UCSF

Social determinants of health (SDOH)—including housing instability, food insecurity, and transportation barriers—significantly impact health outcomes, especially among vulnerable pediatric and opioid-affected populations. Despite their importance, SDOH remain under-documented in structured fields of electronic health records (EHRs), limiting effective clinical and policy responses. This dissertation explores the utility of natural language processing (NLP), particularly transformer-based models, for improving the identification and analysis of SDOH in unstructured clinical notes across three distinct studies.In Chapter 1, a comparative analysis evaluates the performance of RoBERTa, ClinicalBERT, and BioBERT in detecting SDOH mentions from pediatric inpatient notes. RoBERTa outperformed other models, achieving an F1-score of 0.83 and recall of 0.99, highlighting its potential for high-sensitivity detection of social risks in pediatric settings. ClinicalBERT and BioBERT underperformed, particularly in handling implicit or nuanced mentions, underscoring the challenges of applying domain-specific models to complex real-world data. Chapter 2 examines the association between housing and transportation-related SDOH and pediatric hospital readmissions. This study compares structured ICD-10-CM Z-code data with unstructured clinical text analyzed via a validated NLP pipeline. Incorporating unstructured data increased the identified prevalence of SDOH from 0.8% to 31.7%, significantly strengthening associations between social risk exposure and both readmission and emergency department utilization. These findings demonstrate the added value of NLP in revealing latent risk factors and improving predictive modeling in pediatric care. Chapter 3 applies zero-shot inference using GPT-4 and LLaMA-3.2 models to classify SDOH in patients with opioid-related disorders (ORDs). While both models excelled at identifying explicit social needs such as housing and transportation, they struggled with mental health categories like anxiety and depression, often misclassifying overlapping symptoms. Prompt engineering and model refinement improved classification performance but revealed persistent challenges in discerning complex psychosocial conditions from clinical narratives. Together, these studies underscore the promise and limitations of transformer-based NLP models in healthcare. While models like RoBERTa and GPT-4 effectively detect explicit SDOH indicators, enhancing specificity and contextual inference remains a critical frontier. By integrating structured and unstructured data, this dissertation highlights how advanced NLP tools can inform risk stratification, guide interventions, and support health equity efforts across diverse patient populations.