(1050) Temporal Subphenotyping of Sepsis Patients With AKI Using Transformer-Based Machine Learning

Introduction: Despite extensive research on the characteristics and outcomes of patients with sepsis who develop Acute Kidney Injury (AKI), the temporal variations at the patient level remain underexplored. This study aims to fill this gap by investigating temporal progression of clinical characteristics following AKI by use of advanced transformer-based clustering methods.

Methods: We conducted a retrospective cohort study using MIMIC-IV database, focusing on adult patients with sepsis who developed AKI within 48 hours of ICU admission. To identify temporal subphenotypes, we employed a machine learning approach utilizing a transformer-based model for clustering multivariate time-series data. We used routinely available clinical characteristics from electronic health records (such as vital signs, labs, vasopressor use, fluid balance etc), collected every 8 hours, to develop these subphenotypes. We collected data for up to 7 days following AKI onset or until ICU discharge, whichever occurred first. Our primary metric for evaluating subphenotypes was the rates of major adverse kidney events within 30 days (MAKE30) defined as a composite of doubling of serum creatinine, new dialysis or mortality within 30 days of AKI onset.

Results: Our study included 12,720 patients, with 58% male and median age of 69 years (IQR: 58-80). Median baseline serum creatinine was 1 mg/dL (IQR: 0.8-1.2). We identified four distinct temporal subphenotypes, comprising 5,839 (46%), 1,185 (9%), 4,890 (39%), and 806 (6%) patients, respectively. Each subphenotype exhibited unique trajectories in terms of vital signs, vasopressor use, lab values, vasopressor use and fluid balance during the study period. MAKE30 rates varied significantly across subphenotypes at 43%, 21%, 29%, and 25%, respectively (p < 0.001).

Conclusions: Our study identified four distinct temporal subphenotypes among sepsis patients with AKI, each characterized by unique clinical trajectories and varying rates of MAKE30. These findings highlight the heterogeneity within this patient population and underscore the potential of transformer-based clustering methods in stratifying patients according to their risk profiles and clinical outcomes.