Health systems are building and buying models to predict readmissions, falls and deterioration. Look at the inputs and you'll mostly find labs, vitals, diagnoses, medications and utilization history.
Function is usually absent. When it does exist, it's buried in clinical notes as free text like "ambulating well," which is close to useless as a feature.
That's a real gap, because how a patient moves is tied closely to independence and to the outcomes patients actually care about. A model can't use a signal it never sees.
The catch is quality. Adding a noisy movement measure doesn't improve a model. It tends to make predictions worse while making them look more confident. Getting this right means standardized tests with the same protocol everywhere, objective capture instead of a stopwatch and a judgment call, structured storage next to the other vitals, and enough readings over time to show a trend.
I go through the clinical side of this in a longer post, which also covers the research behind gait speed as a predictor.
Disclosure: I'm VP of Operations at BioMech Health, which works in movement measurement.
Top comments (0)