Overview
Serious illness conversations between patients with cancer and their care teams are essential for ensuring that medical care aligns with a patient’s personal goals and values. However, evaluating how often and when these discussions occur across large populations has historically been difficult because details are buried within unstructured clinical notes in electronic health records.
Researchers utilized a privacy-compliant artificial intelligence model, Claude 4.5 Sonnet, to analyze electronic health records from 1,994 patients with advanced non-small cell lung cancer in the Flatiron Health Research Database. The AI successfully identified that 91% of patient charts contained documentation of at least one conversation topic. While discussions about prognosis and palliative care typically occurred within a month of diagnosis, hospice was rarely documented until a median of one month before death. When patients chose to stop cancer-directed treatment, personal preferences—such as maintaining quality of life and spending time at home—were the leading reason cited.
Why this matters
This study demonstrates that advanced language models can extract nuanced, qualitative clinical details at scale to evaluate end-of-life care quality. The findings reveal that while clinicians regularly discuss diagnosis and prognosis early, hospice is often introduced late, offering an opportunity to improve communication frameworks so patients and families can make timely, goal-aligned decisions.