ENERGY-Matched Heart-Rate Sensing Schedules for Pre-Symptomatic Infection Detection on Low-Cost Wearables

Abstract

Most wearable infection-detection studies have assumed that more sensing is better, and focused on deviations in resting heart rate preceding self-reported symptoms; however, not much consideration has been given to when sensing is needed when there is a limited budget for the amount of energy the device can expend. This study asks: Is there a difference in pre-symptomatic warning performance when the timing of the heart-rate acquisition is considered in addition to the amount of sensing? Two publicly released data sets (Stanford Phase 1 and Stanford Phase 2) were re-analyzed using a preregistered causal pipeline. The final frozen base cohort included 38 participants; 30 contributed to the primary paired comparison after arm-specific calibration requirements. On the primary, modelled-energy budget (E3), sparse schedules were employed with 7 acquisition bursts per day of 10 minutes each. The confirmatory comparison tested nocturnally sensing (S3) against uniform block (S2) through the personalized time-of-day-normalized resting-heart-rate detector, with an identical alert burden (no more than two alert-days/person-month) and with a hierarchical paired endpoint of detection and warning time. S3 did not outperform S2: W/L/T = 8/12/10, θ = 0.433 (95% BCa CI 0.300-0.583), exact two-sided p = 0.5034. Detection rate was 63.3% for S3 and 66.7% for S2 and median warning time was 13 and 14 days respectively for the detected cases. Due to H1 failure, the fixed-sequence gate was closed and the remaining schedule comparisons were estimation-only. For all sensitivity analyses reported in H1, θ was kept at or below 0.5. These results do not demonstrate equality, but they do not offer any evidence that heart-rate acquisition focused at night is superior to an energy-matched uniform heart-rate acquisition schedule in warning about infection prior to symptoms onset in this sample. The study presents a framework for evaluation of sensing placement, alert burden, missingness and modeled device energy that is reproducible and suggests the need for larger prospective validation and measured hardware-energy experiments.