New Protocol Evaluates Deep Imputation for Wearable Device Data.

Wearable devices offer continuous health monitoring, but the data they generate often suffers from structured missingness, where features from the same physical sensor drop out simultaneously. Existing deep imputation methods like BRITS and SAITS have not been thoroughly evaluated on multimodal physiological data under these realistic missingness conditions. Current benchmarks typically use random-point holdout protocols, which incorrectly assume missingness is independent across features and time. This research introduces a new evaluation protocol designed to address these limitations. Using data from an epilepsy patient recorded via a Garmin smartwatch, the protocol mines contiguous missing-run templates from training data. These templates are then stratified by per-feature gap-length quantiles and injected as block masks, preserving the co-missingness structure inherent in real-world wearable data. The study found that a matched training protocol, which exposes models to the same realistic missingness distribution, significantly improved BRITS's performance, reducing severe-gap MAE by 43%. While no single model dominated, linear interpolation was optimal for slow-moving features over short gaps, extended BRITS excelled with dynamic cardiac features in moderate and severe gaps, and SAITS better preserved the ground-truth distribution despite higher MAE. The findings underscore that model rankings are highly dependent on evaluation designs, establishing critical steps for developing superior imputation strategies for future multi-sensor wearable datasets.