Predicting Persistent Creatinine Elevation After the First Recorded ICU Furosemide Event in Acute Heart Failure: A Retrospective Cohort Study

Abstract Background During decongestion for acute heart failure, an early creatinine rise may resolve or remain elevated. At an intensive care unit (ICU) reassessment, clinicians cannot know from one value whether a first rise will later persist. Existing furosemide-timed models address death or organ support, and broader heart-failure models address acute kidney injury; neither directly answers this question. We examined later confirmation of the first observed rise at 12 and 24 hours after a recorded ICU furosemide event. Methods We studied a single-center retrospective Medical Information Mart for Intensive Care IV (MIMIC-IV) cohort. Time zero was the first positive-dose furosemide event in ICU-linked records, not verified initiation of hospital diuretic treatment. Separate ridge logistic models used baseline creatinine, creatinine change, cumulative urine output, and mean arterial pressure. Internal validation assessed discrimination, calibration, and prediction error; additional analyses examined initial-rise status, a simple creatinine rule, and the interval from ICU admission to the recorded event. Results Of 784 patients, 481 (61.4%) had a classifiable creatinine course. At 12 and 24 hours, the classifiable risk sets comprised 433 patients/63 events and 383/57; 38.1% and 57.9% of event patients had already shown a first rise. In complete cases, optimism-corrected areas under the receiver operating characteristic curve were 0.783 (bootstrap percentile range 0.721–0.845; 368/52) and 0.840 (0.791–0.900; 357/51). Held-out areas after within-fold predictor imputation were 0.790 and 0.846. The simple rule based only on a rise already recorded had areas of 0.650 and 0.728. Adding cumulative furosemide dose did not consistently improve prediction. Conclusions Among patients with enough creatinine follow-up to classify the outcome, routine measurements separated higher from lower risk of later-confirmed persistent worsening renal function. The models may help frame future studies of scheduled renal reassessment, but their performance has not been established for patients without classifiable outcomes or in another hospital.