Generalizable deep clustering based on Bi-LSTM with applications to sepsis and acute kidney disease populations

Despite the abundance of subphenotype clustering studies on sepsis and acute kidney injury (AKI), few models consider the real-time information of clinical features. The lack of supervision may lead to patient subgroups being derived as clusters without the stratification of patients based on the outcome of interests. The sensitivity of the dimension in clustering methods is generally ignored, so clusters lack robustness. In this study, we propose an ensembled outcome-driven bidirectional long short-term memory autoencoder (BiLSTM-AE) architecture with high robustness and transferability to identify subphenotypes. BiLSTM-AE learns the advanced representation of the time-series clinical features by co-training the encoder and a weak predictor to achieve the risk-stratified clustering of patients. Clusters of a variety of dimensions are ensembled to combine global and local information. Four different datasets from three public datasets, MIMIC-III-AKI, MIMIC-IV-sepsis, eICU-AKI, and eICU-sepsis, were used to assess the method's effectiveness in clustering and prediction. Compared to baseline approaches including latent class analysis (LCA), subgroups generated by BiLSTMAE exhibited the highest mortality risk ratios between subgroups: the mortality for subphenotypes 1, 2, and 3 of BiLSTM and LCA was 6.91%, 17.53%, and 75.56% vs. 13.2%, 14.4%, and 19.7% for MIMIC-III-AKI. The prediction metric area under the receiver operating characteristic curve was 0.86 for MIMIC-IIIAKI, 0.91 for eICU-AKI, 0. SS for MIMIC-IV-sepsis, and 0. S9 for eICU-sepsis. Additionally, clinical evaluation of BiLSTM-AE generated subgroups revealed more meaningful distributions of member characteristics across subgroups. Thus, the method is an effective means to consider the real-time information of clinical features.