Date of Award
Master of Science (MS)
Successful activity recognition in patients with motor disabilities can improve patient care by providing researchers and clinicians with valuable information on patient movements and quality of life in real-world settings. Understanding the everyday activities of patients is important for rehabilitation. For researchers, having convenient, objective, and continuous data can drastically improve outcome measures to better compare therapies, and ultimately make recommendations. For clinicians, individual assessment of compliance and outcomes outside the clinic can be more objective, permitting much more tailored recommendations to patients. Most importantly, for individual patients, activity recognition can make this improved health care possible by simply having patients wearing a small sensor, minimizing the need for clinical visits but reaping all the benefits of tailored healthcare.
There are many activity trackers available in the market. But most of them have been designed for healthy subjects. Studies have shown that activity tracking systems designed for healthy subjects can perform poorly on mobility-impaired populations, like those with incomplete spinal cord injury (iSCI) due to their unique patterns of movement. Because iSCI patient populations move in distinct ways, algorithms can and should be specifically tailored for them. By applying machine learning to collect movement data from this specific patient population, we demonstrate how an iSCI-specific system can improve activity recognition with this population.
Traditional activity recognition approaches analyze individual clips of accelerometer data to perform activity recognition. These static classifiers are easier to construct, as each clip of data is treated independently, but the structure of events over time is lost. This thesis attempts to improve upon the standard static classification method by augmenting these static classifiers with a dynamic state estimation model—a hidden Markov model (HMM). An HMM takes into account not only the information present in a clip of sensor data, but also the context of that clip over time, which leads to a higher classification accuracy. By using an HMM to go over the predictions made by the static classifier, unlikely sequences of events can be removed and corrected.
Data were collected from thirteen ambulatory incomplete spinal cord injury subjects who were instructed to perform a standardized set of activities while wearing a waist-worn accelerometer in the clinic. Activities included lying, sitting, standing, walking, wheeling, and stair climbing. The accelerometer data was parsed into two-second clips and a standard set of time-series features were extracted from those clips. Those features were then analyzed by a static classifier to produce probabilistic estimates of the likely activity the subject was performing. Those estimates were then input as observations into the HMM to reclassify ambiguous or improbable sequences of activities made by the static classifier. Multiple classifiers and validation methods were used to assess the ability of the machine learning techniques.
Using within-subject cross validation, static classifiers provided a classification accuracy of 86.3%. By adding another layer of a hidden Markov model, the accuracy improved an additional 2.6% to 88.9%. In subject-wise cross validation, a hybrid static classifier and HMM model gave the highest classification accuracy of 64.3%, a 1.2% improvement over the model using only static classifiers. Our prediction accuracy was subtle because we dealt with activities that are almost undistinguishable: sitting and wheeling, walking and stair climbing.
Individuals with impaired movements can benefit from improved activity recognition to more objectively, conveniently, and continuously measure patient outcomes. Such measures support therapists, clinicians, and clinical researchers to select the right physical or drug therapies, and further refine selected therapies to improve mobility in patients.
Sok, Pichleap, "Activity Recognition for Incomplete Spinal Cord Injury Subjects Using Hidden Markov Models" (2016). Master's Theses. 3272.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License.
Copyright © 2016 Pichleap Sok