Predicting AKI in critical patients: An interpretable model based on albumin and fluid balance.

Background: Acute kidney injury (AKI) is a clinically complex syndrome with a high incidence and mortality rate in the intensive care unit (ICU). Early identification of high-risk patients and timely intervention are crucial.

Objective: A local database was used to construct a model that predicts the incidence of AKI in ICU patients within 48 hours.

Materials and methods: We conducted a study involving 9,628 critically ill patients at Zhejiang Provincial People's Hospital and divided the cohort into derivation and validation groups. We collected and analyzed demographic data, vital signs, laboratory tests, medications, clinical interventions, and other information for all patients, resulting in a total of 232 variables. Six different machine learning algorithms were employed to construct models, and the optimal model was selected and validated.

Results: A total of 2,441 patients were included, of whom 1,138 (46.62%) met the AKI criteria. A model was derived that included 16 variables such as albumin transfusion, fluid balance, diastolic blood pressure (DBP), partial pressure of oxygen (PO₂), blood glucose (GLU), platelet (PLT), baseline serum creatinine (bSCr), serum sodium, age, epinephrine, proton pump inhibitor (PPI), intra-abdominal infection, anemia, diabetes, glycerin fructose, and nutritional pathway. The area under the receiver operating characteristic curve (AUC) was 0.822. Subgroup analysis revealed the impact of blood pressure fluctuations on AKI. Additionally, the study demonstrated a bidirectional effect of albumin and fluid balance on AKI.

Conclusion: This model is highly accurate and may facilitate the early diagnosis of and interventions for AKI.