Patient Similarity Analysis For Diagnostic Decision Support

Analogical reasoning is one of ways that physicians make diagnostic decisions.Making diagnoses by analogical reasoning refers to the process that physicians analyze the patient's condition such as symptoms,physical signs,lab tests,compare them with the theoretical model or the empirical model of diseases,and then obtain the results of diagnoses after differentiation and reasoning.One of the core challenges of analogical reasoning for diagnostic decision support is to accurately and qualitatively assess similarity among patients.Current research methods generally use traditional distance measurements to calculate patient similarity based on lab test data.The limitations of this approach lie in failing to utilize various types of clinical data,failing to consider that the patient similarity is variable from varying perspectives,and failing to consider whether the process is in line with the physicians' thinking logic.How to use multiple types of clinical concepts and develop mathematical models conforming to clinical analogical reasoning to calculate patient similarity for different scenarios remains a challenge.The researches on the patient similarity calculation method by utilizing multi-type clinical data for diagnostic decision support are conducted in this dissertation.Main contents and conlusions are listed as follows:(1)A method for normalizing clinical data in multiple semantic spaces for patient similarity analysis is proposed.Clinical data containing medical knowledge can not be directly applied to calculation,first of all,should be standardized.In this thesis the characteristics of four types of clinical data including health records,lab tests,diagnoses,and drugs are discussed,and corresponding normalization methods are poposed.A pipeline for constructing a clinical term vector from the health records to represent the semantic content is proposed,within which symptoms are identified and negation is detected.The z-score method introducing normal reference range and the one-hot coding method are utilized to normalize the lab test data.Methods for automatic encoding diagnoses with hierarchical categorical ICD codes,and methods for automatic encoding drugs based on the correlation between drugs and diagnoses are proposed.(2)A method for calculating the similarity between diagnoses concept and diagnosis sets based on the hierarchical semantic space is proposed.There are large number of hierarchical concepts in clinic,among which the diagnosis is one of the most important concepts and a significant evaluation index of patient similarity.By using the hierarchical,clinical knowledge-rich ICD codes of diagnoses,a method for computing the similarity between diagnoses and diagnosis sets based on information content of ICD codes is proposed.There are multiple ways for the implementation of each step and the applicable scenarios and performance of combinations of three steps are discussed.This method is generally applicable for similarity measurement of other clinical concepts in the hierarchical semantic space.(3)A method of measuring patient similarity for diagnostic decision support is proposed.Predicting the diagnosis set is a multi-label classification problem.The routine method is to transform this problem into multiple independent binary classification tasks,which ignores the correlation between diagnoses.Inspired by the psycological structure mapping theory,in this thesis a diagnosis prediction model based on patient similarome is designed,where the multi-label problem is transformed into a single-value regression problem.To meet the structure consistency constraint,the model input is the attribute similarity between patient pairs,that is,the feature vectors obtained by comparing clinical terms,lab tests and preliminary diagnoses one by one.The model output if the similarity between ICD codes of patient pair's discharge diagnosis set.The relation similarity between patient pairs is obtained by training the model with big clinical data.Finally,diagnosis hypotheses are generated based on positive analogy,and diagnosis hypotheses are excluded based on negative analogy.The final comprehensive predicting result is a set of diagnoses.The model retains the correlation between diagnoses.Experimental results on a real-world task demonstrate the superiority of proposed method over baselines.Patient similarity can be used to construct a patient similarity network.This thesis explores how to use content-rich patient similarity network for diagnostic decision support.The structure of the patient similarity network is established according to the preliminary diagnosis similarity.The content of nodes is health record text.The processing method of the text confirming the structural consistency constraint neglects text content other than clinical terms.Inspired by the psychological constraint satisfaction theory,the processing method that constructs content embedding of health records by using word embedding can better represent the semantic content and meet the semantic similarity constraint.The graph convolutional neural network-based method is proposed to predict classification labels of patients by learning representations of nodes.The patient representation and patient similarity is specific to the classification task,which satisfies the semantic pragmatic constraint.This method is applicable to other patient similarity network-based multi-class classification problems by using content-rich patient similarity network.Experimental results on a real-world task demonstrate the superiority of proposed method over baselines.