Deep-learning time-series anomaly detection of acute kidney injury from creatinine–eGFR trajectories in the ICU

Comprehensive Summary

The study develops and evaluates a deep learning model to detect acute kidney injury (AKI) in ICU patients by analyzing how kidney function changes over time. Researchers used two large ICU databases encompassing over 80,000 and 140,000 admissions from MIMIC-III/IV and eICU Collaborative Research Database, respectively. Daily measurements of serum creatinine and estimated glomerular filtration rate (eGFR) were assembled into rolling seven-day windows, and a deep learning algorithm called the Anomaly Transformer was trained to learn what typical and abnormal kidney function trajectories look like. An anomaly score was generated based on the creatinine-eGFR trajectories. Anomaly scores largely reflected the real clinical deterioration seen in patients, even in cases that did not yet meet conventional AKI diagnostic criteria, which was a phenomenon termed by authors as “hidden AKI”. Anomaly scores increased in a stepwise fashion across KDIGO kidney injury severity categories and were higher in patients who required dialysis or died within 24-96 hours. The model achieved an area under the receiver operating characteristic (AUROCs) of 0.83, 0.82, 0.81, and 0.80 for predicting kidney replacement therapy at 1, 2, 3, and 4 days, respectively. Mortality prediction had AUROCs of 0.62-0.66, which was a modest value and expected given that creatinine is a substandard predictor of death. Overall, the study suggests that AI can be trained to detect red flags in lab trajectory patterns and bring to attention clinically meaningful kidney injury before it is clinically apparent.

Outcomes and Implications

Acute kidney injury detection in the ICU has long relied on static creatinine thresholds, which may be masked from effects of fluid dilution, uncertain baseline values, and non-linear kidney function trajectories. The study highlights the role that deep learning can play in advancing patient-centered critical care, which as seen here, where researchers analyzed rolling creatinine and eGFR trajectories in ICU patients to detect clinically significant kidney injury before conventional criteria were met. Unsupervised Anomaly Transformer is an algorithm that is capable of flagging abnormalities in kidney function in patients with near-term dialysis initiation. Currently, KDIGO staging thresholds, which are fixed values, are relied on as the standard for assessing kidney injury. However, these methods may fail to appreciate the dynamic deterioration that unfolds across a week. AI-driven time-series analyses can provide valuable real-time risk stratification and escalation decisions in a timely manner. Thus, clinicians can utilize the anomaly score to initiate earlier nephrology consultations, reassessment of nephrotoxic medications, or dialysis planning: all before a patient reaches a critical threshold. For healthcare systems, especially smaller or remote facilities with limited specialist resources, AI can be integrated to enable more equitable triage and prioritization. This study is retrospective in design and relies on only two readily available lab values and has yet to be validated in a prospective clinical workflow. Nonetheless, the potential of trajectory-based deep learning is clear and can play a serious role in managing ICU patients.