Deep Learning Flags 'Hidden' Acute Kidney Injury in ICU Patients by Reading Creatinine Trajectories

The unit of analysis in this study is small: seven days of kidney blood work, laid out one day at a time. For each stretch, a research team built a two-line record, serum creatinine and an estimated filtration rate called eGFR, and asked a computer model whether the shape of those seven days looked ordinary or strange.

The problem they were chasing is what they call hidden acute kidney injury. In intensive care, doctors usually diagnose sudden kidney damage by checking whether creatinine crosses fixed cut-offs. The authors argue those rules can miss injury that matters clinically, because of uncertainty about a patient's baseline, dilution effects, or trajectories that differ from patient to patient. The study asked whether watching the trend, rather than a single number, could surface those cases.

The work was a retrospective analysis of de-identified public data, exempt from review by the Institutional Review Board of Seoul National University Hospital. The model was developed and internally tested on two databases, MIMIC-III and MIMIC-IV, and then checked externally on the multi-center eICU Collaborative Research Database. Admissions needed at least two kidney measurements. The team excluded patients who started kidney replacement therapy or died within 24 hours of ICU admission, and those with end-stage kidney disease. The MIMIC data were split randomly into 80 percent for training, 5 percent for validation and 15 percent for testing. The external check on eICU ran without refitting or recalibrating the model.

Each admission's data was placed on a daily grid. Missing days were filled by linear interpolation, with small random perturbations added to unobserved creatinine values, within plus or minus 11.4 percent for values of 1 mg/dL or higher and 0.114 mg/dL for lower values. Series were padded at the start and cut into rolling seven-day windows, matching the seven-day interval in the KDIGO creatinine definition. For every window, the team recorded whether kidney replacement therapy began or the patient died within 24, 48, 72 or 96 hours of the window's last day.

The model itself is an unsupervised Anomaly Transformer. It never saw outcome labels. It learned to reconstruct typical creatinine and eGFR trajectories, and its anomaly score rises when its attention pattern departs from a simple prior favoring nearby time points and the sequence is harder to rebuild. The cutoff was set at the 95th percentile of training scores. A window counted as anomaly-detected AKI when creatinine rose on the final day and the score met that threshold.

The central finding, as the authors summarize it, is that higher anomaly scores were linked to a greater chance of starting dialysis or dying within the next few days, including in cases where standard criteria were not met at that same moment. The team compared scores across KDIGO categories and assessed predictive performance with AUROC. They also sorted windows into four groups by whether the anomaly flag and KDIGO stage 2 or higher were present, using odds ratios against a group with neither. Illustrative cases included a patient who did not reach stage 2 but was flagged by the model and began kidney replacement therapy within 48 hours.

The authors are careful about what their numbers mean. Capture rates are calculated per window, and because windows overlap within an admission, they describe how often a rule was positive among outcome windows, not patient-level sensitivity. Robustness checks varied how missing data were handled, tried thresholds at 1, 2.5 and 10 percent tails, and shortened the input window from two to six days. The source text excerpt ends before reporting those results.

The authors frame the signal as a complement to current definitions, not a replacement. Because the work is retrospective and window-based, prospective validation would be needed before any bedside use alongside KDIGO criteria.