Leveraging Large Language Models to Identify Lung Cancer Patients with Unregistered World Trade Center Disaster Exposure

OBJECTIVE: Leveraging large language models (LLM), we identified an unregistered subpopulation of individuals with World Trade Center-related exposure and lung cancer who were not previously captured in registries, and assessed how this exposure impacted patients' disease course. METHODS: Associations between exposure type and smoking history, cancer stage, mutation status, disease progression, and survival were statistically analyzed. RESULTS: The highest proportion of never-smokers was observed among residents, compared to first responders and commuters (19% and 24%; p = 0.005). Residents had more than twice the risk of disease progression (HR = 2.14, p = 0.008) and an elevated risk of death (HR = 2.43, p = 0.03). Only EGFR mutations were significantly associated with exposure type (p = 0.01). CONCLUSIONS: This work highlights that LLM can capture a greater population of WTC survivors, including genetics.