Data-driven Gait Analysis Using Dynamic Mode Decomposition for Monitoring Chemotherapy-induced Peripheral Neuropathy

Abstract

Chemotherapy-induced peripheral neuropathy (CIPN) is a prevalent adverse effect of cancer treatment, known to downgrade sensory and motor function and significantly increase fall risk among survivors. Subtle gait alterations—such as reduced step length, prolonged stride times, and increased double-support phases—often accompany CIPN. However, these changes also occur in other conditions, making early detection of CIPN challenging using conventional gait metrics. To address this challenge, this dissertation develops a data-driven framework that links an individual's unique walking patterns to the onset of treatment-induced abnormalities. Central to this approach is the application of dynamic mode decomposition (DMD) and machine learning models to kinetic and kinematic signals during the gait cycle. Particularly, DMD is well-suited for analyzing gait signals, as it provides a low-dimensional, equation-free representation of high-dimensional, time-resolved biomechanical data in an interpretable way. This data-driven gait modeling enables the identification of individual-specific dynamical gait features that are not captured by standard spatiotemporal gait parameters, facilitating early detection of abnormalities. The dissertation comprises three interrelated studies. First, baseline gait features were established for cancer survivors prior to chemotherapy by decomposing their plantar pressure data into dominant DMD mode features—frequency, decay rate, and initial amplitude. These features were used in a machine learning model that accurately identified individuals with 86–89% accuracy, confirming the uniqueness and consistency of each subject's gait dynamics. Second, a longitudinal monitoring framework was developed to detect deviations from baseline using permutation-based statistical tests applied to the distributions of DMD features. This method detected early gait deviations in 87% of participants by their second or third treatment visits, enabling timely identification of emerging CIPN symptoms. Third, the framework was extended by integrating kinetic (plantar pressure) and kinematic (IMU-based acceleration and angular velocity) datasets using various DMD variants and state-space system identification techniques. This multi-sensor approach reveals the impact of foot-ground interaction on whole-body dynamics, offering a more comprehensive understanding of gait adaptation under neuropathic stress. Collectively, these contributions establish a robust, interpretable, and scalable framework for personalized gait analysis. By capturing individual characteristic gait baseline and detecting deviations in the dynamical features, this work offers preemptive, AI-driven diagnostic tools in healthcare, particularly for monitoring conditions like CIPN that impact mobility and quality of life.