Individual Length Of Stay Prediction Based On Patient Clinical Similarity

EMR data is an objective record of the whole process of patients’ treatment,which contains lots of clinical information closely related to the patient’s condition.It provides data resources for the development of intelligent clinical decision support.However,there are many problems in real EMR,such as data sparsity,low quality and so forth.In addition,the text data in EMR also contains rich clinical information,but it has not been fully used in the research of clinical decision support.In this paper,we will fully consider all kinds of data in the electronic medical record and,use machine learning and data mining technology study the relevant methods of medical decision support.This paper proposes a personalized length of stay(LOS)prediction method based on patient clinical similarity,which is an organic whole composed of many algorithms,including data processing,patient similarity measurement,feature selection,prediction model,model interpretation.The key technologies are studied,mainly including:(1)Research on data cleaning and structure method of EMR.A series of data cleaning algorithms are designed to standardize the low quality structured EMR data;Doc2vec based on character granularity is proposed to transform unstructured text data into vector representation.(2)Research on the measurement method of clinical similarity of patients.In this paper,we propose a more comprehensive method to measure the clinical similarity of patients,which makes full use of various patient characteristics in the electronic medical record to represent the patient’s condition,and effectively evaluates the distance between these characteristics,further,to realize the similarity calculation between any two patients’ conditions.(3)Research on prediction algorithm of LOS based on patient similarity.In this paper,based on the proposed similarity measurement method,the personalized prediction model is established for each patient from the similar sample queue.Additionally,the key features affecting the prediction target are identified by feature selection technology,and the prediction model is learned by supervised learning technologies such as random forest,XGBoost,light GBM and decision tree.Finally,in order to enhance the understanding of the decision-making results of the individual prediction model,this paper studies the application of SHAP visualization method to enhance the explanatory power of the model in terms of feature level and the individual decision-making results level.In this paper,the proposed LOS prediction method is applied to the real EMR data of cirrhotic patients with ascites,and the performance of the method is evaluated by multi group data experiments.The main results are as follows:(1)the 15 key features identified in this study can effectively represent all the features and to predict LOS of cirrhotic patients with ascites and to establish the final model.(2)The prediction model of LOS based on patient similarity in this paper can show better stability and effect,and also shows the effectiveness of the similarity measurement method in personalized diagnosis and treatment decision-making.(3)The performance of the four personalized prediction models can reach the same level as the general model,or even better than the general model.(4)It is found that the low level of "prealbumin" in cirrhotic patients with ascites is most related to the longer LOS.