Title: Adaptive Data Selection Enhances Wearable Predictions When Baseline Performance Is Low

Abstract:

While adaptive sensing techniques that filter data samples are gaining traction in wearable health platforms to boost predictive accuracy within constrained data limits, their individualized efficacy remains unclear. This study assesses the impact of adaptively choosing time windows for model training under strict measurement constraints, utilizing a longitudinal dataset that incorporates heart rate, activity levels, and ecological momentary assessment (EMA). We measured performance enhancements compared to random sampling via the F1 score and area under the receiver operating characteristic curve (AUROC).

The results indicate that adaptive methods deliver significant AUROC improvements—reaching as high as 0.7—for users with initially poor performance, whereas those with strong baseline metrics experienced minimal or even detrimental effects. A strong inverse relationship was observed between adaptive gains and baseline performance across all sensing modalities (Pearson r = -0.67; Spearman p = -0.62). At the individual level, the majority of participants saw AUROC improvements ranging from 60% to 80% depending on the modality, though F1 score enhancements were more modest and inconsistent.

These findings suggest that adaptive sensing is not universally advantageous; rather, it offers the most substantial value in scenarios where baseline performance is weak. Consequently, our results advocate for targeted deployment approaches that customize adaptive sensing protocols based on a user’s baseline performance to maximize the efficiency of wearable health monitoring.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC