Generating fake-but-anatomically-correct movement data turns out to be a decent fix for one of physical therapy AI's persistent problems.
A team of researchers has published a method that uses musculoskeletal simulations to produce synthetic data for inertial measurement unit (IMU) sensors — the accelerometers and gyroscopes embedded in wearables and phone-based fitness tools. The approach modifies movement trajectories within anatomically plausible limits, then labels the resulting data automatically by combining inverse kinematics with a rule-based scoring system. Tested across four datasets of varying complexity, the augmented data closely resembled real-world recordings and improved classification accuracy, generalization to new subjects, and fine-tuning from just a handful of patient examples.
This matters because the data problem in physiotherapy AI is genuinely hard: patients move differently, clinicians disagree on what "correct" form looks like, and collecting labeled reps from real patients at scale is slow and expensive. Synthetic augmentation sidesteps the collection bottleneck and removes the need for a human to score every repetition — but only if the fake data is physically believable, which is exactly what the simulation constraint is designed to enforce.
The researchers are careful to note that the gains vary depending on how balanced and unambiguous a dataset already is — meaning this is not a universal fix, and publications that apply it to clean benchmark data will likely see smaller improvements than clinics working with messy, imbalanced real-world collections. The method is promising, but the gap between a working research prototype and a tool a physiotherapist can actually trust is still doing a lot of heavy lifting here.