Human-Led Causal Modeling in Mental Health

Psychiatric diagnosis represents a complex intersection of clinical judgment, social factors, and institutional practices. While artificial intelligence and algorithmic approaches increasingly influence healthcare decision-making, their application to psychiatric diagnosis often remains limited by an inability to capture the social complexity of diagnostic processes. This paper presents an innovative methodological framework for understanding how Borderline Personality Disorder (BPD) diagnoses are produced, demonstrating how human-led Structural Causal Modeling (SCM) can illuminate the social, institutional, and professional factors that shape diagnostic practices. Drawing on critical realism and Bayesian epistemology, we argue that purely data-driven algorithmic approaches to understanding psychiatric diagnosis remain trapped at the level of correlation, unable to ascend to true causal understanding. We propose instead a socio-technical framework that positions SCMs as tools for integrating expert knowledge, empirical evidence, and social theory. This approach begins with human-generated Directed Acyclic Graphs (DAGs) that represent theoretical understanding of causal relationships, enhanced by structural equations and computational validation. Unlike conventional algorithmic approaches that rely solely on pattern recognition in historical data, our framework explicitly incorporates professional expertise and sociological understanding of how diagnostic categories emerge and are applied. Using BPD as a case study, we demonstrate how this methodology can reveal the complex interplay between social structures, professional practices, and diagnostic decisions. BPD's controversial history and documented diagnostic disparities across demographic groups make it particularly suitable for examining how social factors influence psychiatric diagnosis. Our approach shows how SCMs can function not only as analytical tools but also as means of communication, making visible the often-hidden social processes that influence diagnostic practices. This work makes three key contributions to sociological understanding of algorithmic approaches in healthcare. First, it demonstrates how human-led causal modelling can reveal social influences on diagnostic practices that purely computational approaches might miss. Second, it provides a methodological framework for integrating sociological theory with advanced computational methods, showing how algorithms can serve as tools for social analysis rather than replacing human judgment. Third, it offers practical insights into how better understanding of causal mechanisms in diagnostic practices could help address systemic biases and inequities in psychiatric diagnosis. The implications extend beyond psychiatric diagnosis to broader questions about the role of algorithms in healthcare. We argue that effective integration of algorithmic tools in clinical practice requires frameworks that can account for the social construction of medical knowledge and the influence of institutional contexts. Our approach demonstrates how sociologically-informed causal modelling can help bridge the gap between computational methods and social understanding, potentially informing more equitable and effective diagnostic practices. This research contributes to ongoing discussions about how sociologists can engage with algorithmic technologies in ways that promote social justice and improved healthcare outcomes. By developing methods that make visible the social production of psychiatric diagnosis, we provide tools for both understanding and potentially intervening in processes that perpetuate diagnostic disparities. This work suggests new directions for how sociological perspectives can inform the development and implementation of algorithmic approaches in healthcare while maintaining critical awareness of their limitations and social implications.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.