Brief
Duke provides high quality health care to its patients. It also provides support for its patients’ health-related social needs (HRSNs). Due to the high volume of patients with HRSNs and the chart review work required to support an intervention for a patient, it is challenging to provide comprehensive care for all Duke patients’ HRSNs. To address this challenge, we use artificial intelligence to automatically create a HRSN Discharge Summary Note for all hospitalized patients, available in the electronic health record on the day after they are discharged. Beginning in November 2025, our clinical teams have relied on this note to improve the speed and completeness of their preparation for a HRSN intervention on a patient.
Problem
Unmet health-related social needs (HRSNs) worsen health outcomes, lead to hospital readmissions, and increase healthcare costs. Yet, the path from identifying a need to connecting a patient with resources remained a complicated gauntlet with multiple failure points. Screening approaches to identify HRSNs were inconsistent, incomplete, and often ineffective at connecting patients to resources. Even when a need was recognized, successful referral rates to Community-Based Organizations (CBOs) through platforms like NCCARE360 typically landed in the 15-30% range. We identified three discrete points of failure along this path: recognizing a domain need, confirming the patient’s assent for help, and placing the referral in NCCARE360.
A core part of this problem was structural. Many social needs were documented only in free-text clinical notes within the electronic health record (EHR), making systematic identification labor-intensive and impractical at scale. Although Large Language Models (LLMs) had shown strong performance in extracting social needs and generating clinical summaries, their capability to synthesize HRSN information from longitudinal medical records had not been evaluated by the time this project received funding. There was a critical need to determine whether LLMs could accurately and efficiently summarize HRSN data to improve identification and intervention in clinical settings.
Identifying and connecting of patients with Health Related Social Needs (HRSN) to resources was a complicated gauntlet with multiple potential points for failure. There was a critical need to determine whether LLMs can accurately and efficiently summarize HRSN data to improve identification and intervention and ultimately close the gap between recognition and referral.
Solution
To address this gap, Duke University’s Department of Medicine and the Duke Population Health Management Office (PHMO), along with the Duke Institute for Health Innovation (DIHI), formed a transdisciplinary team to develop and implement the HRSN Discharge Summary Note Solution. At the time of a patient’s discharge from the hospital, our solution automatically generates a HRSN Discharge Summary Note within Maestro Care (Duke’s instance of Epic, the electronic health record) using relevant data from the patient’s hospitalization and past medical and social history. The tool integrates the ChatGPT-4.1 model via Microsoft Azure, a Large Language Model (LLM), to synthesize both free-text documentation from clinical notes and structured screening data captured in flowsheets by nurse teams. It applies a standardized discharge summary template and prompt to produce consistent outputs via the LLM. The evaluation of the LLM-generated output focused on hospitalized patients with diverse social needs profiles and varying admission complexity to ensure applicability across a broad range of cases. HRSN Discharge Notes were evaluated for quality on a 5-point Likert scale by physicians and clinical social workers across domains including:
- Completeness
- Accuracy
- Clarity
- Conciseness
- Credibility
- Overall usability
The findings were incorporated back into the prompt to improve the model’s output across all domains.
Outcomes
Through the evaluation of the HRSN Discharge Summary’s refined output, the solution’s HRSN Discharge Summary note scored well across all domains: completeness 4.75±0.49, accuracy 4.85±0.24, clarity/readability 5.00±0.00, handling of missing data 4.90±0.21, and overall satisfaction 4.85±0.24.
The HRSN Discharge Summary went live on November 1, 2025. From November 1, 2025 to March 1, 2026, HRSN Discharge Summary Notes were automatically generated for 13,491 hospitalizations across the DUHS hospitals. The Transitions Team utilizes the HRSN Discharge Summary Note in their preparation for follow up calls with patients who have identified HRSNs.
Impact
The impact of the HRSN Discharge Summary Note is improved care for more of Duke’s patient population in need. Specifically, this solution increases the clinical capacity and reduces the cognitive burden for the DukeWELL Care Transitions Team. Time savings for the Population Health Care Manager and others are being collected, with current estimates of 975 hours annually. This translates to a total cost avoidance of $60K (hourly pay) for FTE effort being redirected to patient care.
IP eCQM #487 requires reporting for social determinants of health (SDOH) screening rates and volume of patients with HRSNs. We anticipate an increase in the number of patients identified with HRSNs. We anticipate the sum of these interventions will support smoother transitions of care, decreases in readmissions, and improvements in health outcomes.
