Transformer-Based Daily Creatinine Prediction Model in Critically Ill Patients with AKI

Background: Acute Kidney Injury (AKI) affects over half of all intensive care unit (ICU) patients, with approximately 42% experiencing persistent AKI and more than 60% failing to recover. Understanding the trajectory of serum creatinine is critical for assessing kidney function and delivering personalized care. This study aims to develop and evaluate a transformer-based model for daily serum creatinine prediction in critically ill patients with AKI, up to 7 days following diagnosis. Methods: We conducted a retrospective observational study utilizing data from the Mount Sinai Health System. Adult patients who developed AKI within 48 hours of ICU admission were included. Patients were excluded if they received dialysis prior to AKI onset, lacked creatinine measurements after 24 hours post-AKI onset, or were discharged from the ICU within 24 hours of admission. The model incorporated a range of features, including demographics (age, sex, race), vital signs (systolic/diastolic blood pressure, heart rate, respiratory rate, temperature, SpO2, weight), and laboratory values (arterial blood gases, metabolic and enzyme panels). Data were aggregated daily, covering one day before and up to seven days after AKI onset. A transformer-based model, designed to handle missing time-series data, was trained on 70% of the cohort and tested on the remaining 30%. Results: The study included 13,970 patients (mean age: 66.3 years; 58.5% male; 21% White, 32.3% Black, 17.9% Hispanic, 28.7% Other). The overall mean difference between predicted and observed daily serum creatinine was 0.01 mg/dL (p = 0.459). Day-specific predictions showed mean differences of -0.01 mg/dL on Day 1 (p = 0.384), -0.01 mg/dL on Day 2 (p = 0.459), 0.00 mg/dL on Day 3 (p = 0.907), 0.01 mg/dL on Day 4 (p = 0.707), 0.05 mg/dL on Day 5 (p = 0.021), and 0.06 mg/dL on Day 6 (p = 0.025). Conclusion: The transformer-based model effectively predicted daily serum creatinine levels in critically ill AKI patients, showing potential to support real-time kidney monitoring and personalized patient management in the ICU. Funding: NIDDK Support