A novel multivariable time series prediction model for acute kidney injury in general hospitalization

Early recognition and prevention are important to reduce the risk of acute kidney injury (AKI). We aimed to build a novel multivariate time series prediction model for dynamic AKI prediction in general hospitalization.

Deidentified electronic data of all patients admitted in Sichuan Provincial Peoples Hospital during 1 January 2019 and 31 December 2019 was retrospectively collected. Variables including demographics, admission variables, lab investigation variables and prescription variables were extracted. The first 50 most frequently detected lab investigation variables were selected as the predictive variables. Features within three previous days were selected to predict the risk of AKI in the next 24 h. The model was built using recurrent neural network (RNN) algorithm integrated with a time series convolution module and an attention convolution module and internally validated using five-fold cross-validation. Area under the ROC curve (AUC) and recall rate were used to evaluate the performance. The model was compared with four other models built using other machine learning algorithms and published machine learning models in literature.

47,960 eligible admissions were identified, among which 2694 (5.6%) admissions were complicated by AKI. Our model has an AUC of 0.908 and a recall rate of 0.869, outperforming models generated by mainstay machine learning methods and most of the published machine learning models.

This study reports a novel machine learning prediction model for AKI in general hospitalization which is based on RNN algorithm. The model outperforms models generated by mainstay machine learning methods and most of the published machine learning models.