Large Language Models for Accurate Mental Health Screening: Identifying Clinical Phenotypes in ED Healthcare Workers

Abstract The emergency department (ED) is crucial to the healthcare system. ED healthcare workers (HCWs) play a vital role in providing essential healthcare under demanding conditions. Traditional mental health assessment tools often rely on clinical interviews and psychometric assessments, which can be time-consuming, costly, and subject to biases. We aimed to identify heterogenous clinical profiles using a person-centered clustering approach based on burnout, depression, and PTSD symptomatology. We additionally investigated whether these phenotypes could be predicted by narratives about work-related future expectations using a Large Language Model (LLM). Based on n = 199 ED HCWs from an ongoing NIH-funded study (R01HL156134), k-means clustering revealed a High- and Low-Symptom Phenotype, significantly differentiating severity levels. Zero-shot LLM prompt engineering accurately predicted those clinical phenotypes from work-related narratives (accuracy = 70.9%; F1-score = 71.8%; sensitivity = 77.1%) based on key domain-specific indicators identified in the LLM’s reasoning. Our approach leverages LLMs based on unstructured narratives, offering an objective, time-efficient alternative that enhances early risk stratification and fosters a stigma-free environment for mental health assessment in high-stress healthcare settings, revealing subtle but meaningful variations in symptomatology. Future research should incorporate indicators such as risk factors and symptom dynamics to refine this tool for scalable, person-centered mental health monitoring across diverse healthcare settings.