Project
Personalized depression dynamics in emerging adults: towards a real-time alert system
New technological advancements have allowed for the collection of intensive longitudinal data (ILD), using, among others, experience sampling methods and passive sensing from smart devices. These data have great potential for giving insights into a person’s moment-to-moment dynamics.
One example that ILD can be helpful for, are the dynamics of individuals with depression. The level and expression of depression has been shown to vary from time to time within the same individual (e.g. Catarino et al., 2022; Liu et al., 2023), and ILD can be helpful in uncovering these dynamics. Furthermore, ILD may prove to be useful for real-time prediction of depressive symptoms, which would have substantial clinical use. Whereas depression is commonly assumed to be a continuous construct, recent research has indicated that depression may be well represented as discrete states, which vary within persons from time-to-time (Liu et al., 2023; Catarino et al., 2022). Moreover, depression is an inherently heterogeneous disorder; there are considerable between-persondifferences in its symptomatic expression (Fried & Nesse, 2015). This implies that a personalized approach is essential.
This project aims to develop and evaluate Hidden Markov Model based methodology for personalized real-time prediction of depressive symptoms. The Hidden Markov Model (HMM, e.g. Zucchini et al., 2017) is a statistical method that can be used to evaluate switches between unobserved (latent) states over time. Earlier research, for example, used the HMM on bipolar data to uncover latent bipolar, depressed, and mixed states (Mildiner-Moraga et al., 2024).
The HMM implicitly assumes that shorter state durations are more likely than longer durations. We argue that for many constructs, including depression, this is an unrealistic assumption. The Explicit Duration HMM (EDHMM) is an extension of the HMM that allows for explicitly modeling the duration for each state and may therefore better describe some phenomena. To further allow for a personalized approach, a Multilevel EDHMM (MEDHMM) is likely valuable. A MEDHMM could then be used to model the dynamics of discrete depressive states over time, while allowing for between-person heterogeneity in expressions of depression, as well as explicitly modeling the duration of specific depressive states. Subsequently, these uncovered dynamics can be used to forecast future depressive symptoms.
The PhD project will consist of the following sub-projects:
- Develop and validate the MEDHMM prediction algorithm through a simulation study
- Collect intensive longitudinal data pertaining to depression dynamics
- Using the MEDHMM, obtain personalized depression dynamics for the collected data
- Evaluate forecasting ability of the MEDMHMM on real depression data
Supervisors
Prof. Dr. Irene Klugkist
Dr. Emmeke Aarts
Dr. William Hale
Financed by
Utrecht University
Period
8 December 2024 – 2 December 2029