Walking is a complex phenomenon that requires coordination between the brain and body. Neurological disorders like Parkinson's disease affect brain regions important for motor coordination and this results in impaired walking and severe episodes of "freezing of gait" that interrupt and halt an individual's motion. Currently, therapies for people with gait impairment and freezing in Parkinson's disease are limited, due to the little understanding of the underlying neural or biomechanical mechanisms associated with impaired gait, and the lack of standardized, patient-friendly, objective tools with which to assess impaired gait in the clinic or in natural settings over long periods of time. This dissertation describes tools to measure, understand, and improve human gait. First, we developed and validated a standardized walking task and statistical model with which to assess gait, and then used this to test the efficacy of different neurostimulation therapies for Parkinson's disease. Second, we developed and tested deep learning models to detect gait impairment using wearable inertial sensors, of which the number and placement consider both patient preferences and model performance. Third, we developed and validated a workflow to reliably measure human motion over long durations with wearable inertial sensors.

Reliably eliciting freezing has been difficult in the clinical setting, which has limited discovery of pathophysiology and documentation of the efficacy of treatments. Measuring freezing and impaired gait have largely relied on subjective clinical rating scales or expert observation. We validated an instrumented turning and walking course that mimics real-life environments that elicit freezing, with the clinical gold standard assessment of freezing of gait. Using the walking and freezing data from participants in this course, we also created and tested a statistical model with which to objectively detect freezing from inertial sensors worn on a person’s ankles. With only two inertial sensors and a linear model, we were able to achieve accuracies of 0.75 area under the receiver operating characteristic. The advantage of this model was that it was interpretable, and revealed four gait features related to freezing behavior. Statistical models like these can provide deeper understanding of gait impairment and are practical;​ they can remove the subjectivity and burden from a clinician having to visually assess patient gait.

Deep brain stimulation, a clinical therapy for Parkinson’s disease, has been shown to have variable effects on freezing and gait impairment. Our standardized gait task and statistical model empowered us to quantitatively test the efficacy of different deep brain stimulation paradigms for treating freezing of gait. For a small cohort of participants, we saw average improvements in the percent time spent freezing on both lower (60 Hz) and high (140 Hz) frequencies of stimulation from no stimulation, though some participants improved more on the lower frequency setting. We then developed a novel, adaptive or closed-loop deep brain stimulation system that can switch between stimulation settings based on our statistical model’s prediction of whether a person is freezing. This system. exceeds the inflexibility of continuous neurostimulation systems and offers promise for better managing impaired gait and freezing.

In addition to being control variables in closed-loop therapies, objective measures of freezing are also useful for monitoring gait impairment outside of the clinic. Monitoring freezing in natural environments has the potential to help us better understand underlying mechanism and inform therapy. Before a FOG-monitoring system is developed, it is important to determine a standard inertial sensor set that people with Parkinson’s disease will reliably wear and to quantify such a set’s freezing detection performance. To facilitate translation, we developed an open-source deep learning model that accounts for patient preferences and model performance with high accuracy (area under receiver operating characteristic curve = 0.80). The model also produces clinically-relevant metrics that correlated with individual human rater’s assessments, like percent time spent freezing (ICC = 0.92) and number of freezing events (ICC = 0.93). We present these preliminary results on a small dataset, and would need to validate on a larger, more heterogenous dataset before implementing in the real world. However, this is an important step toward at-home monitoring that could greatly supplement clinical walking assessments and will progress the community toward adopting a general, open-source model for freezing detection.

The ability to measure joint kinematics in natural environments over long durations using inertial sensors could extend the aforementioned work, enabling long-term, at-home monitoring and personalized treatment of neurological and musculoskeletal disorders. However, drift, or the accumulation of error over time, inhibits the accurate measurement of movement over long durations. We developed an open-source workflow to estimate lower extremity joint kinematics from inertial measurement unit data capable of assessing and mitigating drift. We validated the workflow in a healthy cohort of participants as they performed common activities for ten minutes. We show our workflow produced inertial-based estimates of kinematics consistent with optical-based measures of kinematics, the current industry standard for human motion capture, and could mitigate drift even during continuous walking. This workflow offers an alternative to current drift-mitigation approaches that rely on explicit sensor recalibration (e.g. sitting or standing still for a few seconds), which are not always feasible during real-life activities.

This dissertation developed, validated, and shared open-source statistical models, data and tools to better measure, understand, and improve human gait. This work demonstrates the utility of combining clinical, biomechanical, and engineering techniques, with the goal of addressing real patient needs in a way that considers their human experiences. In sharing this work with the community, we hope that others will build upon the work presented here.