Personalized Data-Driven State Models of the Circadian Dynamics in a Biometric Signal.

Area of Science:

• * Computational biology
• * Chronobiology
• * Machine learning

Background:

• * Circadian rhythms are crucial for well-being but difficult to track individually.
• * Existing population-level models lack accuracy for personalized dynamics.
• * Understanding individual circadian variations is key for health applications.

Purpose of the Study:

• * To develop a personalized latent state model for circadian dynamics.
• * To accurately relate circadian inputs (e.g., light) to biometric signals.
• * To validate the model's ability to capture individual circadian oscillations.

Main Methods:

• * Combined autoencoder and recurrent neural network for latent state modeling.
• * Utilized light as circadian input and actigraphy for biometric signals.
• * Validated model performance against melatonin measurements for circadian phase.

Main Results:

• * Developed low-dimensional latent state models that accurately reconstruct and predict biometric signals.
• * Demonstrated that learned latent state oscillations correlate with individual circadian rhythms.
• * Showcased the model's effectiveness in capturing personalized circadian dynamics.

Conclusions:

• * Proposed a novel technique for personalized circadian system modeling.
• * The method offers potential for individualized feedback control and health studies.
• * Accurate individual modeling of circadian rhythms is achievable with advanced machine learning.