Section 1. The UCSF Health Problem 

Problem Statement: Across UCSF Health, nearly 59,000 hospital discharge summary (DCS) narratives are manually written each year. As one of the longest, yet most important forms of clinical documentation, the DCS places substantial documentation burden on inpatient providers. Unfortunately, as both the literature and inpatient provider experience have shown,¹ producing a high quality DCS in a timely manner is challenging. Whether providers are pressed for time, or are simply not aware of all of the details of a patient’s hospital encounter (the discharging physician at UCSF, is on average the last of 3 sequential physicians having cared for a patient, and therefore may not be aware of all events throughout the encounter), physician-written discharge summaries are not error free, as we recently demonstrated in a UCSF study accepted in JAMA Internal Medicine and is in pre-print.² (Figure 1) Poor quality discharge summaries may then have multiple downstream ripple effects impacting subsequent quality of care and patient safety. 

Background: The hospital DCS is an accounting of important hospital events and treatments that must be identified, synthesized, and composed by the discharging provider. A substantial contributor to documentation burden (often described as an epidemic³) is the DCS, a uniquely time intensive source of burden affecting not only all inpatient providers directly, but - depending on the quality of the summary – downstream providers too (e.g. primary care physicians, Skilled Nursing Facility physicians, and subsequent inpatient providers who commonly rely on reviewing the prior DCS when readmitting a patient).

The sequelae of this documentation burden include reduced face-to-face time on inpatient care, increased medical error rates, reduced document quality, physician burnout, and attrition. In one study, over 44% of hospital physicians reported not having sufficient time to compose high quality discharge summaries.¹ Burden that is unique to the DCS derives from the need to review notes, procedures, and events throughout the hospital encounter (the longer the encounter, the more difficult the task); synthesize; reconstruct; and manually compose a problem-by-problem narrative of important hospital events and treatments. The discharging physician is often in the position of trying to reconstruct events that occurred before his or her care for the patient. Furthermore, because the hospital environment is a busy setting, with the physician taking care of multiple patients and being paged on average every 15 minutes (internal study), composing a summary is done either while the physician is on service in a highly interruptive environment - or done after hours in “pajama time.” Ultimately, not only is documentation burden recognized by leading healthcare organizations as a critical systemic problem, but it has also been designated by UCSF Health as a high priority IT Portfolio Initiative for FY2025 here. Additionally, by increasing the efficiency of DCS narrative production, an important secondary benefit of this proposal is helping facilitate discharge readiness earlier in the day, thereby opening inpatient beds earlier and decompressing the number of patients boarding in the emergency department (also a UCSF priority). Therefore, this proposal aligns with several existing UCSF priorities.

Section 2. How might AI help? 

AI assistants are already being piloted in many health systems to reduce documentation burden through ambient scribing (USCF is currently piloting this technology), and AI-generated draft responses to inbox messages, noting statistically significant reductions in burden and burnout.⁴ Although AI’s potential for higher-stakes use cases such as medical decision making is still being explored, LLMs are well-known to excel in lower-stakes use cases such as medical summarization.⁵ We have already demonstrated safety and feasibility of using Versa for generating the DCS narrative by reading through all hospital encounter notes and comparing GPT-4 Turbo’s output to physician-generated summary narratives in a retrospective study that has been accepted for publication by JAMA Internal Medicine. According to processes outlined by the UCSF AI Governance Committee here, the next step following the retrospective analysis we did for the JAMA IM paper, would be a prospective pilot, which we are hereby proposing.

While the proposed pilot targets DCS for the Hospital Medicine service (as in our JAMA IM paper), the value of an LLM-drafted discharge summary has tremendous potential to scale across all inpatient specialty services at UCSF.

Section 3. How would an end-user find and use it?

An end-user would invoke the LLM-generated DCS narrative directly in the existing discharge summary workflow in APeX. (Figure 2A) As discussed with the APEx Enabled Research (AER) team on March 10, 2025, any inpatient provider writing a hospital DCS could optionally launch the LLM simply by following the usual workflow and clicking on “Discharge Summary” that would then be followed by the option to choose an LLM generated draft for provider review. Because Versa is HIPAA-compliant, UCSF is uniquely positioned to do this work and become a national leader in LLMs for the DCS.

Section 4. Embed a picture of what the AI tool might look like (Figure 2)

Figure 2A shows the existing APeX workflow for generating a discharge summary. The inpatient provider selects “Discharge Summary” from within the Discharge Navigator tab, opening a Discharge Summary template. Many of the items in the template (e.g. medication list, referrals, etc.) will auto-populate from APeX. However, the summary narrative (“History (with Chief Complaint)” and “Brief Hospital Course by Problem”) must currently be manually composed by the inpatient provider. This is the section that involves the most documentation burden. Figure 2B shows sample LLM output where the drafted discharge summary narrative would be placed and available for provider review. Furthermore, Epic has the ability to optionally include citations /links associated with LLM-generated statements to information sources from the hospital encounter, thereby facilitating review and verification by the provider.

Section 5. What are the risks of AI errors? 

The proposed use of AI for the DCS involves human-in-the-loop. Namely, inpatient providers must review LLM-drafted DCS much in the same manner as they must review drafted inbox replies to patients. It remains the inpatient provider’s responsibility to ensure the accuracy of the DCS. Nevertheless, in our retrospective study, not only did we find that providers reviewing the summaries (blinded as to whether the summary was LLM- or physician-generated) had equal preference for both (χ2 = 5.2, p=0.27), and that they had similar overall quality ratings (3.67 [SD 0.49] vs 3.77 [SD 0.57]; p=0.674), but that there was no difference in the harmfulness scores associated with each LLM vs. physician error. While accurate medical summarization is important, (we found that physicians also made errors of omission, inaccuracy, and even hallucinations), the stakes of AI use for medical summarization with human-in-the-loop are lower than for AI used in medical decision making. Finally, identifying errors in the LLM-generated content could optionally be accomplished by enabling a reporting button within APeX with which inpatient providers could report potential errors for further investigation and mitigation. We also hope to be one of the first use cases for UCSF’s Impact Monitoring Platform for AI in Clinical Care (IMPACC), which is being built for continuous, automated, longitudinal AI monitoring.

Section 6. How will we measure success? 

There are many potential approaches to deploying the intervention and metrics to measure success. In this high level overview, it is critical that the deployment approach minimizes the friction to existing clinical workflows that could limit adoption and user satisfaction. We therefore propose an observation cohort approach allowing each inpatient provider to have the option to choose an LLM-drafted summary for each patient. Although a randomized controlled trial (RCT) – either randomized at the provider level (certain providers opted in for all of their patients) or at the patient level (same provider could use the LLM drafts for some patients but not others) – would offer the most rigorous evidence, an RCT could cause substantial workflow friction.

Therefore, we will measure success in an observational cohort study on two main domains:

1. Measurements using data already being collected in APeX: Leveraging the APeX audit logs, an area of deep expertise in our research group, we will assess the burden on inpatient providers using the LLM-drafted content by: 1. Measuring the amount of manual editing of the LLM-drafted narrative the provider makes (character count and % of narrative text edited as proxies for burden), and 2. Time savings as measured by the amount of time spent either editing LLM-drafted narratives or composing provider-generated narratives.
2. Measurements not necessarily available in APeX, but ideal to have: In any new technology deployment, user satisfaction is important to capture. If possible, we will attempt to capture user satisfaction with the LLM-generated narrative process by collecting Net Promoter Scores from inpatient providers using the LLM-drafted narratives, as well as those who are downstream recipients (e.g. PCPs, SNF physicians). Also ideal, if possible, would be to capture in APeX, inpatient provider flags of potential errors (inaccuracies, omissions, hallucinations), not only as a measure of quality and safety, but for investigators to identify concerns, and develop mitigation strategies. Optionally, the investigator team can quantify these error types, or (in a more scalable approach) we could potentially embed a multi-agentic platform such as CrewAI (crewai.com) to allow multiple LLMs to check one another for errors.

Section 7. Describe your qualifications and commitment:

This project is led by Dr. Benjamin Rosner, MD, PhD, FAMIA. Dr. Rosner is a hospitalist, a clinical informaticist, an AI researcher within DoC-IT, and the Faculty Lead for AI in Medical Education at the School of Medicine. He has years of experience developing, testing, and deploying digital technologies into healthcare, and he was the lead author of the retrospective study that underpinned the evidence for this solution. Dr. Rosner has worked directly with AER through the Digital Diagnostics and Therapeutics Committee (DD&T) for 6 years.

Charu Raghu Subramanian, MD is a Clinical Informatics Fellow, a hospitalist, and a co-lead author on the retrospective research study that underpinned the evidence for this pilot. She holds certifications in Epic Clarity and as an Epic Builder.

Citations 

1.    Sorita A, Robelia PM, Kattel SB, et al. The ideal hospital discharge summary: A survey of U.S. physicians. J Patient Saf. September 6, 2017. doi:10.1097/PTS.0000000000000421

2.    Williams CYK, Subramanian CR, Ali SS, et al. Physician- and Large Language Model-Generated Hospital Discharge Summaries: A Blinded, Comparative Quality and Safety Study. medRxiv. September 30, 2024. doi:10.1101/2024.09.29.24314562 https://www.medrxiv.org/content/10.1101/2024.09.29.24314562v1 (also accepted for publication in JAMA-IM, with anticipated publication in May, 2025)

3.    Hobensack M, Levy DR, Cato K, et al. 25 × 5 Symposium to Reduce Documentation Burden: Report-out and Call for Action. Appl Clin Inform. 2022;13(2):439-446. doi:10.1055/s-0042-1746169

4.    Garcia P, Ma SP, Shah S, et al. Artificial Intelligence-Generated Draft Replies to Patient Inbox Messages. JAMA Netw Open. 2024;7(3):e243201. doi:10.1001/jamanetworkopen.2024.3201

5.    Van Veen D, Van Uden C, Blankemeier L, et al. Adapted large language models can outperform medical experts in clinical text summarization. Nat Med. 2024;30(4):1134-1142. doi:10.1038/s41591-024-02855-5

Summary of Open Improvement Edits

• Added suggested measurement concepts to allow inpatient providers to flag potential errors (inaccuracies, omissions, hallucinations)
• Added option for investigator team to quantify errors
• Added option for multi-agentic platform to identify and quantify errors
• Added Medrxiv link to manuscript in citation #2 (to be published in JAMA-IM in May, 2025)