The work performed during the outgoing phase at the UC Berkeley commenced by reviewing recent work in the field of clinical idiographic modeling. This process resulted in notable revisions to the project’s design. First, the review highlighted the importance of base rates to allow the generation of accurate, reliable, and useful prediction models. Hence, Study 1’s population was set to be clinically diagnosed socially anxious participants. Second, the review pointed out the difficulties in generating accurate models of emotional states. Hence, a specific effort to characterize compound emotional states was initialized. Third, the review drew attention to the ambiguous nature of psychological targets (e.g. avoidance), making them harder to predict and intervene upon. Hence, Study 2 has been redesigned to employ concurrent models that guide clinician-administered single-session interventions.
During Study 1's data collection, two smaller projects were launched. The first reviewed the impact of timing on affect dynamics research in a chapter titled ‘A Close Look at the Role of Time in Affect Dynamics Research’. The second analyzed ecological momentary assessment (EMA) data from an open trial, discovering that the differentiation of negative emotions predicts treatment outcomes. This finding, published in Frontiers in Psychology, illustrated the utility of pre-treatment EMA for identifying patient characteristics that forecast treatment success.
Additionally, as part of an effort to characterize emotional states, a new approach to model compound states has been developed – a Moment-Oriented Approach. It employs unsupervised clustering techniques to delineate archetypal patterns of momentary mixtures of emotions. The approach allows the identification of regularities that exist both within individuals and across the sample. Moreover, the approach reduces multiple emotional features to presence/absence scores providing both conceptual clarity and efficiency in subsequent data-analytic procedures. The moment-oriented approach has been implemented in multiple datasets, yielding promising results.
A new series of personalized single sessions was developed for Study 2, a randomized controlled trial targeting mild-to-moderate anxiety and depression, to test the effectiveness of customized sessions chosen through detailed data analysis. These sessions focus on psychosocial needs such as emotional stability and self-esteem, identified using a conditional entropy algorithm and EMA data. After completing therapist training and initial piloting, a small group of participants has successfully participated, affirming the method's viability.
In the project's incoming phase at KU Leuven, we recruited 70 high self-criticism participants for a 21-day EMA. Here again, we incorporated person-specific items and applied means ensure a balanced distribution of the target variable. We also included location features derived from participants' GPS. Using these data, we generated person-specific models to predict self-criticism. Using elastic net regression and k-fold cross-validation, we achieved a median AUC of 0.68 though model accuracy varied widely. Despite observing a moderate correlation between the number of items retained and model accuracy, the small sample size limited our ability to identify reliable predictors of accuracy.
Leveraging the person-specific models, we developed a JITAI focused on self-criticism, activating self-compassion exercise prompts based on prior data. Implemented in a micro-randomization design, this system either triggered exercises upon predicting self-criticism or did not, enabling an in-depth efficacy analysis. Despite a small control group for comparison and low real-time exercise usage by some participants, the intervention showed some effects. These results indicate the need for enhanced participant engagement and further data collection emphasizing adherence and intervention engagement.
The results of the project were presented at four scientific conferences, including SPR (twice), APS, and ICPS. Additionally, results were communicated via Twitter, and presented at KU Leuven seminars.
