METATHESIS Project

Title	METATHESIS: Modelling pathological gait with machine learning for treatment selection support.
Dates	01.09.2024 – 31.08.2028
Principal investigator	Alexandros Kalousis (HES-SO Genève)
Other investigators	Stéphane Armand (UNIGE/HUG), Lionel Blonde (HES-SO Genève), Marys Revaz (UNIGE/HUG), Hugues Vinzant (UNIGE/HES-SO Genève)
Institutional collaborations	DMML (HES-SO Genève), K-LAB (UNIGE/HUG),
Funding	815,897 CHF (FNS Health Research and Wellbeing UAS and UTE - 2023)
Keywords	Reinforcement learning, Human gait analysis and modelling, Kinesiology, Machine learning, Simulation, Pathological gait, Generative modelling
Website	Not available.
Related articles	Coming soon!

Abstract
Gait disorders have a large prevalence and induce a dramatic burden on healthcare and are widely spread within the patient population ranging from the elderly and adults (e.g., stroke (CVA), myopathies, multiple sclerosis (MS)) to children with Cerebral Palsy (CP). Gold standard practices for functional diagnosis and treatment selection require analysis by highly specialized personnel with deep technical knowledge and have high costs. The clinical experts assess the gait function by means of instrumented gait analysis, referred as Clinical Gait Analysis (CGA). These assessments are complemented by a clinical evaluation of the neuromusculoskeletal impairments. However, an accurate personalized functional diagnosis, i.e., identifying causes of gait deviations in the neuromusculoskeletal impairments, is difficult to establish and very often several treatment plans may be eligible for reducing motor disabilities. Selecting the most appropriate plan is a complex non-standardized endeavor. Different medical teams can use different decision-making processes, based on the clinical assessment of the patient and, to a large extent, on the physicians' knowledge and expertise, often leading to subjective decisions. Foreseeing the effect that a treatment will have on a patient's locomotion is quite challenging. As a result, outcomes are not always as desired.CP is the most frequent motor disability of childhood, affecting 17 million worldwide. The estimated healthcare and socioeconomic lifetime costs for CP subjects are between 800,000 and 860,000. The high treatment costs, the irreversibility of some interventions, and the large social impact of the treatment on the life of the patients and their caregivers motivates additional investments in the process of clinical decision making; 25% of CP patients whose treatment followed the CGA-based surgical recommendations, experience a negative outcome and only 37% of the cases have a clinically meaningful improvement over natural progression. Additionally, over the last decades, the proportion of positive outcomes of orthopedic surgeries that improve gait function in children with CP has stagnated. In this project, we will tackle heads on the challenges that the treatment of gait disorders has, starting with applying it to CP.It is therefore the objective of the project to: significantly improve the current medical practice and the treatment outcomes for gait disorders through patient tailored diagnosis and treatment by significantly reducing the number of cases that lead to a negative treatment outcome. To do so we will support clinicians in: (i) developing a structured and holistic understanding of a patient's condition, and in (ii) selecting the optimal treatment for any patient.We will address these objectives by developing the digital patient twin to deliver accurate, robust, and explainable gait modeling and treatment selection support. Central to the development of the digital patient twin will be the use of machine learning and in particular generative learning and reinforcement learning algorithms and models. We will train these models on patient data from the clinical partner which include, information on neuromusculoskeletal impairments, diagnosis, gait data (e.g., kinematics, kinetics, EMG, video), medical imaging and notes, information on administered treatments and patient Follow-up.Digital patient twin: We will develop patient-tailored models to support treatment selection. The overarching approach that we will follow is given a patient's description, namely clinical information as well as musculo-skeletal information, the learned models should reproduce the patient's gait as this is captured by the kinematics curves. By modifying different parameters describing the patient, e.g., muscle lengthening and stiffness, we will be simulating the effect of different treatments and the trained models will output the expected kinematics. We will inform and constrain the learning models using domain knowledge. We will consider generative models with latent variables such as Variational Autoencoders. Parts of the latent space can be assigned specific semantics corresponding to factors, e.g., neuromusculoskeletal impairments, that are known to relate to gait deviations. The decoders will incorporate explicit biomechanical models generating gaits that are consistent with physics and human physiology. In the reinforcement learning approaches we will use biomechanical models/simulators to learn patient-specific controllers/policies that accurately reproduce a patient's gait. The learned policies will produce only physics- and physiology-consistent gait. Such models allow us to explore the effect of different treatments by modifying the patient's properties. By explicitly incorporating domain knowledge the learned models will be to a significant extent understandable since they will carry explicit semantics. We will provide further explanations of the models' decisions using saliency that indicate which part of the input data are responsible for the model's decision.The successful achievement of the targeted results of METATHESIS will have a very significant practical impact: to help improve the mobility of several million significantly disabled people, whose overall quality of life will also be considerably improved. Preserving and improving their mobility is a major challenge that our society is facing and treatment outcomes have not improved since several decades. In METATHESIS we aim to deliver innovative tools to address this necessity, providing clinicians with treatment selection support based on explainable gait modeling.