A deep learning-based, unsupervised method to impute missing values in electronic health records for improved patient management

Electronic health records (EHRs) often suffer missing values, for which recent advances in deep learning offer a promising remedy. We develop a deep learning-based, unsupervised method to impute missing values in patient records, then examine its imputation effectiveness and predictive efficacy for peritonitis patient management. Our method builds on a deep autoencoder framework, incorporates missing patterns, accounts for essential relationships in patient data, considers temporal patterns common to patient records, and employs a novel loss function for error calculation and regularization. Using a data set of 27,327 patient records, we perform a comparative evaluation of the proposed method and several prevalent benchmark techniques. The results indicate the greater imputation performance of our method relative to all the benchmark techniques, recording 5.3-15.5% lower imputation errors. Furthermore, the data imputed by the proposed method better predict readmission, length of stay, and mortality than those obtained from any benchmark techniques, achieving 2.7-11.5% improvements in predictive efficacy. The illustrated evaluation indicates the proposed method's viability, imputation effectiveness, and clinical decision support utilities. Overall, our method can reduce imputation biases and be applied to various missing value scenarios clinically, thereby empowering physicians and researchers to better analyze and utilize EHRs for improved patient management.