Know the Patient, Not Just the Disease: Modeling Human Mental States Through Graph-Based Relational Reasoning

Mental health is inherently relational, encompassing social factor interactions, longitudinal treatment dynamics, and interdependencies between symptoms. However, current AI systems fail to capture this complexity by treating patient data as independent features. This creates a fundamental mismatch between the relational nature of mental health and the capabilities of current AI systems. The result is models that miss critical interdependencies, such as how social isolation exacerbates depressive symptoms or how medication side effects interact with existing conditions, limiting their clinical utility and accuracy. Addressing this mismatch requires rethinking how AI systems represent and reason over clinical data. This position paper argues for a fundamental architectural shift toward graph-enhanced AI agents that combine relational reasoning with fast inference. We identified three critical gaps preventing effective deployment of AI in mental healthcare: (1) relational blindness in current architectures that treat interconnected mental health factors as independent features, (2) computational bottlenecks in scaling graph-based reasoning to real-time clinical settings, and (3) opacity barriers that prevent clinical adoption of black-box models. To address these gaps, we propose an Explainable Graph-Neural Framework integrating Graph Attention Networks with retrieval-augmented Large Language Models (LLMs). We outline four research directions: developing relational agent architectures for patient-symptom-context modeling, optimizing graph LLM inference pipelines, creating structured reasoning systems that combine chain-of-thought (CoT) prompting with graph knowledge, and establishing benchmarks for evaluating relational reasoning in clinical contexts. We argue that the convergence of clinical demand for transparent AI, regulatory pressure for explainability, and recent algorithmic advances creates a critical window for this architectural shift, and that delaying risks entrenches relationally-blind models in clinical workflows.