Dual-core mutual learning between scoring systems and clinical features for ICU mortality prediction.

Perpetually improving mortality prediction in intensive care units (ICUs) via the implementation of eHealth evaluation approaches has become a major research hotspot in the field of medical data mining for the purpose of saving lives. Recently, researchers have attempted to achieve improved prediction accuracy by using only deep learning-based techniques. However, some problems remain. (1) Most of the existing methods utilize independent clinical features to predict mortality by eliminating the correlations between the latent features, but this technique may fail to comprehensively capture and evaluate the statuses of patients. (2) Several clinical features are needed to ensure strong prediction accuracy, but most methods only use static features that are not extendable. (3) An effective practical framework that unifies traditional ICU scoring systems and state-of-the-art deep learning methods to predict mortality is also lacking. (4) Moreover, the interpretability of existing deep learning-based methods needs to be further improved. Therefore, we propose a novel d ual-core m utual l earning f ramework (DMLF) between ICU scoring systems and clinical features for mortality prediction. In particular, we mutually utilize s equential o rgan f ailure a ssessment (SOFA) scores and clinical measurement features to learn a unified model for enhancing the accuracy and interpretability of our DMLF. Experiments conducted on five real-world disease datasets show that the DMLF achieves significantly better prediction accuracy and area under the receiver operating characteristic curve (AUROC) values than six baselines and four state-of-the-art methods. Moreover, clinicians utilize a familiarized SOFA system to conduct mortality prediction and achieve increased interpretability, which benefits the adoption of the proposed framework in real clinical scenarios. [ABSTRACT FROM AUTHOR]