Towards personalized prevention of stress-related disorders with computational phenotyping

Author: Eliana Vassena
Affiliation(s): Radboud University / Donders Center for Cognitive Neuroimaging
Keywords: Computational phenotyping, stress-related disorders, motivation, computational modeling, reward and punishment sensitivity

Research question(s):

The goal of this study is to investigate whether motivational phenotyping (i.e., differences in sensitivity to reward, punishment, volatility, and effort measured in behavior during tasks involving motivation) predicts vulnerability to stress-related disorders. To this end, we will adopt a multi-step approach. First, we will compute standard behavioural task parameters reflecting motivation, to replicate usual findings for these tasks (see below for a more detailed explanation of the 4 tasks). Second, we will leverage an existing computational model of motivation (Silvetti et al., 2018), to create a digital twin for each participant, by estimating underlying motivation parameters based on behavior at the tasks (individual sensitivity to reward, punishment, volatility, and effort). Third, we will test whether the obtained motivational parameters predict individual well-being in real life, with outcome measures on positive affect, negative affect, resilience, and physiological arousal (measured with wearable devices). Fourth, we will examine whether motivational parameters also capture vulnerability at later time points during follow-up assessments. Additionally, we will compare prediction based on model-based motivational parameters with prediction based on typical mental health scales (Self-Report Inventory of Depressive Symptomatology (IDS-SR)  Rush et al., 1996, State and Trait Anxiety Inventory (STAI-S) Spielberger et. al., 1983). The goal is to determine whether computational phenotyping is more sensitive in detecting vulnerability to stress-related disorders as compared to mental health scales (currently widely used). Below a description of each step follows.

Standard behavioural analysis of 4 HBS motivation-related tasks. As a first step, we will conduct a set of replication analyses. We expect to replicate the behavioral findings that these tasks intend to capture. The outcome of these analyses will be provided to the HBS team, in case these measures are relevant for other researchers. To quantify sensitivity to reward, punishment, effort, and volatility, we leverage four tasks from the HBS dataset: the Foraging task, the Contextual Fear Generalization Task, the Columbia Card Task, and the Emotion Regulation Task.

• Foraging task (Constantino & Daw, 2015): This task investigates how individuals track reward rates and make decisions related to exploration and exploitation to optimize rewards while minimizing opportunity costs. We expect to replicate the 2015 findings: people tend to exploit more (rather than explore), as compared to what the optimal strategy would be (based on the marginal value theorem). Additionally, we expect participants to exploit more when the travel time is longer.
• Contextual Fear Generalization Task (Andreatta et al., 2015; Klumpers et al., 2010): This task measures fear generalization and the avoidance tendencies of individuals for feared versus safe stimuli. For our purposes, one specific measure is relevant for motivation, namely the Approach/Avoidance Task. Here, we expect that participants will choose to avoid viewing in the feared condition more often than in the safe condition, especially if the anticipated reward is low.
• Columbia Card Task (Figner et al., 2009; Figner & Weber, 2011): This task measures risk-taking in environments with varied levels of gain amount, loss amount, and loss probability. We expect to replicate the findings of Figner and colleagues (2009): we expect that individuals will turn fewer cards in the low reward gain trials, in the trials where the loss amount was high, and in the trials where the probability of encountering a loss card was high.
• Emotion regulation task (Kanske et al., 2011): We expect to replicate the findings of the study by Kanske and colleagues (2011):  subjective ratings of negative emotional stimuli should be assessed as less negative when the participants are asked to reappraise compared to passive viewing.

Digital twin creation: Next, we will adopt a computational phenotyping approach, using the Reinforcement Meta Learner (RML) developed by Silvetti and colleagues (2018). The RML is a neurobiologically-inspired computational system that simulates decision-making in situations of varying reward and punishment incentives, required effort, volatility, and temporal delay. The RML simulates behaviour, capturing choices that individuals make under certain task circumstances, and hence can be used to capture behaviour across different tasks. Using this model, we aim to derive the individual model-based parameters (sensitivity to reward, punishment, volatility, and effort) for each individual from their behavioral performance on behavioural tasks. The rationale for combining 4 tasks is based on the fact that each single task only captures a part of the relevant motivational processes. Using the obtained model-based parameters, we can then create a digital twin: a model that makes decisions in the way the participant does. This is an important methodological step, because this digital twin approach allows showing that the model can reproduce individual behaviour, and also that in the future it may be possible to use this digital twin for individualized simulations (for example to predict the effect of a specific intervention or treatment on a specific individual).

Motivational phenotyping to predict real-life well-being. The model-based parameters obtained in step B will be used to predict individuals’ well-being in real life. Here we test whether we can use the model-based parameters to detect individuals with lower positive affect, higher positive affect, lower resilience (computed as residual-based resilience score, Kalisch et al., 2017), and overall increased (Roos et al., 2021) or decreased (Tutunji et al., 2023) physiological arousal. The physiology analysis is more exploratory, as previous studies have reported both higher arousal and lower arousal under stress in different studies.

Subsequently, we will test the predictive value of the digital twins over time, by testing whether parameters at the 1^st assessment predict real-life well-being measures 4 months later (2nd assessment) and 8 months later (3rd assessment). Importantly, we will compare the predictive ability of the digital twin with normally used mental health scales (IDS-SR, STAI-S), to test whether the computational phenotyping approach has an added benefit.

Exploratory analyses: 1. Using machine learning methods, we will attempt to classify more vulnerable individuals as compared to less vulnerable individuals based on model parameters (in two separate attempts, one with real-life outcome measures as ground truth, and one with IDS and STI as ground truth). 2. With a clustering approach, we will attempt to identify sub-groups within the sample and qualitatively describe these different phenotypes. This will be very informative for future research, as this method, if successful, may be used for personalized selection of treatment (or preventative intervention in the future).

Risk assessment and contingency plan: The proposed computational phenotyping method is novel. Estimating motivational model parameters based on 4 different tasks may not be successful. Should this be the case, a risk mitigation strategy is in place. We will then estimate parameters for the tasks separately and apply.

Link: OSF preregistration

Abstract:

Stress-related disorders, including depression and anxiety, affect more than 500 million people worldwide (James et al., 2018). Despite increased research efforts, persistent challenges of heterogeneity within disorders and comorbidity between disorders result in high non-response rates in treatment, relapses, and the need for invasive interventions (Kalisch et al., 2017; MacQueen et al., 2017). This affects individuals’ quality of life and places a burden on their families and the healthcare system. This alarming scenario highlights a growing need to shift the focus away from treatment toward prevention strategies to improve individuals' well-being and avoid chronic outcomes.

To fill this gap, this project aims to develop an early vulnerability tracking method based on motivational phenotyping. Motivation is notably altered in stress-related disorders, and these impairments affect patients’ daily functioning. Individuals with depression are less sensitive to rewards, more sensitive to punishments, and generally reluctant to exert effort (Horne et al., 2021; Treadway et al., 2012). Anxious individuals are also more sensitive to punishments and tend to exert excessive efforts to avoid them (Hulsman et al., 2021). These alterations are present to various degrees in healthy individuals (Ang et al., 2017; Apps et al., 2015), and contribute to the development and maintenance of maladaptive behaviours. For example, reduced effort to obtain rewards leads to less exposure to positive events, potentially preventing recovery. We propose that capturing these alterations is the key to timely predict individual vulnerability and has the potential to indicate which underlying dysfunction could be targeted with preventive intervention.

Leveraging a computational phenotyping approach, we will quantify individual variations in sensitivity to reward, effort, punishment, and volatility, and based on these motivational indexes, predict individual resilience and well-being in real life (concurrently and longitudinally). This method holds great promise for capturing vulnerability to stress-related disorders, and to ultimately develop personalized prevention.