응급실 정신건강의학과 협진 환자의 재내원 관련 요인 및 예측: ChatGPT 기반 자연어 분석을 이용한 탐색적 연구

Factors and Predictors of Revisit Among Psychiatric Consultation Patients in the Emergency Department: An Exploratory Study Using ChatGPT-Based Natural Language Analysis

Objectives: This study aimed to identify the factors associated with psychiatric emergency de\-partment revisits within three months through OpenAI-based natural language analysis of medi\-cal records and develop a predictive model for revisits. Methods: A retrospective chart review was conducted by analyzing medical records of a general hospital over a 5-year period, from March 1, 2020 to March 1, 2025, using a GPT-4o mini API (ap\-plication programming interface). The main reasons for psychiatric consultation were categorized into the top 10 causes, which included the following: suicidality (ideation or attempt), depressive symptoms, anxiety or panic symptoms, substance or alcohol use problem, cognitive or mental fluc\-tuation, impulsiveness or aggression, sleep disturbance, psychiatric consultation for medical illness, acute psychosis symptoms, other or unspecified. The socioeconomic factor variables included eco\-nomic difficulties, low educational attainment, family conflict or stress, interpersonal problems (work/school), sexual assault or abuse, physical illness or chronic pain, and exposure to celebrity suicide. Results: The revisit group had a significantly lower mean age than the control group, and the pro\-portion of female patients was higher among those who revisited (adjusted residual [AR]=-2.00). Higher revisit rates were observed among patients with persistent mood disorder (AR=+11.18), acute stress or adjustment disorder (AR=+2.62), suicidality (AR=+6.94), sexual assault or abuse (AR=+6.80), and interpersonal problems (AR=+4.66). These were also presented as major predic\-tive variables in the Shapley Additive exPlanations analysis. Conclusion: The risk of revisits could be predicted using only the information from psychiatric emergency medical records. An automated prediction system based on the narrative data in the electronic medical records could be developed.