Personalized Forecasting and Just-in-Time Interventions for Repetitive Negative Thinking: A Proof-of-Concept Study
This research project examines whether person-specific prediction models can improve the timing and effectiveness of just-in-time adaptive interventions for rumination. This proof-of-concept study integrates intensive ecological momentary assessment, idiographic machine-learning models, and a micro-randomized trial to test whether brief interventions are more effective when delivered at personally predicted moments of elevated rumination risk. Participants with elevated trait rumination completed two phases of ecological momentary assessment. Data from the first EMA phase were used to construct a separate forward-prediction model for each participant. The models used information available at one assessment, including lagged self-report and temporal features, to estimate the likelihood of rumination at the subsequent assessment. Following a one-week interval, participants completed a 14-day intervention phase. After each completed EMA prompt, the participant’s prediction model classified the subsequent assessment period as either a predicted rumination-risk or non-predicted moment. At each decision point, participants were micro-randomized to receive an intervention prompt or no prompt, with person-specific randomization probabilities that depended on prediction status and the participant’s Phase 1 rumination base rate. The primary objective is to determine whether prediction status moderates the proximal effect of intervention prompts. Specifically, the study tests whether receiving an intervention prompt, compared with receiving no prompt, produces greater reductions in subsequent rumination and down or depressed mood at predicted rumination-risk moments than at non-predicted moments. Binary rumination and distress outcomes are analyzed using estimating-equation methods (EMEE), whereas depressed mood is analyzed using weighted and centered least squares (WCLS). The study also examines broader changes in psychological functioning across the study period. Primary distal outcomes are rumination (RRS) and symptoms of depression, anxiety, and stress (DASS-21). These measures were assessed at baseline, after the first EMA phase and before the intervention period, immediately after the intervention period, and one month later. Linear mixed-effects models are used to evaluate changes over time, with the primary comparison assessing change from the pre-intervention assessment to the immediate post-intervention assessment. The expected outcome is that intervention prompts will have stronger immediate effects when delivered at moments identified by the personalized models as carrying elevated rumination risk. Such a finding would suggest that idiographic prediction provides clinically meaningful information beyond simply determining whether an intervention is delivered. The study also expects reductions in rumination and broader psychopathology symptoms following the intervention period, with improvements maintained at the one-month follow-up. Because the trial is a proof-of-concept study, the prediction-by-intervention interaction is interpreted as an initial test of personalized intervention timing rather than a definitive efficacy test. The project is intended to provide evidence regarding the feasibility of integrating individualized forecasting with real-time digital intervention and to inform the development of larger, more highly powered trials of personalized just-in-time interventions for repetitive negative thinking.