Unique Time-Series Patterns of Behavioral and Psychological Factors in Late-Life Depression: A Computational Psychiatry Approach with Hidden Markov Models

• Published in: The American journal of geriatric psychiatry 2024-04, Vol.32 (4), p.S50-S51
• Main Authors: Faruque, Saurab, Mizuno, Akiko, Wang, Linghai, Wu, Minjie, Schweitzer, Noah, Stahl, Sarah, Aizenstein, Howard
• Format: Article
• Language: English
• ISSN: 1064-7481; 1545-7214
• DOI: 10.1016/j.jagp.2024.01.122

Traditional approaches for characterizing changes in psychopathological constructs (e.g., depression) often focus on elucidating how the individual construct changes within the time-series data, while statistically accounting for other related constructs (e.g., anxiety, loneliness) that may a have shared/common variance. While these methods are useful for exploring associations among the isolated signals of those constructs, these classical frameworks fall short in providing insights into the comprehensive system-level dynamics underlying changes of observable psychological/behavioral constructs. Hidden Markov Models (HMM) are a statistical model that enable us to describe the sequential relations among multiple observable constructs. The structure of “hidden state” of HMM, in particular, closely aligns well with our objective of investigating the unobservable brain operations (i.e., latent variables) of these changes. This alignment offers an alternative approach that is potentially a more insightful methodology for analyzing time-series data. By integrating with other computationally advanced techniques, this study aimed to illustrate how one psychological/behavioral construct leads to another over time among older adults during the first year of the COVID-19 pandemic. Our study also aims to investigate the difference in the sequential patterns between depressed and non-depressed older adults. We analyzed the data of 823 older adults (age over 55, mean = 68.2 ± 6.95, 66% women) from the COVID-19 Coping Study by the University of Michigan Institute for Social Research. The participants completed a monthly survey measuring the levels of depression [Center for Epidemiological Studies Depression Scale (CES-D-8)], anxiety (Beck Anxiety Inventory), loneliness (UCLA Loneliness Scale), and exercise (score of 0-6 in 30-min increments) for 12 months between April 2020 and June 2021. The participants who reported a score less than two on the CES-D-8 (at baseline) were defined as depressed (25% of the sample). First, we applied the k-means clustering algorithm (k=5) to all four measures (depression, anxiety, loneliness, and exercise; standardized via z-score) to define five “states/clusters”. In the HMM, transition probabilities between the “states” were computed using the Viterbi algorithm to identify the most probable sequence of transitions, referred to as “change patterns.” T-tests were conducted to compare how each of the five “change patterns” differed between