Bridging Social Determinants and Opioid Risk via Heterogeneous Graph Learning

Abstract

The opioid crisis continues to impose substantial public health burdens across the United States and beyond, necessitating novel computational frameworks capable of integrating multidimensional risk factors. Social determinants of health (SDOH), encompassing economic instability, housing insecurity, educational attainment, and social isolation, have emerged as critical yet underutilized predictors of opioid use disorder (OUD) onset and escalation. Conventional machine learning (ML) approaches largely treat these factors as independent variables, failing to capture the relational complexity embedded in real-world patient and community networks. This study proposes a heterogeneous graph learning (HGL) framework that encodes SDOH-related entities and their interactions through a multi-relational graph structure, enabling nuanced risk stratification at both individual and population levels. By integrating electronic health records (EHR), census-level socioeconomic data, and clinical outcome registries, the proposed model constructs heterogeneous graphs where nodes represent patients, communities, clinical events, and social variables, while edges encode diverse relationship types. Experiments conducted on a large-scale real-world dataset demonstrate that the HGL model outperforms state-of-the-art baselines by up to 14.3% in area under the receiver operating characteristic curve (AUROC), with notable improvements in identifying high-risk populations in socioeconomically disadvantaged areas. These findings underscore the potential of heterogeneous graph learning as a powerful tool for SDOH-informed opioid risk prediction, paving the way for targeted and equitable intervention strategies.