Artificial Intelligence in Early Detection of Acute Kidney Injury: Current Evidence, Methodological Landscape, and Future Perspectives

Acute kidney injury (AKI) remains one of the most diagnostically challenging conditions in clinical nephrology. Conventional AKI criteria rely on a delayed rise in serum creatinine (SCr) and/or a fall in urine output, both of which reflect functional deterioration rather than structural injury and may lag behind the onset of tubular damage by 24 to 48 hours. Artificial intelligence (AI) and machine learning (ML) have attracted considerable interest as tools for reducing the diagnostic delay, and the volume of published prediction models has grown rapidly over the past decade. This narrative review examines the current evidence across the main algorithmic approaches—classical ML, deep learning architectures, hybrid frameworks, and explainable AI (XAI)—as well as the data sources and biomarker inputs that feed them. We pay particular attention to external validation, calibration, alert fatigue, equity considerations, and the practical challenges of embedding these tools into real clinical workflows. Despite area under the receiver operating characteristic curve (AUROC) values ranging from 0.76 (Vernocchi et al., urine-output deep learning model) to 0.92 (Tomasev et al., large-scale VA system model), the gap between model performance on held-outtest sets and genuine prospective clinical benefit has not yet been bridged. We discuss what will be required to close it.