Research On Readmission Prediction By Fusion Unstructured Data In EHR

Unplanned readmission(referred to as readmission)after a patient is discharged is the main source of the cost of the healthcare system,which is usually regarded as an indicator of the healthcare quality and hospital performance.Readmission after discharge will not only expose patients to some hospital-acquired infections,but also increase the economic burden of patients and their families,and increase the cost of care.Therefore,the hospital readmission rate has become a key issue for healthcare researchers.The quantification and early identification of unplanned readmission risk will help to improve the quality of care during hospitalization of patients,reduce the economic burden of patients and hospitals,and alleviate the suffering of patients.At present,there has been a large amount of research work on readmission prediction,but most of the work is based on structured data,ignoring the potentially valuable information on unstructured clinical texts.Aiming at the problem that unstructured clinical texts are not fully used in current readmission prediction research,this paper proposes a readmission prediction method based on the unstructured discharge summary.Firstly,the discharge summary of the sequence is transformed into a vector representation by using the pre-training model,and then the semantic information on the discharge summary is further extracted by using the bidirectional LSTM,and finally,the features extracted from the previous steps are classified.This method can learn the deep semantic knowledge of the discharge summary,which is very useful for clinical prediction tasks such as readmission prediction.Our model can also predict readmissions in different time windows,and if the hospital updates the patient’s hospitalization information during the patient’s admission,our method can also dynamically update the patient’s readmission risk.Finally,experiments on the MIMIC-III dataset show that our method has achieved the best results in AUROC and AUPRC compared to other methods of readmission prediction.In view of the problem that most current researches on readmission prediction are based on structured data and ignore unstructured clinical text information,we propose a readmission prediction method that integrates multimodal data with EHR unstructured data.Firstly,the unstructured clinical notes need to be processed into textlevel vectors by Doc2 Vec.Then for structured static information,a static information encoder is used to encode it into a vector.For structured time signals,given the problem that the machine learning method cannot directly process time-series data,therefore,we use the aggregation method to process it into a vector representation.Then the three vectors are fused to form the patient representation,which is input into the ensemble machine learning model for classification.Finally,experiments on the MIMIC-III dataset show that the readmission prediction method that integrates multimodal data has achieved better results in AUROC and AUPRC.At the same time,the results on the prediction of length of stay and in-hospital mortality show that the method has strong portability.