Ultra-efficient causal deep learning for Dynamic CSA-AKI Detection Using Minimal Variables

ABSTRACT

Cardiac surgery-associated Acute Kidney Injury (CSA-AKI) is a significant complication that often leads to increased morbidity and mortality. Effective CSA-AKI management relies on timely diagnosis and interventions. However, many cases of CSA-AKI are detected too late. Despite the efforts of novel biomarkers and data-driven predictive models, their limited discriminative and generalization capabilities along with stringent application requirements pose challenges for clinical use. Here we incorporate a causal deep learning approach that combines the universal approximation abilities of neural networks with causal discovery to develop REACT, a reliable and generalizable model to predict a patient’s risk of developing CSA-AKI within the next 48 hours. REACT was developed using 21.5 billion time-stamped medical records from two large hospitals covering 23,933 patients and validated in three independent centers covering 30,963 patients. By analyzing the causal relationships buried in the time dimensions, REACT distilled the complex temporal dynamics among variables into six minimal causal inputs and achieved an average AUROC of 0.93 (ranging from 0.89 to 0.96 among different CSA-AKI stages), surpassing state-of-the-art models that depend on more complex variables. This approach accurately predicted 97% of CSA-AKI events within 48 hours for all prediction windows, maintaining a ratio of 2 false alerts for every true alert, improving practical feasibility. Compared to guideline-recommended pathways, REACT detected CSA-AKI on average 16.35 hours earlier in external tests. In addition, we have established a publicly accessible website and performed prospective validation on 754 patients across two centers, achieving high accuracy. Our study holds substantial promise in enhancing early detection and preserving critical intervention windows for clinicians.