Deciphering Critical Transition for AKI Deterioration by STIfuse: Fusing Dynamic and Static Representations

Abstract Acute kidney injury (AKI) is a life-threatening condition with notably high in-hospital mortality, characterized by a three-phase progression: a relatively stable state (without imminent life threat), a critical pre-deterioration state (the time window which may be identified and in which timely intervention could prevent upcoming irreversible deterioration), and a deteriorated state (severe conditions such as death or rescue events). Accurate early prediction of AKI deterioration and deciphering the critical transition remain challenging because of complex, high-dimensional, noisy, and incomplete clinical data, as well as limited mechanistic understanding of disease progression. To solve this problem, we propose STIfuse (Spatiotemporal Information fusion), a unique dual-view framework that integrates deep learning (DL) and spatiotemporal information transformation (STI) based on Takens’ embedding theorem and critical slowing down (CSD) theory. The DL backbone encodes multi-dimensional static representations to capture structural dependencies, while the STI branch transforms high-dimensional spatial information into a low-dimensional dynamic representation. These views are fused by a multivariate Gaussian distribution: the means capture statistical levels for diagnosis, and the covariances reflect collective physiological fluctuations for early deterioration prediction. An optimal-transport-distance-based contrastive loss is employed to learn discriminative dynamic features between ˈcriticalˈ and ˈstableˈ states. Pretrained on the MIMIC-IV dataset and validated on five external cohorts (including MIMIC-III (with two independent cohorts), eICU, HiRID, and NWICU), STIfuse demonstrated superior performance in predicting 72-hour deterioration. Further analysis of single-cell multi-omics data study uncovered inflammatory cascades linking comorbidities to AKI. STIfuse provides an interpretable, robust and scalable framework for early warning of AKI deterioration prediction, highlighting its potential in utility and scalability for clinical foundation models.