Search

BackgroundDischarge instructions are a key form of documentation and patient communication in the time of transition from the emergency department (ED) to home. Discharge instructions are time-consuming and often underprioritized, especially in the ED, leading to discharge delays and possibly impersonal patient instructions. Generative artificial intelligence and large language models (LLMs) offer promising methods of creating high-quality and personalized discharge instructions; however, there exists a gap in understanding patient perspectives of LLM-generated discharge instructions. ObjectiveWe aimed to assess the use of LLMs such as ChatGPT in synthesizing accurate and patient-accessible discharge instructions in the ED. MethodsWe synthesized 5 unique, fictional ED encounters to emulate real ED encounters that included a diverse set of clinician history, physical notes, and nursing notes. These were passed to GPT-4 in Azure OpenAI Service (Microsoft) to generate LLM-generated discharge instructions. Standard discharge instructions were also generated for each of the 5 unique ED encounters. All GPT-generated and standard discharge instructions were then formatted into standardized after-visit summary documents. These after-visit summaries containing either GPT-generated or standard discharge instructions were randomly and blindly administered to Amazon MTurk respondents representing patient populations through Amazon MTurk Survey Distribution. Discharge instructions were assessed based on metrics of interpretability of significance, understandability, and satisfaction. ResultsOur findings revealed that survey respondents’ perspectives regarding GPT-generated and standard discharge instructions were significantly (P=.01) more favorable toward GPT-generated return precautions, and all other sections were considered noninferior to standard discharge instructions. Of the 156 survey respondents, GPT-generated discharge instructions were assigned favorable ratings, “agree” and “strongly agree,” more frequently along the metric of interpretability of significance in discharge instruction subsections regarding diagnosis, procedures, treatment, post-ED medications or any changes to medications, and return precautions. Survey respondents found GPT-generated instructions to be more understandable when rating procedures, treatment, post-ED medications or medication changes, post-ED follow-up, and return precautions. Satisfaction with GPT-generated discharge instruction subsections was the most favorable in procedures, treatment, post-ED medications or medication changes, and return precautions. Wilcoxon rank-sum test of Likert responses revealed significant differences (P=.01) in the interpretability of significant return precautions in GPT-generated discharge instructions compared to standard discharge instructions but not for other evaluation metrics and discharge instruction subsections. ConclusionsThis study demonstrates the potential for LLMs such as ChatGPT to act as a method of augmenting current documentation workflows in the ED to reduce the documentation burden of physicians. The ability of LLMs to provide tailored instructions for patients by improving readability and making instructions more applicable to patients could improve upon the methods of communication that currently exist.

Abstract Objectives This study presents a design framework to enhance the accuracy by which large language models (LLMs), like ChatGPT can extract insights from clinical notes. We highlight this framework via prompt refinement for the automated determination of HEART (History, ECG, Age, Risk factors, Troponin risk algorithm) scores in chest pain evaluation. Methods We developed a pipeline for LLM prompt testing, employing stochastic repeat testing and quantifying response errors relative to physician assessment. We evaluated the pipeline for automated HEART score determination across a limited set of 24 synthetic clinical notes representing four simulated patients. To assess whether iterative prompt design could improve the LLMs’ ability to extract complex clinical concepts and apply rule‐based logic to translate them to HEART subscores, we monitored diagnostic performance during prompt iteration. Results Validation included three iterative rounds of prompt improvement for three HEART subscores with 25 repeat trials totaling 1200 queries each for GPT‐3.5 and GPT‐4. For both LLM models, from initial to final prompt design, there was a decrease in the rate of responses with erroneous, non‐numerical subscore answers. Accuracy of numerical responses for HEART subscores (discrete 0–2 point scale) improved for GPT‐4 from the initial to final prompt iteration, decreasing from a mean error of 0.16–0.10 (95% confidence interval: 0.07–0.14) points. Conclusion We established a framework for iterative prompt design in the clinical space. Although the results indicate potential for integrating LLMs in structured clinical note analysis, translation to real, large‐scale clinical data with appropriate data privacy safeguards is needed.

Abstract Background: Incarceration is a significant social determinant of health, contributing to high morbidity, mortality, and racialized health inequities. However, incarceration status is largely invisible to health services research due to inadequate clinical electronic health record (EHR) capture. This study aims to develop, train, and validate natural language processing (NLP) techniques to more effectively identify incarceration status in the EHR. Methods: The study population consisted of adult patients (≥ 18 y.o.) who presented to the emergency department between June 2013 and August 2021. The EHR database was filtered for notes for specific incarceration-related terms, and then a random selection of 1,000 notes was annotated for incarceration and further stratified into specific statuses of prior history, recent, and current incarceration. For NLP model development, 80% of the notes were used to train the Longformer-based and RoBERTa algorithms. The remaining 20% of the notes underwent analysis with GPT-4. Results: There were 849 unique patients across 989 visits in the 1000 annotated notes. Manual annotation revealed that 559 of 1000 notes (55.9%) contained evidence of incarceration history. ICD-10 code (sensitivity: 4.8%, specificity: 99.1%, F1-score: 0.09) demonstrated inferior performance to RoBERTa NLP (sensitivity: 78.6%, specificity: 73.3%, F1-score: 0.79), Longformer NLP (sensitivity: 94.6%, specificity: 87.5%, F1-score: 0.93), and GPT-4 (sensitivity: 100%, specificity: 61.1%, F1-score: 0.86). Conclusions: Our advanced NLP models demonstrate a high degree of accuracy in identifying incarceration status from clinical notes. Further research is needed to explore their scaled implementation in population health initiatives and assess their potential to mitigate health disparities through tailored system interventions.

BackgroundChat Generative Pre-trained Transformer (ChatGPT) is a 175-billion-parameter natural language processing model that can generate conversation-style responses to user input. ObjectiveThis study aimed to evaluate the performance of ChatGPT on questions within the scope of the United States Medical Licensing Examination Step 1 and Step 2 exams, as well as to analyze responses for user interpretability. MethodsWe used 2 sets of multiple-choice questions to evaluate ChatGPT’s performance, each with questions pertaining to Step 1 and Step 2. The first set was derived from AMBOSS, a commonly used question bank for medical students, which also provides statistics on question difficulty and the performance on an exam relative to the user base. The second set was the National Board of Medical Examiners (NBME) free 120 questions. ChatGPT’s performance was compared to 2 other large language models, GPT-3 and InstructGPT. The text output of each ChatGPT response was evaluated across 3 qualitative metrics: logical justification of the answer selected, presence of information internal to the question, and presence of information external to the question. ResultsOf the 4 data sets, AMBOSS-Step1, AMBOSS-Step2, NBME-Free-Step1, and NBME-Free-Step2, ChatGPT achieved accuracies of 44% (44/100), 42% (42/100), 64.4% (56/87), and 57.8% (59/102), respectively. ChatGPT outperformed InstructGPT by 8.15% on average across all data sets, and GPT-3 performed similarly to random chance. The model demonstrated a significant decrease in performance as question difficulty increased (P=.01) within the AMBOSS-Step1 data set. We found that logical justification for ChatGPT’s answer selection was present in 100% of outputs of the NBME data sets. Internal information to the question was present in 96.8% (183/189) of all questions. The presence of information external to the question was 44.5% and 27% lower for incorrect answers relative to correct answers on the NBME-Free-Step1 (P