Automatic multi-IMU-based deep learning evaluation of intensity during static standing balance training exercises.

The 'ground truth' was based on physical therapists watching videos, which may not fully capture the clinical judgment made during in-person interaction. The dataset was skewed toward lower intensity ratings (most exercises were rated 1 or 2), which may have reduced the model's accuracy in identifying very high-intensity exercises. The study included a mix of healthy adults and those with neurological conditions, but the small sample size and specific sensor setup limit generalizability to all patient populations or home environments. The model was trained on static standing exercises only; it has not been tested on dynamic movements or functional tasks.

Funded by the National Science Foundation and the Precision Health Investigators Awards at the University of Michigan. No commercial conflicts of interest were reported.

Do not assume this technology is ready for clinical use or that it replaces therapist judgment. The study validated a model's ability to mimic therapist ratings of exercise difficulty, not the model's ability to predict patient outcomes or fall risk. Furthermore, the accuracy dropped when predicting high-intensity exercises because they were rare in the dataset.