Machine Learning For Enhanced Robotic-Assisted Stroke Assessment
Loading...
Authors
Akbarifar, Faranak
Date
2025-09-30
Type
thesis
Language
eng
Keyword
Stroke assessment , Kinarm Robotic assessment , Machine learning , Deep learning , time series forecasting , uncertainty estimation , Explainability
Alternative Title
Abstract
Stroke remains one of the leading causes of long-term adult disability worldwide, creating an urgent need for precise and objective neurological assessment methods to guide therapeutic interventions. Conventional assessment techniques primarily rely on clinician observations, which are inherently subjective and often lack the granularity required for personalized rehabilitation. To address these limitations, robotic platforms such as the Kinarm Exoskeleton Lab offer detailed, high-resolution kinematic data that promise more accurate, objective, and sensitive evaluations of motor impairment.
In this thesis, we leverage advanced machine learning techniques to extract clinically relevant insights from Kinarm robotic data, addressing four major research challenges: (1) the limited exploration of machine learning approaches for interpreting robotic data, (2) shortcomings of existing one-class normative models which overlook subtle motor impairments, (3) challenges related to label ambiguity that impact generalization, and (4) the lengthy nature of full Kinarm assessments that could fatigue stroke survivors.
First, we successfully distinguished between stroke and control participants using Kinarm kinematics, employing unsupervised representation learning methods that identified and visualized subtle distinctions within clinical categories, revealing fine-grained patterns of impairment undetectable by standard assessment tools. Second, we introduce uncertainty-aware machine learning techniques to refine stroke and control classifications, significantly enhancing the model’s sensitivity to subtle neurological impairments by systematically handling ambiguous data points. Our findings indicate that filtering out the top 10\% most uncertain data samples significantly increases sensitivity in distinguishing minimally impaired stroke survivors from healthy controls. Finally, we investigate foundation models for time series forecasting, enabling the prediction of additional robotic task trials based on limited initial data. This approach significantly reduces the required assessment duration without compromising diagnostic accuracy.
Collectively, this work demonstrates that integrating robotic data with innovative machine learning techniques not only enhances the sensitivity and objectivity of stroke impairment assessment but also paves the way toward personalized and efficient rehabilitation strategies.
Description
Citation
Publisher
License
Queen's University's Thesis/Dissertation Non-Exclusive License for Deposit to QSpace and Library and Archives Canada
Intellectual Property Guidelines at Queen's University
Copying and Preserving Your Thesis
This publication is made available by the authority of the copyright owner solely for the purpose of private study and research and may not be copied or reproduced except as permitted by the copyright laws without written authority from the copyright owner.
Attribution-NonCommercial 4.0 International
Intellectual Property Guidelines at Queen's University
Copying and Preserving Your Thesis
This publication is made available by the authority of the copyright owner solely for the purpose of private study and research and may not be copied or reproduced except as permitted by the copyright laws without written authority from the copyright owner.
Attribution-NonCommercial 4.0 International