Why do smartwatch health gaps break standard AI repair tools?
A new study shows that missing data from wearable sensors doesn't scatter randomly — it vanishes in blocks when a shared sensor fails. Standard benchmarks miss this pattern, so they rank imputation models wrong.
The problem with how we test
Wearable device data enables continuous health monitoring. But it suffers from structured missingness: features sharing a physical sensor drop out together. Deep imputation methods such as BRITS and SAITS have seen limited evaluation on multimodal physiological data under realistic missingness. Existing benchmarks use random-point holdout protocols that incorrectly assume missingness is independent across features and time.
What the new protocol does
Using data from a person with epilepsy recorded on a Garmin smartwatch, the authors develop an evaluation protocol that mines contiguous missing-run templates from training data. The protocol then stratifies these templates by per-feature gap-length quantiles before injecting them as block masks that preserve the original co-missingness structure. A matched training protocol exposing models to the same missingness distribution reduces BRITS's severe-gap MAE by 43%.
The models behave differently
Linear interpolation suits slow-moving features over short gaps. Extended BRITS lowers MAE on dynamic cardiac features in moderate and severe gaps. SAITS better preserves the ground-truth distribution by Jensen-Shannon distance despite higher MAE. No single model dominates.
Where the explanation stops
Model rankings depend on evaluation designs. The protocol transfers to future multi-sensor datasets, but single-participant results leave scaling across populations, brands, or sensor suites untested.
The sources this article was written from
- "arXiv:2608.24436v1
https://arxiv.org/abs/2608.24436v1"