인공지능 프로그램 기반 병원 중증환자관리 평가 지표 및 개선 근거 연구

Abstract

Background The Rapid Response System (RRS) aims to prevent critical incidents like cardiac arrest and unexpected death by enabling immediate medical interventions when patients deteriorate unexpectedly. Objective This study examines the impact of RRS implementation at our hospital, introduced in April 2019, on key metrics: cardiopulmonary resuscitation (CPR), unexpected ICU transfer, and unexpected death. It also assesses the effect of an AI solution, added in November 2022, on these outcomes. Data from January 1, 2012, to June 30, 2024, were analyzed using ARIMA, with metrics evaluated via AUROC, recall, and precision. Results After implementing RRS, a statistically significant reduction was observed only in CPR, with no significant changes in unexpected ICU transfers or deaths. The AI solution also showed a statistically significant reduction in CPR incidence but did not significantly impact the other metrics. AI performed worse than clinicians in predicting CPR and unexpected death (lower AUROC) but better for predicting unexpected ICU transfers. There were no major differences in prediction accuracy among models like MAES, NEWS, and MEWS. It is expected that AI-based RRS systems will improve over time. In some cases, the AI system detected critical issues that clinicians missed due to off-hours or lack of real-time monitoring. Discrepancies between automated and manual record systems highlighted limitations in evaluating real-time predictions. The proactive nature of monitoring studies introduces bias, as interventions can skew the comparison of predictive results. This study observed lower predictive performance compared to earlier validation studies, likely due to differences in the dataset and model performance over time. Conclusion Both the RRS and AI solution significantly reduced CPR incidences but did not impact other metrics. While the AI’s predictive performance did not meet expectations, biases in evaluation methods and data characteristics should be addressed in future studies. Additionally, optimizing RRS, particularly for off-hours, will help prevent critical incidents and improve patient outcomes.