A systematic review of computational modeling of interpersonal dynamics in psychopathology

Abstract

Interpersonal dynamics have long been acknowledged as critical for the development and treatment of mental health problems. While recent computational approaches have been argued to be uniquely suited for investigating such dynamics, no systematic assessment has been made to scrutinize this claim. Here we conduct a systematic review to assess the utility of computational modeling in the field of interpersonal psychopathology. Candidate studies (k = 4,208), including preprints and conference manuscripts, were derived from five databases (MEDLINE, Embase, PsycINFO, Web of Science and Google Scholar) up to May 2025. A total of 58 studies met inclusion criteria and were assessed in terms of the validity, performance and transparency of their computational modeling. Bayesian modeling was the most common approach (k = 18), followed by machine learning (k = 17), dynamical systems modeling (k = 13) and reinforcement learning (k = 10). These approaches revealed several interpersonal disruptions across various mental health conditions, including rigid social learning in mood conditions, hypo- versus hyper-mentalizing in autism versus psychotic conditions and polarized relational dynamics in personality conditions. Despite these insights, critical challenges persist, with few studies reporting comprehensive performance metrics (16%) or adopting open science practices (20%). We discuss these challenges and conclude with more optimistic messages by suggesting that when rigorously and transparently conducted, computational approaches have the potential to advance our understanding of psychopathology by highlighting the social underpinnings of both mental health and disorder. This systematic review provides insights into interpersonal dynamics of psychopathology when using various computational approaches and highlights the key challenges in the field of social computational psychiatry, including the need to standardize computational tasks, apply diverse computational models to the same datasets, report more comprehensive performance metrics and adopt open science practices to enhance transparency.