Predicting hospice eligibility among dementia patients using language models

Abstract

Introduction: Alzheimer's disease and related dementias (ADRD) are increasing in prevalence, and access to potential benefits of hospice care remains challenging. Large language models (LLMs), like GPT-4o, applied to electronic health records (EHRs) could support decisions by estimating mortality risk.

Methods: We analyzed patients with ADRD diagnosis from two academic medical centers. GPT-4o was used to estimate 6-month mortality risk from discharge summaries without any retraining or preprocessing. We used Cox regression to assess associations between predictions and time to death.

Results: Of 9872 individuals, 3563 (36%) died within 6 months. GPT-4o predictions stratified risk of death within 6 months (log-rank p < 0.001, area under the curve [AUC] = 0.79); predictions were strongly associated with mortality in Cox regression models (adjusted hazard ratio [aHR] = 31.02 95% confidence interval [CI] 27.44-35.08, p < 0.001) with similar results between sites.

Discussion: GPT-4o can stratify mortality risk using routinely generated documentation, potentially facilitating hospice referral decisions, but more prospective work is needed.

Highlights: Large language models (LLMs) can estimate 6-month mortality in patients with dementia. GPT-4o estimates of mortality risk from discharge summaries were highly discriminative (area under the curve [AUC] = 0.79). Predictions may support hospice referral decisions.

Keywords: Alzheimer's disease; clinical decision support; dementia care; electronic health records (EHRs); end‐of‐life care; hospice eligibility; large language models (LLMs); mortality prediction; natural language processing (NLP); prediction modeling.