Invalidity of light sensor data in field studies and a proposal of an algorithmic approach for detection and filtering of non-wear time
Wearable light sensors are increasingly used in intervention and population-based studies investigating the consequences of environmental light exposure on human physiology. An important step in such analyses is the reliable detection of non-wear time. We observed in light data that days with less wear-time also have lower variability in the light signal, and we sought to test if the standard deviation of the change between subsequent samples can detect this condition. In this study, we propose and validate an easy-to-implement algorithm designed to discriminate between days with a non-wear time >4h ('invalid days') vs. ≤4h ('valid days') and investigate to which extent values of commonly used physiologically meaningful light variables differ between invalid days, valid days, and algorithm-selected non-wear days. We used 83 days of light data from a field study with high participant compliance, complemented by 47 days of light data where free-living individuals were instructed not to wear the sensor for varying amounts of time. Light data were recorded every two minutes using the pendant-worn f.luxometer light sensor; validity was derived from daily logs where participants recorded all non wear time. The algorithm-derived score discriminated well between valid and invalid days (area under the curve (AUC): 0.77, 95% CI: 0.67-0.87). The best cut-off value (i.e., highest Youden index) correctly recognized valid days with a probability of 87% ('sensitivity'), and invalid days with a probability of 63% ('specificity'). Values of various light variables derived from algorithm selected days only (median: 264.3 (Q1: 153.6, Q3: 420.0) for 24h light intensity (in lux); 496.0 (404.0, 582.0) for time spent above 50-lux) gave values close to those derived from self reported valid days only. However, these values did not significantly differ when derived across all days compared to self-reported valid days. Our results suggest that our proposed algorithm discriminates well between valid and invalid days. However, in high compliance cohorts, distortions in aggregated light measures of individual-level environmental light recordings across days appear to be small, making the application of our algorithm optional, but not necessary.