Title: Longitudinal Multimodal Sensing of Physical Activity and Well-Being in Older Adults

Abstract:

The deployment of wearable and mobile technologies facilitates the ongoing, real-time surveillance of human health and behavior within naturalistic environments. Nevertheless, developing predictive models from longitudinal, multimodal datasets poses significant difficulties, especially when the objective involves complex or clinically based endpoints. This study introduces a longitudinal, multimodal investigation involving 66 older adults, executed under authentic, real-world conditions. The research integrates clinical evaluations with behavioral monitoring and data collected via wearable sensors, offering a distinctive chance to examine a demographic group that is frequently underrepresented in long-term, "in-the-wild" studies.

Leveraging this unique dataset, we analyze how the correlation between captured signals and target variables influences predictive accuracy across various health-oriented tasks. We establish a comprehensive evaluation framework that encompasses objectives with varying degrees of observability: predicting Activity Levels, estimating Sleep Duration, and classifying Sleep Apnea Severity. Our findings demonstrate a distinct hierarchy of predictability. Behavioral targets that are highly observable yield strong results, achieving a macro-F1 score of 65%. In contrast, more abstract health outcomes prove difficult to predict, although they show steady progress compared to baseline models. Furthermore, explainability analyses indicate that historical features consistently serve as the most valuable predictors, underscoring the critical importance of longitudinal data in these assessments.

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