DEEP LEARNING IN STROKE REHABILITATION COMPENSATORY MOVEMENT DETECTION: FROM LABORATORY BENCHMARKS TO CLINICAL REALITY

Stroke rehabilitation is going through some pretty major changes right now. We're seeing a gradual move away from those subjective clinical scales toward more objective, automated assessments—though it's not happening as fast as some people hoped. Traditional methods like the Fugl-Meyer Assessment (FMA) are still considered the gold standard in clinics, but they've got some real problems. They're time-consuming, different raters often give different scores, and they tend to hit ceiling effects that miss subtle improvements patients are making.This review looks at important studies from 2016 to 2025. We trace how automated detection systems evolved from those lab-based depth sensors to RGB cameras you can use at home and wearable IMUs. We compare different sensing technologies and look at public datasets. There's also this "sim-to-real" gap issue - basically, models trained on healthy actors often don't work well with real patients, which is a big problem. The algorithms have changed a lot too, going from traditional machine learning to foundation models. Clinical metrics have also expanded beyond simple detection to things like severity grading.At the end, we talk about what this all means for tele-rehabilitation and why getting clinicians to actually adopt these technologies is still really challenging. We suggest a roadmap that focuses on privacy-preserving collaboration and long-term validation studies, though honestly implementing all of this won't be easy.