LLM-in-the-Loop clinical query language execution on unstructured data in electronic health records.

Presentation Time: 02:45 PM - 03:00 PM

Abstract Keywords: Adaptive Clinical Decision Support, Artificial Intelligence/Machine Learning, EHR Implementation and Optimization
Primary Track: AI and Care Outcomes
Programmatic Theme: Clinical Decision Support and Analytics

Despite recent advances, large language models (LLMs) alone remain insufficient for Clinical Decision Support (CDS). Current Clinical Query Language (CQL) execution engines rely on structured data and cannot process unstructured clinical notes. We propose a method that integrates LLMs to support unstructured data within CQL. The code, prompts, LLM, and CQL execution engine are all open source and available for use.
Speaker(s):
Bellraj Eapen, MD, PhD.
University of Illinois at Springfield

Author(s):

Computable Phenotyping for Severe Sepsis. A Hybrid Approach using LLM.

Presentation Time: 03:00 PM - 03:15 PM

Abstract Keywords: Adaptive Clinical Decision Support, Data Science, Quality Measures and eCQMs / Quality Improvement, Quality Measures and eCQMs / Quality Improvement
Primary Track: AI and Care Outcomes
Programmatic Theme: Clinical Decision Support and Analytics

Timely and accurate recognition of sepsis in emergency departments and inpatient settings is crucial for improving patient outcomes, reduce mortality rates, and improve consistency and timeliness of documentation and reporting. Implementing an algorithmic model to consistently and accurately identify patients meeting the Centers for Medicare & Medicaid Services (CMS) Severe Sepsis and Septic Shock Early Management Bundle (SEP-1) measures, is of paramount importance in clinical practice. Such a model can significantly enhance early detection, streamline treatment protocols, and optimize resource allocation. In this presentation, we will delve into our experience utilizing large language models for real-time chart reviews to ascertain sepsis diagnosis based on clinical documentation. We will explore how these advanced natural language processing techniques can be seamlessly integrated into a computable phenotype algorithm that adheres to CMS criteria for sepsis. This innovative approach has the potential to revolutionize sepsis recognition, enabling healthcare providers to intervene promptly and effectively, ultimately saving lives and improving overall patient care.
Speaker(s):
Parsa Mirhaji, MD, PhD
Albert Einstein College of Medicine

Author(s):

Clinical Decision Support and Educational Interventions to Promote Editing of AI Generated Text

2025 Clinical Informatics Conference 25x5 Presentation

Presentation Time: 03:15 PM - 03:30 PM

Abstract Keywords: EHR Implementation and Optimization, Usability and Measuring User Experience, Patient Safety, SAFER guidelines
Primary Track: AI and Care Outcomes
Programmatic Theme: Usability, Efficiency, and User Experience

While large language models (LLMs) can be utilized to simplify medical jargon and complex language within discharge summaries, LLMs are prone to hallucinations and require clinician editing. Our usability studies with this model demonstrated more editing compared to within live clinical scenarios. We demonstrate the importance of several clinical decision support (CDS) systems and educational interventions to promote clinician editing in order to successfully implement such a model in a large healthcare system.
Speaker(s):
Priyanka Solanki, MD
NYU

Author(s):
Priyanka Solanki, MD - NYU; Jared Silberlust, MD MPH - NYU Langone Health; Christopher Sonne, MD - NYULMC; William Small, MD, MBA - NYU Langone Health; Nilufar Tursunova, MD - NYU Langone Health; Lucille Fenelon, MSN, MHA - New York University Health Center; Kellie Owens, PhD - University of Pennsylvania Perelman School of Medicine; Alyssa Gutjahr, BS - NYU Langone Health; Marisa Lewis, RHIA - NYU Langone Health; Jonah Zaretsky, MD - NYU Langone Hospital Brooklyn; Jonah Feldman, MD, FACP - NYU Langone Health;

GPT is not Alone. Physicians Hallucinate Too. A Blinded Comparative Quality and Safety Study of Physician- and Large Language Model-Generated Hospital Discharge Summaries

2025 Clinical Informatics Conference 25x5 Presentation

Presentation Time: 03:30 PM - 03:45 PM

Abstract Keywords: Artificial Intelligence/Machine Learning, Documentation Burden, Disruptive and Innovative Technologies, Patient Safety
Primary Track: AI and Care Outcomes
Programmatic Theme: Organizational Challenges

High quality hospital discharge (DC) summaries improve patient outcomes, but reconstructing the clinically important events, particularly for lengthy hospital encounters or ones in which there have been multiple attending physicians over time, is a time consuming process that contributes to documentation burden, and detracts from face-to-face care. Large language models (LLMs) offer the promise to summarize clinical notes from hospital encounters and generate draft summaries. We conducted a blinded comparative quality and safety study of physician- and LLM-generated DC summaries using 100 historical hospital encounters. The concatenated corpus of all clinical notes from each encounter (minus the existing DC summary) served as a reference against which: 1. The original physician DC summary, and 2. An LLM-generated summary (GPT-4 Turbo) were compared. 22 blinded physician reviewers evaluated the DC summaries for errors (inaccuracies, omissions, and hallucinations), potential for harm from errors, concision, comprehensiveness, coherence, and overall quality and preference. There were an average of 2.91 (SD 2.54) and 1.82 (SD 1.94) unique errors per LLM- and physician-generated summary respectively (p<0.001). LLM summaries had a greater number of inaccuracies and omissions than physician summaries, but a similar number of hallucinations. There was no difference in the potential for harm per error. LLM-generated summaries were more concise (p<0.001), more coherent (p=0.019), but less comprehensive (p<0.001) than physician-generated summaries. Overall, there was no difference between LLM- and physician-generated summaries in quality or in reviewer preference.Our findings highlight the potential of LLMs to draft hospital discharge summaries for clinician review and editing.
Speaker(s):
Benjamin Rosner, MD, PhD, FAMIA
UCSF

Author(s):
Christopher Williams, MB BChir - UCSF; Charumathi Raghu Subramanian, MD - UCSF;