JMIR Publications

ABSTRACT

Background:

Nonintrusive monitoring through WiFi signal sensing (WiFi sensing) has been proposed as an alternative to wearables and cameras to support aging in place while preserving privacy. However, a persistent gap remains between laboratory performance and real-world deployment, where the scarcity of labeled data constrains supervised approaches.

Objective:

This study aims to evaluate the operational feasibility and perceived usefulness of a personalized, unsupervised WiFi-based anomaly detection system in a real-world, longitudinal deployment. Specifically, we examine how this technology can support care, follow-up, and social support for older adults by identifying behavioral deviations without the need for wearable devices or cameras.

Methods:

We deployed a WiFi-based monitoring system in the homes of 32 older adults (analytic cohort defined by availability of valid sleep and activity data). The analytics engine combined (1) a consensus novelty detection ensemble (CNDE) with SHAP-based explainability for acute sleep deviations and (2) trend analysis using moving average convergence divergence (MACD) for sustained changes; additionally, Prophet was used to detect atypical durations of stays. We distinguished technical alerts (device status/connectivity; resolved by technical support) from care alerts (signals about routines, sleep, and presence; intended for the care team). Operational outcomes were summarized from alert logs and from quality-controlled sleep/activity data available in the database.

Results:

Median follow-up was 326 days (IQR 247–410). We generated 2,214 care alerts (excluding technical system alerts), predominantly stay-based activity alerts (45.4%) and sleep alerts (daily sleep quality and sleep location alerts; 47.6% combined). Most alerts were mild (75.8%), with smaller proportions of moderate (10.4%) and severe (13.8%). Operationally (without clinical ground truth), the overall mean rate was 0.240 alerts per participant-day (combined sleep/activity exposure), with the burden concentrated in 18.9% of participant-days (≥1 alert). When stratified by domain using stream-specific exposure, sleep alerts showed 0.156 alerts per participant-day (13.9% of participant-days with ≥1 alert) and a higher proportion of high priority (severe+moderate: 42.2%), whereas activity alerts showed 0.126 alerts per participant-day (12.2% of participant-days with ≥1 alert; high priority: 4.3%). Care professionals reported practical usefulness for prioritization, proactive outreach, and early detection of functional changes, supported by multiple success cases reported during follow-up. We present selected illustrative vignettes, including an acute severe sleep anomaly temporally aligned with a reactive anxiety episode and a sustained deterioration trend that prompted follow-up and led to identification of an underlying health issue affecting sleep.

Conclusions:

In a real-world deployment, a personalized, unsupervised WiFi-based approach can be integrated into workflows and can generate actionable signals with a manageable operational load. These findings support the feasibility of unsupervised artificial intelligence as proactive support in gerontechnology focused on care and social support, although prospective designs with formal evaluation protocols for care processes and person-centered outcomes will be needed.