Feature Learning For Clinical Pathway Analysis

Clinical Pathway is an important management tool for the standarization and nor-malization of patient treatment process.It plays a crucial role in care quality improvement and expense control.However,most of existing clinical pathways are designed by do-main experts,which are not only time-consuming,but also non-adaptive for individual requirements when applied for making clinical planning.It brings a lot of difficulties for the application of clinical pathways.Due to the exponential growth of clinical data,data-driven researches about clinical pathway analysis are receiving increasingly attention.While the clinical data are extremely complex and diverse,so that the feature learning is of great importance for these researches.Because different analysis tasks need different features,we study the feature learning problem for two typical kinds clinical pathway analysis tasks.One is the " looking back" task about clinical pathway mining,and the other one is the "looking forward" task about clinical pathway individualized planning.The primary contribution of this study are summarized below:(1)The goal of clinical pathway mining is to discover the treatment process model with high generalization and rich temporality from history data,which can reflect most of patients’ treatment patterns.Considering the day-based management characteristic of clinical pathway,learning the features of each day is critical for clinical pathway mining.This study proposed a topic modeling based feature learning methods for clinical data.Each clinical day would be mapped to topic based features.The topic assignments constraint,topic correlation restriction and medical ontology has been incorporated into the topic modeling algorithm to improve the quality of discovered topics,which lays a foundation for clinical pathway mining.(2)Clinical pathway individualized planning refers to make a clinical planning for future events based on patient’s current clinical information.It based on an accurate prediction model.The more informative features which can accurately and completely reflect the information in data,the better performance of the prediction,and thus more suitable for the planing task.This study proposed a deep learning based feature learning method which contains three different kinds of representations.A self-supervised learning network is constructed according to the different relations among them.The concept about absolute time has been incorporated into the LSTM network structure,which can better capture the temporal information for the representation learning.The comparative evaluations with state-of-the-art methods on real-world clinical dataset shows encouraging results.(3)The output from deep learning model,which is a "black box",can be hardly understood and reasoned.It is unacceptable in clinical domain.This study designed an interpretability optimization strategy for the proposed deep learning based feature learning method.On the one hand,a two layer attention mechanism based on bidirectional LSTM is proposed to locate the core clinical days and the core dimensions of the clinical days.On the other hand,the topic modeling is used to make the representation dimensions topical,so that the dimensions can be easier to understand.The optimized deep learning model gains significant improvement on interpretability.