Explainable time-series deep learning models for the prediction of mortality, prolonged length of stay and 30-day readmission in intensive care patients

• DOI: 10.3389/fmed.2022.933037
• Publication Date: 2022-09-28T04:38:15Z
• Authors: Yuhan Deng; Shuang Liu; Ziyao Wang; Yuxin Wang; Yong Jiang; Baohua Liu
• Keywords: Brier score; SAPS II

In-hospital mortality, prolonged length of stay (LOS), and 30-day readmission are common outcomes in the intensive care unit (ICU). Traditional scoring systems machine learning models for predicting these usually ignore characteristics ICU data, which time-series forms. We aimed to use deep with selective combination three widely used predict outcomes.A retrospective cohort study was conducted on 40,083 patients from Medical Information Mart Intensive Care-IV (MIMIC-IV) database. Three models, namely, recurrent neural network (RNN), gated (GRU), long short-term memory (LSTM) attention mechanisms, were trained prediction in-hospital LOS, variables collected during initial 24 h after admission or last before discharge. The inclusion based systems, APACHE II, SOFA, SAPS predictors consisted vital signs, laboratory tests, medication, procedures. randomly divided into a training set (80%) test (20%), model development evaluation, respectively. area under receiver operating characteristic curve (AUC), sensitivity, specificity, Brier scores evaluate performance. Variable significance identified through mechanisms.A total 33 enrolled mortality LOS 36,180 prediction. rates occurrence 9.74%, 27.54%, 11.79%, In each outcomes, performance RNN, GRU, LSTM did not differ greatly. Mortality achieved AUCs 0.870 ± 0.001, 0.765 0.003, 0.635 0.018, top significant co-selected by Glasgow Coma Scale (GCS), age, blood urea nitrogen, norepinephrine mortality; GCS, invasive ventilation, nitrogen LOS; ethnicity readmission.The prognostic established our good ICU, especially addition, GCS as most important factors strongly associated adverse events.