Abstract: Objective: To develop a predictive model for the timing of fluid removal in sepsis patients, aiming to provide a reference for clinical identification of the optimal fluid removal time. Methods: We retrospectively selected 17,186 observation windows of 532 ICU sepsis patients in a tertiary grade A hospital in Beijing from January 2019 to December 2023 to form the training set. Thirty-six predictor variables were included after screening. Four machine learning algorithms were employed to construct predictive models, and the optimal model for predicting fluid removal timing in ICU sepsis patients was identified. External validation of the model was conducted on a prospectively collected dataset of 2534 observation windows of 79 sepsis patients from January to December 2024. Results: The Random Forest model demonstrated the best overall performance, with an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.797 and a Brier score of 0.202. Meanwhile, in the external validation set, it achieved an AUC of 0.786 and a Brier score of 0.213. The top 10 variables ranked by importance: daily net balance, daily output, 24-hour urine output, 4-hour urine output, central venous pressure, 24-hour net balance, cumulative net balance, serum creatinine level, N-terminal pro-brain natriuretic peptide, and cumulative output. Conclusion: The Random Forest-based predictive model exhibits the best performance and can guide healthcare professionals in identifying the optimal timing for fluid removal.

Key words: ICU; sepsis; fluid removal; machine learning; predictive model