Sex-Specific Diagnostic Subtypes in Adolescents Hospitalized for Substance Use Disorders Revealed by Transformer-Based Clustering

Substance use disorders (SUD) are a leading cause of psychiatric hospitalization among adolescents, yet the underlying diagnostic profiles and comorbidities remain poorly characterized. Here, we applied a transformer-based language model to 4,849 hospital discharge records from adolescents (aged 11–18) admitted with mental health and SUD in Spain between 2016 and 2020. We generated dense clinical embeddings and identified sex-specific diagnostic clusters using unsupervised learning. Male subgroups were primarily defined by externalizing disorders such as conduct disorder, ADHD, and psychosis, often associated with cannabis and alcohol use. Female subgroups exhibited higher rates of borderline personality traits, mood instability, suicidal ideation, and stress-related conditions. A unified embedding space revealed a clear sex-based separation in diagnostic patterns. These findings highlight the potential of transformer-based embeddings to uncover latent clinical phenotypes and inform precision approaches to adolescent mental health care. These sex-specific subtypes could enhance early intervention strategies and support the development of tailored treatment programmes for youths with SUD.