In researching neurological injury and disease, it is imperative that we measure the right theoretical constructs in the right way. For instance, in a genome-wide association study of “recovery” following stroke, the hits we get across the genome will depend on how that recovery phenotype is defined. The NIH Stroke Scale is one of the most widely used instruments for measuring stroke severity and recovery, but when aggregated to a total score it loses specificity (i.e., two people can have the same total score for very different reasons). And when taken at the individual item level it loses sensitivity (i.e., the 5-level ordinal motor score for the arm has strong ceiling/floor effects and lacks resolution compared to continuous measures). Thus, we often face constraints between the informational value of a measure relative to the cost of its collection. To that end, I have worked on many different projects exploring the costs/benefits of measuring behavioral phenotypes in neurology. 

Example Publications:

•  Lohse, KR, Miller, AE, Bland, MD, Lee, J-M, & Lang, CE (2025). Association between real-world actigraphy and post-stroke motor recovery. Stroke. https://doi.org/10.1161/STROKEAHA.124.050229

• Lang CE, Holleran CL, Strube MJ, Ellis TD, Newman CA, Fahey M, DeAngelis TR, Nordahl TJ, Reisman DS, Earhart GM, Lohse KR, Bland, MD. Improvement in the capacity for activity versus improvement in performance of activity in daily life during outpatient rehabilitation. Journal of Neurologic Physical Therapy. 2023 Jan 1;47(1):16-25.

• Aldridge, CM, Braun, R, Lohse, KR, de Havenon, A, Cole, J, Cramer, SC, Lindgren, AG, Keene, KL, Hsu, F-C, & Worrall, BB. (2024). Genome-wide association studies of three distinct recovery phenotypes in mild stroke. Neurology. 102(3). doi: https://doi.org/10.1212/WNL.0000000000208011

Rehabilitation is fundamentally about change within a person over time. Statistically speaking, however, these temporally dependent data violate the assumptions of many statistical tests and require specialized tools for analysis. People often use the term “longitudinal data” to refer to few data points (e.g., <10) that are collected over a long timescale (e.g., days or months apart). In contrast, people use the term “time-series data” to refer to large numbers of data points (e.g., hundreds to millions) that are sampled at a very high density (e.g., milliseconds or microseconds apart). Although these different data types actually exist on a continuum, it is still useful to talk about them separately as differences in the sampling rate and number of observations make them amenable to different types of analyses. For instance, time-series data can be transformed into the frequency domain with Fourier analysis and future data can be predicted with various autoregressive models. In contrast, longitudinal data might be analyzed with linear or non-linear mixed-effect regression models, with the specifics of the model depending on the nature of the outcome (e.g., binary, ordinal, or interval/ratio) and the “shape” of the trajectory (e.g., linear, exponential, or sigmoidal). Although we do not create these mathematical tools, we apply them to rehabilitation problems and create instructional materials for rehabilitation researchers to use them. 

Example Publications:

• Albertson, AJ, Landsness, EC, Eisfelder, M, Young, BM, Judge, B, Brier, MR, Euler, MJ, Cramer, SC, Lee, J-M, & Lohse, KR (in press). Stroke is associated with regional and age-specific changes in periodic and aperiodic cortical activity. Experimental Physiology. https://doi.org/10.1101/2024.11.07.622359   

• Lohse KR, Kozlowski AJ, Strube MJ. Model Specification in Mixed-Effects Models: A Focus on Random Effects. Communications in Kinesiology. 2023. https://doi.org/10.51224/cik.2023.52

• Lohse KR, Shen J, Kozlowski AJ. Modeling longitudinal outcomes: A contrast of two methods. Journal of Motor Learning and Development. 2020 Jan 27;8(1):145-65.

As with many fields, rehabilitation has seen astronomical growth in the amount and complexity of the data we produce. For instance, physiological data from EEG or accelerometry data from inertial sensors contain highly structured data (e.g., voltages or forces in discrete intervals of time) in very dense samples (e.g., 250-1,000Hz for minutes or hours of recording). In contrast, electronic health records contain loosely structured data from millions of individuals all with complex data types that all have unique relationships to each other and may or may not be recorded over time. This means that researchers and their students are facing increasingly large and complex data sets. In my research group, we want to give researchers the tools and training to work with their own data effectively. More than any one project, we want to make sure that data are Findable, Accessible, Interoperable, and Reusable (FAIR) in rehabilitation science. As part of that effort, I am part of the educational leadership team for the Reproducible Rehabilitation (“ReproRehab”) program funded by NCMRR, I collaborate with other researchers at WUSTL to harmonize and archive large research datasets, and I am the Director of the Resource Core for the NIH-funded (P50) Data Science and Analytics for Precision Rehabilitation Center (DAPR, https://dapr.usc.edu/). 

Example publications:

• Lohse, K.R., & Kliethermes, S. (2025). Approaching significance: Statistical guidance for authors and reviewers. Journal of Neurologic Physical Therapy. https://doi.org/10.1097/npt.0000000000000526

• Lohse, K.R. (2025). Taking steps toward open data in motor control, learning, and development. Journal of Motor Learning and Development. https://doi.org/10.1123/jmld.2024-0081

• Lohse KR, Schaefer SY, Raikes AC, Boyd LA, Lang CE. Asking new questions with old data: the centralized open-access rehabilitation database for stroke. Frontiers in Neurology. 2016 Sep 20;7:153.

  • The interactive SCOAR website: https://keithlohse.github.io/SCOAR_data_viz/.