INTEGRATING WEARABLE GAIT ANALYSIS FOR INFORMED DECISION MAKING IN LATE-STAGE OSTEOARTHRITIS: A FRAMEWORK FOR FREE-LIVING ASSESSMENT

Abstract

Late-stage knee osteoarthritis (OA) is a growing musculoskeletal disease affecting millions of older adults. Objective clinical assessments may be a means to improve surgical outcomes but often requires dedicated laboratory equipment and space. Wearable sensors may be easier to collect but are limited by more complex analysis and interpretation of results. This thesis introduces a modular, open-source framework that for collecting gait with bilateral shank inertial measurement units (IMUs). The processing pipeline automates data alignment, incorporates deep learning gait segmentation, stride event detection, and metric extraction, enabling seamless analysis across laboratory and free-living settings. Three studies establish the framework’s value. Study 1 retrained an existing ResNet + BiLSTM using healthy and OA datasets. The model reached ~97 % classification accuracy and decreased walking bout fragmentation compared to a heuristic frequency method, especially at slower walking speeds. Study 2 demonstrated strong in-lab agreement between motion capture- and sensor-derived spatiotemporal and kinematic variables. However, week-long free-living recordings revealed systematically slower and more variable gait, confirming that laboratory snapshots may overestimate real-world mobility. Notably, peak mediolateral shank angular velocity, a native IMU metric, remained well-correlated with Oxford Knee Score, highlighting its clinical promise. Study 3 delivered the first longitudinal, head-to-head sensitivity comparison between measurement systems in 42 arthroplasty patients. Metrics from motion capture were able to capture early postoperative gains, whereas data from IMUs tracked day-to-day function. Collectively, these findings show that pairing laboratory precision with ecological breadth from inertial sensors could yield a richer picture of OA gait than either modality alone, while also demonstrating strengths and weaknesses of both measures. The framework’s sensor-agnostic design, evidence for clinically relevant native IMU variables, and demonstration of complementary sensitivity advance the field toward scalable, data-driven monitoring and personalised rehabilitation.