An Interpretable Temporal Convolutional Network Model for Acute Kidney Injury Prediction in the Intensive Care Unit.

Acute kidney injury (AKI), a common disease in the intensive care unit(ICU), which is one of the most common critical diseases, has a great impact on patients’ life safety and length of stay. No interventions to improve outcomes of established AKI have yet been developed, so prevention and early diagnosis are key. In the past, traditional machine learning methods such as logistic regression and XGBoost mainly used laboratory data and the statistical features of vital signs data to predict the occurrence of AKI. These methods have achieved good predictive performance, but the statistical values obviously cannot represent all the information of vital signs data. Therefore, a TCN-based model is proposed in this paper. The patient's vital signs data in the ICU, laboratory examination data, and statistical features of monitoring data are fused as input to the model. Hilbert-Schmidt independence criterion (HSIC) is introduced into our model to discover more useful temporal features by calculating the independence between statistics and monitoring data. In addition, we also explore the impact of changes in vital signs on the occurrence of AKI based on complete vital signs data. Numerical experiments were carried out on the public data set MIMIC-III. Compared with the four baseline models (Logistic Regression, Support Vector Machine, Long-Short Term Memory Network, and XGBoost), our model showed excellent performance. In the ablation experiment, the effectiveness of HSIC in mining temporal features is proven. Moreover, the proposed model can effectively discover important changes in monitoring data.