Study On A Hemodialvsis Mortality Prediction Model Based On Anomaly Detection

Most end-stage renal disease(ESRD)patients rely on hemodialysis(HD)to maintain their life,who are facing serious financial burden and a high risk of mortality Due to the current situation of the healthcare system,a large number of HD patients are lost to follow-up,making the identification of patients with a high mortality risk a thorny issue,which brings great challenges to model training.This paper proposed an HD mortality prediction approach based on abnormal detection using longitudinal electronic health record(EHR)data,which can solve the problem of data imbalance caused by loss of follow-up.By introducing the long short-term memory(LSTM)network,we developed LSTM deep autoencoder(LSTM-DAE),the autoencoder can model the correlation between time-varied EHR data,capture the changes in the physical condition of HD patients,and predict the risk of death of patients.It adopted semi-supervised anomaly detection theory to learn from single class data,that is,using only the survival patients to train the model and identifying the dead patients by the autoencoder reconstruction errors.The EHR data were from the Kidney Disease Center of the First Affiliated Hospital of Zhejiang University,and 36/72/108 continuous HD sessions were used to predict the mortality in the prediction window of 90/180/365 days.The performance of LSTM-DAE was compared with six baseline models including logistic regression,support vector machine,random forest,LSTM classifier,isolation forest,and stacked autoencoder.When predicting 90-day mortality using 36 continuous HD sessions,1200 patients were included(survival:1055,dead:145),the area under precision-recall curve(AUPRC)for LSTM-DAE was 0.78,the recall was 0.84 and the F1-score was 0.76,outperforming the baseline models.When varying the observation window length or the prediction window length,experimental results showed that LSTM-DAE displayed significant better performance than baseline models.Through a variable importance analysis,the dialysis time was determined to be the most influential feature in the prediction model.The proposed approach leveraged temporal information in EHR data appears to improve the ability of detection of patients with high mortality risk from imbalanced datasets,and can be used for clinical early warning.