While current mathematical models of human circadian rhythms accurately predict circadian phase responses to light in controlled laboratory experiments, they show reduced performance in the real world, especially among shift workers with irregular schedules and downstream erratic light diets. The source of the discrepancy between in-laboratory and ambulatory performance remains unclear. We evaluate the impact of initialization strategy, recording duration, and light exposure characteristics on model performance using wearable data from both individuals on regular schedules and shift workers. We implement a probabilistic initialization framework to account for unknown starting phase and assess model performance in prediction of phase from light input data against an in-lab measure of circadian phase (dim light melatonin onset). In participants with regular schedules, accuracy improved with longer recordings, while shift workers show no accuracy gains when having more nights of data. Light exposure patterns differed significantly between groups, with brighter and more regular day-to-day light exposure being weakly to moderately associated with improved model estimates, whereas fragmented patterns of light exposure increased uncertainty. These findings suggest that current models require adaptation, particularly in light sensitivity, to generalize to free-living, irregular conditions and support robust, scalable circadian tracking in real-world populations.
Weed et al. (Thu,) studied this question.