Making Sense of Sensor Data

Every year, approximately 12.2 million people worldwide experience a stroke, placing a significant burden on individuals, families, and healthcare systems. Despite rehabilitation, walking problems remain common, impacting quality of life. Inertial Measurement Units (IMUs) are small, portable sensors capable of capturing gait features such as asymmetry, which are not assessed in conventional tests. This thesis explored how IMU-based gait measurements in clinical stroke rehabilitation can monitor progression and improve prognostic models, thereby personalizing rehabilitation. Before evaluating their added value, we examined the reliability of IMU-based gait and balance measurements—an essential prerequisite for clinical application. In Chapter 2, we assessed the test–retest reliability of gait features measured with IMUs during a 2-minute walk test (2MWT) in 31 individuals after stroke. Three IMUs were placed on both feet and the lower back. Results showed that gait could be measured reliably during rehabilitation. In Chapter 3, we assessed the reliability of balance features in 40 participants measured with one IMU across five balance conditions (1 sitting, 4 standing) with a 24-hour retest interval. Both sitting and standing balance could be measured reliably in people after stroke. In clinical rehabilitation, gait speed is often used to estimate walking ability in daily life. Chapter 4 evaluated whether IMU-based gait features provide additional value beyond gait speed for estimating daily-life walking ability. Longitudinal data included a 2MWT and two days of daily-life measurement with an ankle-mounted IMU. Gait speed was weakly to moderately associated with daily-life gait features, and IMU features did not improve estimations, suggesting a difference between walking capacity and performance. Calculating IMU-based gait features involves theoretical assumptions that may not always hold. In Chapter 5, we explored a data-driven approach using Variational AutoEncoders (VAEs) to reduce IMU data to a small set of latent features. Four latent features had good–excellent reliability and differed significantly between stroke patients and healthy controls. However, group-level changes during rehabilitation were limited. Chapter 6 mapped gait recovery over time and examined associations between patient characteristics and the rate of change. Participants were measured every three weeks, collecting gait features, balance features, and clinical test outcomes. About 40% improved gait speed, but other gait features showed limited change. Associations between patient characteristics and recovery rate were minimal, possibly due to large individual variability hindering group-level analysis. In Chapter 7, we predicted community walking ability six months after stroke using patient characteristics, clinical tests, and IMU-based gait and balance features. Significant predictors included age, trunk stability (TCT), affected leg strength (MI), and initial gait quality and quantity. Admission assessments had greater predictive value than discharge assessments, suggesting early measurements can effectively predict long-term walking outcomes. In conclusion, this thesis examined the clinimetric properties of IMU-based measurements of gait and balance in clinical stroke rehabilitation. While IMUs can reliably measure these features, we found no added value of gait quality measures beyond conventional tests. In contrast, measuring gait quantity is simpler, yields easily interpretable outcomes, and may provide additional insights into movement quality. Future research should focus on structural implementation of IMU-based gait quantity measurements to enable more personalized rehabilitation strategies for people after stroke, both short- and long-term.