Predicting hospice eligibility among dementia patients using language models

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Geriatric Medicine

Background:

• Alzheimer's disease and related dementias (ADRD) prevalence is rising, yet hospice care access remains a challenge.
• Electronic health records (EHRs) contain valuable data for clinical decision support.
• Large language models (LLMs) show potential for extracting insights from unstructured EHR data.

Purpose of the Study:

• To evaluate the ability of GPT-4o, a large language model, to estimate 6-month mortality risk in patients with ADRD.
• To determine if LLM-generated mortality risk predictions can aid in hospice referral decisions.

Main Methods:

• Analysis of 9872 patients diagnosed with ADRD from two academic medical centers.
• Utilized GPT-4o to estimate 6-month mortality risk directly from discharge summaries without retraining.
• Employed Cox regression to assess the association between GPT-4o predictions and actual mortality.

Main Results:

• GPT-4o predictions effectively stratified 6-month mortality risk (log-rank p < 0.001, AUC = 0.79).
• Predictions showed a strong association with mortality (adjusted hazard ratio = 31.02, p < 0.001).
• Performance was consistent across both participating medical centers.

Conclusions:

• GPT-4o can accurately stratify mortality risk in dementia patients using routine clinical documentation.
• These LLM-driven risk estimates show promise for facilitating timely hospice referral decisions.
• Further prospective research is warranted to validate these findings in clinical practice.