Multi-Modal Continuous Monitoring in Skilled Nursing Facilities During COVID-19: A Feasibility Study with Exploratory Biomarker Discovery for Long COVID Prediction

Background: Approximately 35% of COVID-19 deaths occurred in skilled nursing facilities (SNFs), where conventional intermittent vital sign monitoring may fail to detect early physiological deterioration. Long COVID affects 10-30% of survivors, with limited ability to predict which patients will develop persistent symptoms during the acute illness phase. We investigated whether continuous wearable device monitoring combined with the Visensia Safety Index (VSI), a proprietary multi-parameter physiological instability score, could predict patient outcomes including Long COVID in high-risk SNF populations. Methods: In this prospective feasibility study (NCT04548895), we enrolled 21 PCR-confirmed COVID-19 patients (age 77.2±9.1 years, 62% female) from two Washington State SNFs between December 2020 and March 2021. Patients were monitored using Biostrap wrist-worn sensors (SpO₂, heart rate, respiratory rate), Emfit under-mattress sensors (heart rate variability, sleep architecture), and the Visensia Safety Index (11,447 measurements across cohort). We extracted 28 time-series features and applied principal components analysis (PCA) to leverage all 11,447 VSI measurements rather than summary statistics alone. Machine learning models were trained using leave-one-out cross-validation. Long COVID was assessed at 6-month follow-up. Given the exploratory nature and small sample size (n=3 Long COVID cases), we report both uncorrected and Bonferroni-corrected p-values for the 10 VSI metrics examined. Results: Patients were discharged to home (n=8, 38%), palliative care (n=9, 43%), or died during acute illness (n=4, 19%). Among 18 survivors at 6-month follow-up, three (16.7%) met WHO criteria for Long COVID; all three had been transferred to palliative care during acute illness, while zero of eight patients discharged to home developed Long COVID. VSI patterns differed significantly by outcome (p=0.017 uncorrected; p=0.17 after Bonferroni correction for 10 comparisons). Comprehensive HRV feature engineering revealed critical device differences: Emfit captured 100% (3/3) of Long COVID cases with 16 HRV features, while Biostrap captured only 33% (1/3) due to intermittent heart rate sampling incompatible with continuous HRV computation. Conclusions: This feasibility study demonstrates that multi-modal continuous monitoring combining wearables, contact-free sensors, and the Visensia Index is technically feasible and well-tolerated in SNF COVID-19 patients. Continuous monitoring detected deterioration 1-2 weeks before clinical recognition, and trajectory-based analysis leveraging all 11,447 measurements provided convergent evidence for Long COVID prediction. However, the small sample size (n=21 patients, n=3 Long COVID cases) and high feature-to-sample ratio (9.3:1) create substantial overfitting risk. All Long COVID findings should be considered hypothesisgenerating and require prospective validation in adequately powered cohorts (≥50 Long COVID cases) before clinical application. Device selection for Long COVID prediction should prioritize Emfit (complete case capture, native HRV features) over Biostrap (incomplete case capture due to technical limitations in this patient population), though Biostrap remains essential for SpO₂- based respiratory monitoring.