PTOtherJournal of neuroengineering and rehabilitation2021

Automatically evaluating balance using machine learning and data from a single inertial measurement unit.

Fahad Kamran, Kathryn Harrold, Jonathan Zwier and 4 others

PMID 34256799

WHAT IT FOUND

A machine learning model using sensor data rated balance exercises more accurately than patients' own self-assessments.

It better identified when exercises were too difficult, potentially making home-based training safer by catching overconfidence before patients attempt unsafe movements.

Key findings

01The best machine learning model achieved an AUROC of 0.806 and accuracy of 56.4%, outperforming patient self-assessments which had an AUROC of 0.665 and accuracy of 43.3%.

02Patients tended to be overconfident in their self-ratings, often rating difficult exercises (label 5) as moderate (label 3 or 4), whereas the model correctly identified these high-risk repetitions.

03Models using raw sensor data (image representation) significantly outperformed models using hand-engineered features in terms of AUROC, suggesting spatial sway patterns are more informative than temporal ones.

STILL TO COME

How it was doneWhat they foundWhat it means for PTs

Read the rest of this summary

You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.

Already have one?

What it does not show

Very small sample size (10 participants) and limited exercise set. Low inter-rater reliability among therapists (Krippendorf's alpha 0.013), meaning the 'ground truth' labels were noisy and inconsistent. Models performed poorly on intermediate difficulty ratings (3 and 4), limiting their utility for fine-tuning exercise progression. The study validates the algorithm's ability to mimic therapist ratings, not its ability to improve actual patient outcomes or prevent falls. Data was collected in a controlled lab setting with safety harnesses, not in the home environment where the tool would be used.

Declared interests

The authors declared no competing interests. The study was supported by non-U.S. government funds.

The easy way to misread this

Do not interpret this as evidence that a commercial sensor system is ready for clinical use or that it has been proven to reduce falls. This was a small validation study of algorithms in a lab setting, not a clinical trial of patient outcomes.

Read it on PubMed →