Causal deep learning for real-time detection of cardiac surgery-associated acute kidney injury: derivation and validation in seven time-series cohorts

BACKGROUND: Cardiac surgery-associated acute kidney injury (CSA-AKI) is a complex complication substantially contributing to an increased risk of mortality. Effective CSA-AKI management relies on timely diagnosis and interventions. However, many cases are detected too late. Despite the advancements in novel biomarkers and data-driven predictive models, existing practices are primarily constrained due to the limited discriminative and generalisation capabilities and stringent application requirements, presenting major challenges to the timely and effective diagnosis and interventions in CSA-AKI management. This study aimed to develop a causal deep learning architecture, named REACT, to achieve precise and dynamic predictions of CSA-AKI within the subsequent 48 h. METHODS: In this retrospective model development and prospective validation study, we included adult patients (aged ≥18 years) from seven distinct cohorts undergoing major open-heart surgery for model training and validation. Data for model development and internal validation were sourced from electronic health records of two large centres in Beijing, China, between Jan 1, 2000, and Dec 31, 2022. External validation was conducted on three independent centres in China between Jan 1, 2000, and Dec 31, 2022, along with cross-national data from the public databases MIMIC-IV and eICU in the USA. To facilitate implementation, we also developed a publicly accessible web calculator and applet. The model's prospective application was validated from June 1, to Oct 31, 2023, at two centres in Beijing and Nanjing, China. FINDINGS: The final derivation cohort included 14 513 eligible patients with a median age of 56 years (IQR 45-65), 5515 (38·0%) patients were female, and 3047 (21·0%) developed CSA-AKI. The external validation dataset included 20 813 patients from China and 28 023 from the USA. REACT reduced 1328 input variables to six essential causal factors for CSA-AKI prediction. In internal validation, REACT achieved an average area under the receiver operating characteristic curve (AUROC) of 0·930 (SD 0·032), outperforming state-of-the-art deep learning architectures, specifically transformer-based and long short-term memory-based models, which rely on more complex variables. The model consistently outperformed in external validation across different centres (average AUROC 0·920 SD 0·036) and regions (0·867 0·073), as well as in prospective validation (0·896 0·023). Compared with guideline-recommended pathways, REACT detected CSA-AKI on average 16·35 h (SD 2·01) earlier in external validation. INTERPRETATION: We proposed a causal deep learning approach to predict CSA-AKI risk within 48 h, distilling the complex temporal interactions between variables into only a few universal, relatively cost-effective inputs. The approach shows great potential for deployment across hospitals with minimum data requirements and provides a general framework for causal deep learning and early detection of other conditions. FUNDING: The Construction Project and the National Natural Science Foundation of China.