Personalizing Loneliness Interventions Using Individual and Social Contexts: An Agent-Based Mental Health Continuum Approach

Loneliness is a growing public health concern with a wide-range of impacts on mental and physical well-being. While most interventions adopt a uniform, one-size-fits-all approach, emerging evidence suggests that the effectiveness of loneliness interventions depends on individual psychological traits and the structure of social environments. Previous work has shown that mental health outcomes are shaped not only by personal vulnerabilities but also by complex feedback processes embedded in social networks. However, it remains unclear which types of interventions work best for which individuals, and how social context modulates these effects. We develop and apply an agent-based model that simulates mental health on a continuum with loneliness as a latent state at its lower end. The model is shaped by baseline resilience, self-reinforcement, and social influence mechanisms, across a realistic synthetic population. We target these mechanisms for specific individuals and varying proportions of their social networks. Outcomes for the individual and sequential neighbors are compared to a counterfactual. We use a random forest model and compute SHAP values that identify which features of the individuals and their social networks predict intervention success.We show that interventions improving baseline resilience consistently benefit individuals, particularly those with low socioeconomic status embedded in moderately diverse networks. In contrast, interventions targeting self-reinforcement or peer influence exhibit greater variability and may produce adverse effects in overly homogeneous or fragmented contexts. Surprisingly, classical network centrality metrics fail to explain these patterns; instead, similarity in baseline and mental health traits across social layers emerges as a strong predictor of intervention efficacy.These findings suggest that personalization based on social and psychological similarity is more informative than topological metrics when designing interventions for loneliness. The results challenge assumptions about “scaling up” interventions and highlight the importance of tailoring to social context and individual vulnerability. More broadly, our approach demonstrates how computational models can capture the complex interplay between individual traits and cumulative, self-consistent social feedbacks, offering a concrete framework for designing socially aware, targeted public health interventions.