Dr.Deep: Interpretable Evaluation of Patient Health Status via Clinical Feature's Context Learning

Deep-learning-based health status representation learning is a fundamental research problem in clinical prediction and has raised much research interest. Existing models have shown superior performance, but they fail to explore personal characteristics and provide fine-grained interpretability thoroughly. In this work, we develop a general health status evaluation approach based on clinical feature's context learning, named Dr.Deep. It embeds the feature sequences separately based on a multi-channel structure. The inter-dependencies among dynamic features and static baseline information are captured by self-attention to form the personal health context and adaptively make use of the features for patients in diverse conditions. Dr.Deep further improves the multi-head mechanism to comprehensively encourage the diversity of the model and depict the health status. The model abstracts the clinical features to enhance the ones which strongly indicate the health status. It thus can maintain a state-of-the-art performance in terms of prediction accuracy while providing qualitative interpretability. Experimental results on two publicly available real-world EMR datasets demonstrate the effectiveness of Dr.Deep. To facilitate the personalized clinical service and verify the reasonability of the model, we also develop an AI-Doctor interaction system to show the patient's health trajectory and the corresponding vital biomarkers when performing prognosis. The medical findings extracted by Dr.Deep are also empirically confirmed by human experts and medical literature. We release our code and the AI-Doctor interaction system at GitHub https://github.com/Accountable-Machine-Intelligence/Dr.Deep and http://47.93.42.104/challenge/100049. © 2021, Science Press. All right reserved.