0005 Validation and Comparison of Mathematical Models to Predict Circadian Phase Among Preschoolers

Abstract Introduction Having a later circadian phase is linked to risks for childhood obesity, which often persists into adulthood and is strongly related to metabolic disease risk-associated conditions, now increasingly observed in early childhood. However, accurately assessing circadian phase in young children is challenging. Mathematical models within applied mathematics provide a promising approach. This study aimed to validate and compare two models, the physics-based Forger and physiologically informed Hannay model, to advance circadian phase predictions using activity inputs in preschoolers. Methods As part of an ongoing randomized controlled trial, 69 parents of 4-year-old preschoolers (55.6% male, Latino 37.5%) completed 7-day wrist-worn actigraphy before a lab visit, during which dim light melatonin onset (DLMO) was assessed. Salivary melatonin was sampled hourly under dim-light conditions (< 5 lx). Actigraphy data were used to predict circadian phase. Model simulations were conducted using Runge-Kutta 45 numerical integration written in Python. Lin’s Concordance Correlation Coefficient (LinCCC) was used to assess agreement between model-predicted DLMO and salivary DLMO. The Mean Absolute Error (MAE) and percentage of participants with a predicted DLMO within 1hr of measured DLMO were calculated. The Bland-Altman method was used to evaluate bias between two methods and to estimate an agreement interval within which 95% of the differences fall. A sensitivity test was also conducted, excluding participants with extended non-wear periods in their actigraphy data. Results The physics-based Forger model and the physiologically informed Hannay model demonstrated similar performance, yielding LinCCC values of 0.47 and 0.41. Following the removal of participants with extended non-wear periods, LinCCC increased to 0.50 and 0.49, respectively. The Forger model achieved a mean absolute error of 44 minutes, with 70% predictions within 1 hour of the salivary DLMO. The Hannay model showed a mean absolute error of 50 minutes, with 71% of predictions within 1 hr. The Bland-Altman analysis showed biases of −0.23 ± 0.95 for the Forger model and −0.32 ± 1.02 for the Hannay model. Conclusion Mathematical models using data collected from wearable devices can be used to predict DLMO among 4-year-old preschoolers. Further research is needed to refine and adapt these adult-based models for children across different ages. Support (if any) P01HD109876