Development and validation of a dual-channel deep learning for continuous acute kidney injury prediction in critically ill patients.

This study aimed to develop and externally validate a real-time, continuous prediction model for 48-h acute kidney injury (AKI) risk in critically ill patients using a dual-channel deep learning model (DC-AKI). The model was developed using electronic health records from 28,099 patients at Beth Israel Deaconess Medical Center and externally validated on two independent cohorts: 3,108 patients from the eICU Database and 2,808 patients from Zhejiang Provincial People's Hospital. Thirty-one time-varying features were updated every 6 h. The DC-AKI model's dual-channel architecture integrated BiGRU networks, convolutional layers, and attention mechanisms to capture multiscale temporal dependencies. The model achieved areas under the receiver operating characteristic curve (AUC) of 0.720 (95% CI, 0.714-0.728) in internal validation, and 0.577 (95% CI, 0.570-0.583) and 0.798 (95% CI, 0.795-0.799) in the two external cohorts. Interpretability analysis via SHAP identified key clinical predictors and individual risk trajectories. In conclusion, DC-AKI demonstrated strong predictive performance in the development cohort and one external validation site, although performance varied substantially across institutions. Further validation and local calibration are warranted to support its clinical deployment.