Deep Learning-Based Anomaly Detection to Identify AKI in Patients in the Intensive Care Unit (ICU)

Background: Traditional AKI diagnosis relies on discrete serum creatinine changes, which may miss subtle dynamic shifts in ICU settings. We propose a deep learning–based anomaly detection approach to capture patient-specific temporal patterns. Methods: This study analyzed ICU patients from the MIMIC-IV and eICU databases, excluding those with end-stage kidney disease or prior kidney replacement therapy (KRT). Serum creatinine and eGFR measured at 24-hour intervals—from 24 h before ICU admission until discharge, death, or KRT initiation—were linearly interpolated with physiologic noise. Kidney injury was defined by an Anomaly Transformer model when final creatinine rise yielded an anomaly score above the 5th percentile. KRT and mortality prediction performance was internally and externally validated. Results: A total of 61,373 ICU patients (81,876 admissions) from MIMIC-IV generated 381,700 seven-day datasets (80% training, 5% validation, 15% test). External validation utilized 494,684 datasets from 124,348 patients (140,237 admissions) in the eICU database. Kidney injury by anomaly detection occurred in 3.91% of test cases, slightly higher than AKI stage ≥2 (3.79%). Anomaly detection outperformed AKI stage ≥2 criteria in predicting KRT initiation at 96h (test F1-score: 0.265 vs. 0.206; external F1-score: 0.161 vs. 0.132). Combining both criteria improved mortality prediction at 96h (test F1-score: 0.175 vs. 0.153 for AKI stage ≥2 alone). Anomaly scores were significantly higher in patients with AKI stage ≥2 and adverse outcomes, consistently across datasets. Conclusion: Anomaly detection captures nuanced kidney dysfunction and outperforms conventional AKI criteria for key ICU outcomes, offering a promising complementary diagnostic tool.