Computational approaches for individual circadian phase prediction in field settings

Knowledge of circadian phase is critical for timing interventions for circadian rhythm disorders, medications, or predicting alertness. Current gold-standard measures of circadian phase are impractical for continuous or real-time tracking. Mathematical modeling offers an alternative, whereby ambulatory monitoring of environmental, behavioral, and/or physiological variables can be used to predict circadian phase. This review examines available approaches for predicting circadian phase, ranging from statistical models to machine learning and dynamical systems models, and evaluates their readiness for individual phase predictions. Multiple models predicted circadian phase with similar accuracy when individuals were stably entrained. However, most models did not generalize, or were not tested, under more challenging conditions (e.g., circadian misalignment). One model performed similarly under a range of conditions: a limit-cycle oscillator model. Most models had been designed to predict circadian phase using group-level assumptions. Future work should focus on model individualization and improved wearables to capture more accurate ambulatory signals.