New Research: Circadian Model Initialization with Wearables
I’m pleased to share our latest publication in the Journal of Biological Rhythms, titled “Circadian phase estimation from ambulatory wearables with particle filtering: accuracy depends on initialization, recording duration, and light exposure.” This work tackles a practical challenge in translating circadian models from the laboratory to real-world wearable sensors data: how do you get reliable estimates circadian phase when you don’t know the starting phase?
Circadian models typically require initial conditions to estimate the timings of phase markers like dim-light melatonin onset (DLMO), but in real-world settings, the true starting phase is unknown, especially in people with irregular schedules and light exposure patterns. In this study, we examined how uncertainty propagates through model estimates over time. We find that for individuals with a stable regular light exposure pattern, the transient period following initialization impacts accuracy for ~14 days. For shift workers with less regular schedules, this period is longer.
We also quantified how many initialization particles are needed for reliable maximum likelihood estimates and explored how light exposure pattern illuminance and temporal structure influence performance. Understanding these behaviors in real-world datasets helps to clarify the limitations of current circadian models, enabling their more appropriate and reliable use with wearable sensors. This sets the foundation for improving future application in both research and clinical settings.
I’d like to thank my PI, Jamie Zeitzer, as well as my collaborators Arec Jamgochian, Melissa St. Hilaire, Philip Cheng, and Mykel Kochenderfer for their support and thoughtful feedback throughout this project. I’m also grateful to the NIH NHLBI for supporting me through a NRSA F31 predoctoral fellowship, enabling this work. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.
