From Correlation to Causation: A Conceptual Framework for Applying Causal AI to Adolescent Depression Research.

Abstract:

Adolescent depression is a growing global concern, yet most neuroimaging and machine learning research in this domain remains correlational, limiting understanding and clinical translation. Existing studies predominantly use cross-sectional designs and black-box predictive models that identify associations but cannot determine whether observed brain features represent causes, consequences, or correlates depressive symptoms. This article proposes a conceptual framework for advancing from correlation to causation through the integration of longitudinal multimodal data and modern causal AI techniques. We outline key methodological limitations in current research including poor reproducibility, lack of temporal resolution, confounding development factors, and limited reproducibility; and describe how causal approaches such as Directed Acyclic Graphs, Granger causality, counterfactual reasoning, Mendelian Randomization, and causal graph neural networks can address these gaps. Building on these tools, we propose a three-phase framework consisting of: (a) Data foundations emphasizing longitudinal neuroimaging, symptom-level phenotyping, and multimodal integration; (2) Causal modeling pipelines using causal discovery, causal GNNs, and counterfactual simulations to identify mechanisms; and (3) translational pathways for Personalized interventions, and early-risk prediction grounded in causal pathways. Ethical considerations related to privacy, consent, fairness, and data governance are also examined, especially given adolescents’ vulnerability and the sensitivity of neural and digital phenotyping data. Collectively, this framework provides a systematic and feasible roadmap for leveraging causal AI to undercover mechanistic pathways, guide interventions, and strengthen the scientific foundation of adolescent depression research.