Identification of AKI Subtypes Using Deep Learning-Based Clustering of Serum Creatinine Trajectories and Clinical Features in Critically Ill Patients

Background: This study aimed to identify subtypes of ICU patients with AKI by clustering serum creatinine trajectories and clinical features, then compare outcomes across clusters. Methods: In this retrospective MIMIC-IV cohort, ICU patients with KDIGO-defined AKI were included; those with end-stage kidney disease or prior KRT were excluded. We gathered creatinine measurements (7 days before to 4 days after AKI onset), demographics, vitals, comorbidities, nephrotoxin and vasopressor use, sepsis status, cardiac surgery history, and labs. Missing values were imputed. Autoencoders combined with Deep Embedded K-Means yielded ten clusters, optimized via the Davies–Bouldin index. We compared 7-, 14-, and 30-day mortality and KRT initiation rates across clusters. Results: A total of 30,096 ICU patients with AKI were included, and ten clusters were identified. Baseline creatinine differed significantly among clusters (p<0.001), highest in cluster 6 (2.90 ± 2.62 mg/dL) and lowest in cluster 2 (1.43 ± 1.10 mg/dL). Nephrotoxic medication use was highest in cluster 7 (46.8%) and lowest in cluster 8 (15.1%). Sepsis incidence was highest in cluster 5 (85.4%) and lowest in cluster 1 (14.1%). Regarding creatinine trajectories, clusters 5 and 7 exhibited abrupt elevations, while other clusters showed gradual increases. Clinical outcomes varied significantly across clusters (all p<0.001). The 7-day mortality rate was highest in cluster 3 (2.44%), followed by cluster 6 (0.53%), while the 7-day KRT initiation rate was highest in cluster 7 (25.00%), followed by cluster 8 (3.45%). Conclusion: Deep learning–based clustering delineated distinct AKI subtypes with divergent trajectories and outcomes, offering potential for tailored risk stratification and targeted research.