Modeling Individual Cyclic Variation in Human Behavior.

Area of Science:

• Computational biology
• Time series analysis
• Machine learning

Background:

• Human health and behavior are influenced by various cycles like circadian rhythms and the menstrual cycle.
• Modeling these cycles in time series data is difficult due to unobserved and unlabeled nature of cycles.
• Multidimensional measurements over time require advanced methods for accurate cycle inference.

Purpose of the Study:

• To introduce Cyclic Hidden Markov Models (CyHMMs) for detecting and modeling cycles in multidimensional, heterogeneous time series data.
• To address challenges in real-world cycle modeling, including multivariate data, missing values, and inter-individual variations.
• To improve the accuracy of cycle length inference compared to existing methods.

Main Methods:

• Developed Cyclic Hidden Markov Models (CyHMMs) capable of handling multivariate data with discrete and continuous dimensions.
• Incorporated robustness to missing data and information sharing across individuals.
• Applied CyHMMs to synthetic and real-world health-tracking datasets.

Main Results:

• CyHMMs demonstrated significantly lower error rates in inferring cycle lengths (58% on simulated, 63% on real-world data) compared to baselines.
• The models successfully captured the progression of individual features/symptoms throughout the cycle.
• Identified distinct clusters of individuals with unique cycle patterns, revealing insights not explicitly provided to the model.

Conclusions:

• CyHMMs offer a powerful and accurate approach for modeling complex biological cycles from time series data.
• The method can uncover medically relevant subgroups, such as birth control users, by effectively handling missing data.
• CyHMMs provide deeper insights into individual health patterns and variations within populations.