To learn or not to learn? Model-based estimation of motor learning in a large, "in the wild", home rehabilitation database.

Motor learning, defined as practice-related processes leading to relatively permanent changes in response capability, is critical for regaining motor function following brain injury. However, quantifying learning in real-world, unsupervised settings remains challenging. Here, we present a model-based framework for estimating the presence and extent of motor learning across multiple tasks and users from a large "in-the-wild" rehabilitation dataset collected via the FitMi sensor system. We analyzed 4,661 episodes from 398 users practicing 20 upper-limb tasks for at least 50 sessions. For each episode, we modeled the effect of daily dose (repetitions) on performance (repetitions per second) using a discrete-time first-order state-space model, in which a latent motor memory evolves through cumulative practice. The model employed a learning rate for memory updates and a forgetting time constant for decay. Simulation-based recovery confirmed the robustness of this procedure despite realistic noise and irregular practice schedules. We utilized a likelihood ratio test (LRT) to compare this learning model against a non-learning null model. Overall, 38.3% of episodes were classified as learning. Among these, 44.6%, which correspond to 17.1% of all episodes, exhibited a time constant exceeding 30 days, consistent with durable motor memory. In contrast, non-learning episodes exhibited short time constants (under 2 days), reflecting transient fluctuations. Task-level analysis revealed no clear dichotomy between learnable and non-learnable tasks. Among users practicing at least five tasks, 18% were non-learners and only 5% exhibited learning on all tasks, suggesting that generalizable improvements were rare. Our findings demonstrate that real-world rehabilitation induces detectable motor learning in specific tasks and users. Future research will incorporate user-level covariates, characterize higher-order learning dynamics, and validate findings against clinical outcomes to establish a link with functional recovery. These results offer a path toward optimizing post-stroke recovery through individualized task selection.