Automated Prompt Optimization Improves Same-Day Discharge Prediction in a Retrospective Inpatient Cohort Study

Automated prompt optimization improved the performance of a large language model for predicting same-day hospital discharge in a retrospective cohort study of adults admitted to a tertiary academic medical center. The study’s abstract reports that the optimized approach achieved higher F1 score and sensitivity than a basic prompt on a hold-out test set, although those gains came with lower positive predictive value and specificity.

Researchers used two independently randomized cohorts of hospitalized patients aged 18 years or older who were admitted between January 1 and December 31, 2024. Patients had hospital stays lasting between 2 and 14 days. The validation set included 860 patients, while the test set included 886. Each model prediction assessed whether a patient would be discharged the same day, using clinical documentation from the 30 hours preceding a 06:00 index time. A baseline GPT-5 prompt produced an F1 score of 0.48 and sensitivity of 0.37 in the validation set.

The researchers evaluated three inference-time optimization strategies: test-time scaling, expert-led prompt engineering, and automated prompt optimization. On the hold-out test set, automated prompt optimization had higher F1 score and sensitivity than the basic prompt, but also lower positive predictive value and specificity. The abstract does not provide numerical test-set performance values. Test-time scaling and expert-led prompt engineering produced minimal improvement. Overall, automated optimization brought the model’s discharge-prediction performance into the range reported in the abstract for prior approaches.

Qualitative analysis, conducted from June through December 2025, identified operational workflows as the most common source of model errors. The study authors concluded that these workflows, rather than medical knowledge alone, accounted for most errors. The available evidence is limited to the journal abstract and does not report the number of errors reviewed, detailed examples, or the model’s operating characteristics beyond those listed above. Because the study was retrospective and based at one tertiary academic medical center, the abstract does not establish whether the observed performance would generalize to other hospitals or patient populations.

Subject of Research: Evaluation of large language models and prompt optimization for same-day hospital discharge prediction in adult inpatients

Article Title: Optimizing Large Language Models for Hospital Discharge Prediction

Article References: Spagnoli, J., Guzman, N. M., Desai, K., Chang, T. H., Bell, A., Gilbert, V., Lehn, S., Parekh, V. I., Sjoding, M. W., & Wong, A. (2026). Optimizing Large Language Models for Hospital Discharge Prediction. Journal of General Internal Medicine. https://doi.org/10.1007/s11606-026-10792-x

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10792-x

Keywords: large language models, hospital discharge prediction, automated prompt optimization, GPT-5, retrospective cohort study, qualitative error analysis, inference-time optimization, electronic health records, Optimizing, Large, Language, Models