Machine-learning-based biomarkers towards customization of robotic rehabilitation treatments for stroke patients

• Ano de defesa: 2021
• Autor(a) principal: Moretti, Caio Benatti
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Acidente vascular cerebral; Aprendizado de máquina; Biomarcadores; Biomarkers; Machine learning; Rehabilitation robotics; Robótica de reabilitação; Stroke
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-24062021-160523/

Coupled with sensors, robotic devices for Stroke rehabilitation describe the motor behavior of patients as kinematic and kinetic data, which are underexplored in the data science and machine learning context, due to the time-consuming task of pursuing enough data volume. Moreover, the definition of biomarkers for a quantitative and deterministic assessment of patient progress remains an open problem in the literature. Four different studies were carried out aiming to address such an issue. We also propose a modular framework for organizing pieces of software for data analysis with the calculations of more than twenty robot-based metrics implemented. Our first study concerns a clinically-naive method for defining a region in the data space, alluding to a state of rehabilitation, based on the uncertainty in the classification of hemiparetic sides of chronic stroke patients. Our second study raised evidences that anodal tDCS may have a detrimental or maladaptive interaction with the affected hemisphere in patients with very severe upper-extremity impairments. Our third study correlates the implemented robot-based metrics with traditional clinical scales, so trained machine learning models can play the same role on a quantitative and deterministic way, eliminating the subjective nature from traditional evaluation methods. We found that having a model to estimate clinical scales only from one kind of robot (shoulder/elbow or wrist) is as good as combining data from both. We found in our fourth study evidences from a clinical perspective on the prediction of clinical outcomes of patients at early stages of the treatment. Our results indicate the possibility of improving the decision-making process by alerting, at the end of the second therapy session, when patients will potentially not present a significant response. The four projects here described enabled to push the state of the art towards the development of biomarkers to evaluate and track patients progress on rehabilitation robotics treatments; proposing standards for simplifying the data sharing; simplifying clinical studies with traditional statistical tools and our framework API and optimize the clinicians decision-making process towards impacting patients budget on rehabilitation treatments for optimizing quality of life.