Search

Profiling the daily activity of a patient in-community is one of the solutions to the world's ballooning healthcare costs and an aging treatment system that is limited by access to care, the inability to monitor home-based practice to provide feedback and the lack of measurement tools that reveal progress. Research thus far has been focused on small scale activity classification that does not address the real challenges: 1) deployment to large and diverse user communities leads to degraded classifier performance due to large activity models; 2) lack of personalization and the inability to train persons involved to use complex systems; 3) activity classification alone cannot deliver the information required by caregivers to scrutinize the skillfulness of movements, determine the progression of recovery and evaluate the patient's overall quality of life.Using wearable inertial sensors, the research in this dissertation provides methods and architectures required for an easy to deploy, low cost end-to-end system that is capable of providing clinically relevant, personalized daily activity profile at multiple levels of granularity: 1) at the highest level, information on the location a person was able to visit; 2) within each location, information about the activities a patient was able to perform; 3) at the lowest level, motion trajectory of each activity with visualization and metrics. These allow physicians to assess and provide feedback on a patient's ability to socialize, their level of exercise tolerance and their compliance to prescription. First, we focused on improving the state of the art in activity classification by introducing a multimodal, hierarchical activity classification toolkit that is less susceptible to performance degradation with large models. Second, we proposed a context-driven, personalized, targeted activity monitoring methodology. Through the definition of context and scenarios, this approach provides personalization, context information and activity to caregivers and further enhances the performance of a traditional activity classifier in terms of speed, accuracy and sensor energy usage. Through multiple iterations, the final system features novel automatic context identification using energy efficient, WIFI augmented GPS. We also developed a prescription model that enables caregivers to prescribe sensors, smartphone and monitoring plan to an out-patient along with the rehabilitation treatment.Third, we developed novel methodologies and implementations required to perform motion tracking and metrics computation for both generic upper and lower body motions and exercise specific motions. The methods developed make use of in-depth biomechanical knowledge and are robust to pathological movement patterns and can provide visualization using a global reference frame without user interaction.Finally, we presented an end-to-end system architecture that synergizes the various components to provide the multi-tiered daily report required. Verification and evaluation of individual component demonstrate the effectiveness of all of our methods individually and the full system is evaluated to show that it is capable of operating with minimal user interaction and training.