Using large language models for extracting stressful life events to assess their impact on preventive colon cancer...

Increase in early onset colorectal cancer makes adherence to screening a significant public health concern, with various social determinants playing a crucial role in its incidence, diagnosis, treatment, and outcomes. Stressful life events, such as divorce, marriage, or sudden loss of job, have a unique position among the social determinants of health.

We applied a large language model (LLM) to social history sections of clinical notes in the health records database of the Medical University of South Carolina to extract recent stressful life events and assess their impact on colorectal cancer screening adherence. We used pattern-matching regular expressions to detect a possible signal in social histories and ran LLM four times on each social history to achieve self-consistency and then used logistic regression to estimate the impact of life events on the probability of having a code in health records related to colorectal cancer screening.

The LLM detected 380 patients with one or more stressful life events and 5,344 patients with no life events. The events with the most negative impact on screening were arrest or incarceration (OR 0.26 95% CI 0.06-0.77), becoming homeless (OR 0.18 95% CI 0.01-0.92), separation from spouse or partner (OR 0.32 95% CI 0.05-1.18), getting married or starting to live with a partner (OR 0.60 95% CI 0.19-1.53). Death of somebody close to the patient (excluding their spouse) increased the chance of screening (OR 1.21 95% CI 0.71-2.05). Many of the observed effects did not reach statistical significance.

Our findings suggest that stressful life events might have a counterintuitive impact on screening, with some events, such as bereavement, were associated with increased screening. Future work should be focused on validating the research findings using data from other health institutions. In addition, expanding the list of stressful life events by including a validated scale of stressful life events for patients from historically marginalized groups is warranted.