8] https://orthoload.com/comprehensive-data-sample/
[9] https://simtk.org/projects/kneeloads
[10] https://cams-knee.orthoload.com/

Number of Students

1