Interpretive Clinical Measurements Prediction and Diagnosis in ICU with Multi-Task Attention-Guided Sequential Feature Learning

During the outbreak of COVID-19, it is particularly important to improve clinical decision making for critically ill patients in intensive care units (ICU). Machine learning has become increasingly important in healthcare analysis with the wide adoption of Electronic Health Records (EHR). Many deep learning algorithms, such as Recurrent Neural Networks (RNNs), have been developed to deal with clinical time series data to predict the health status of patients, such as mortality, ventilation, and phenotype, etc. Despite RNN being successful in medical tasks, the special traits of data, such as long temporality, high dimensionality, existence of missing values, and multi-labels, have brought challenges to either shallow or deep learning algorithms. Motivated by the fact that the progress of diseases may vary over time, to address these challenges, we proposed a multi-task attention-guided recurrent neural network for learning sequential features (MASFL) model, which employs recurrent units and an attention mechanism. The recurrent units can memorize the past progressive information of patients after entering ICU, while the attention mechanism can selectively utilize this information for different prediction tasks. Meanwhile, the attention also helps to interpret the relationship between the clinical time series data and the prediction result. Furthermore, in order to solve undesirable local solutions for the patient’s length-of-stay task, we develop an ordinal regression loss to address the issue. According to the experiments on MIMIC-III, we demonstrate that the proposed model can improve the prediction performance over state-of-the-art prediction approaches.