Mining social determinants of health documentation patterns for genitourinary cancer patients using natural language processing

Mounting evidence demonstrates that social determinants of health (SDOH) significantly impact cancer care, contributing to persistent disparities in prevention, diagnosis, treatment access and survival rates that disproportionately affect vulnerable populations. However, SDOH are frequently underreported in clinical notes, obscuring barriers for patients to receive appropriate care. The systematic extraction from clinical notes remains challenging. We developed an advanced Natural Language Processing (NLP) based Artificial Intelligence (AI) platform to automatically extract SDOH from genitourinary cancer clinical notes. We analyzed 757,757 clinical notes from 5,585 patients (prostate:n=3,772; bladder:n=619; renal:n=1,194, based on ICD) using the AI platform. The platform used NLP to extract 19 types of SDOH features (e.g. race, ethnicity, health literacy, substance abuse, financial strain, social isolation, etc) across four note types: Progress Notes, Consults, Care Plans, Patient Instructions. The AI platform extracted 767,000 SDOH mentions from 4,464 patients (80% of patients). 140 clinical notes across 21 patients were randomly selected to evaluate AI accuracy. Of 140 notes, AI found SDOH features in 131 notes and clinicians from University of Illinois Cancer Center manually determined that 124 notes contained fully accurate SDOH extractions, while 7 notes had partially inaccurate SDOH extractions, yielding 95% accuracy (124/131). The study showed high prevalence of mentions of health literacy (96.71%), substance use (93.15%), and mood/affect issues (81.41%). 65% patients had minimal SDOH documentation (1.4 features per patient) and 35% patients had multiple SDOH documentation (6.9 features per patient). Among 2,900 patients who were diagnosed with genitourinary cancer in 2018-2023 (317,143 SDOH mentions after removing boilerplate SDOH mentions), by analyzing COVID-19 impact (pre vs post March 2020), we found that documentation of demographic SDOH decreased, e.g. home address (-37%), race (-34%), ethnicity (-36%), but the social SDOH increased, e.g. social isolation (+65%), financial strain (+76%), stress (+63%). This study shows the feasibility and effectiveness of using NLP to systematically extract SDOH from genitourinary cancer clinical notes. The high prevalence of documented SDOH underscores their potential impact on cancer care delivery and emphasizes the importance of routine SDOH documentation for every patient. COVID-19 caused a significant shift in SDOH documentation patterns with more pandemic-exacerbated social barriers. Future work should focus on validating the clinical impact of SDOH-informed decision making. Additional directions include leveraging large language models (LLMs) to infer the presence or impact of SDOH features and analyzing extracted SDOH information for clinical and societal insights.