Effectiveness of ChatGPT-4 in predicting the human decision to send patients to the postoperative intensive care unit: a prospective multicentric study.

Background: Postoperative ICU admission is crucial in surgical patient management, impacting morbidity and mortality outcomes. Traditional prediction methods, such as the ASA physical status classification, are subjective and prone to variability. This study evaluates the effectiveness of ChatGPT-4 in predicting postoperative ICU admission needs using comprehensive preoperative patient data.

Methods: In this prospective multicentric study, data from 406 patients aged 18 and older were analyzed. Patients requiring emergency surgery and those with insufficient information were excluded. Preoperative data, including demographics, medical history, laboratory results, and imaging findings, were collected and anonymized. ChatGPT-4 was configured to predict ICU admission needs based on this data. The model's predictions were compared with actual ICU admissions using Chi-Square, confusion matrix, and one-sample t-tests.

Results: ChatGPT-4 model predicted 128 patients for ward care and 278 for ICU admission. Among the predicted ICU cases, 160 were correctly identified as ICU, while 118 were observed to need ward care. The overall accuracy of the model was 0.645, with a specificity of 0.464, sensitivity of 0.860, and an F1 score of 0.690. A chi-square test revealed a significant result (P=0.000).

Conclusions: The findings demonstrate that ChatGPT-4 can effectively predict postoperative ICU needs, providing a valuable tool for clinical decision-making. While the model showed strong agreement with actual ICU admissions, refinement is needed to improve the accuracy of ICU stay duration predictions. Integrating AI in preoperative assessments could enhance objectivity and efficiency, contributing to optimized patient care and resource allocation. Further validation across diverse patient populations is recommended.