Feature Explanations in Recurrent Neural Networks for Predicting Risk of Mortality in Intensive Care Patients

Abstract

Critical care staff are presented with a large amount of data, which made it difficult to systematically evaluate. Early detection of patients whose condition is deteriorating could reduce mortality, improve treatment outcomes, and allow a better use of healthcare resources. In this study, we propose a data-driven framework for predicting the risk of mortality that combines high-accuracy recurrent neural networks with interpretable explanations. Our model processes time-series of vital signs and laboratory observations to predict the probability of a patient’s mortality in the intensive care unit (ICU). We investigated our approach on three public critical care databases: Multiparameter Intelligent Monitoring in Intensive Care III (MIMIC-III), MIMIC-IV, and eICU. Our models achieved an area under the receiver operating characteristic curve (AUC) of 0.87–0.91. Our approach was not only able to provide the predicted mortality risk but also to recognize and explain the historical contributions of the associated factors to the prediction. The explanations provided by our model were consistent with the literature. Patients may benefit from early intervention if their clinical observations in the ICU are continuously monitored in real time.

FLICU: A Federated Learning Workflow for Intensive Care Unit Mortality Prediction

“…For each ICU patient i, we identify one of two critical time points, i.e., the time of death in the ICU (positive class) or the time of discharge from the ICU (negative class). Given a fixed time window W (8h, 16h, 24h, and 48h), for each D ij ∈ D i , and based on Pattalung et al [13], we only consider the medical events that occurred within W hours before the last recorded vital sign or lab value. Knowing the time of ICU death/discharge, we assume that the most important information about the critical event is at the end of the ICU stay, and we want to explore how much of this information (window size) is really relevant based on the predictive performance achieved by the model.…”

“…For each client k ∈ [1, K], the local FL models h k F L (•) are trained using their local cohort D k , and each set of local weights w k is then passed to the central server. The local model consists of two parallel input channels (one for vitals and one for labs), with one Conv1D layer followed by one Flatten layer (kernel sizes 1 and 8) each, for the 1DCNN model, or with 3 RNN layers of 16 units each for the three sequential neural network classifiers (according to [13]). For all classifiers, we perform batch normalization subsequentially, followed by a concatenation of the outputs and fusion of the concatenated outputs via two fully connected layers.…”

“…We use the MIMIC-III dataset [21] for this study, which is a publicly available critical care database containing deidentified clinical data of patients admitted to an ICU at the Beth Israel Deaconess Medical Center (BIDMC) from 2001 to 2012. We follow the approach described in Pattalung et al [13] for the pre-processing and feature extraction steps, expanding on the publicly available code of the mimic-code GitHub repository [22], [23]. We extract patient demographic information for pre-processing and labeling (icustay id, first icu stay, first careunit, length of stay icu, deathtime icu), as well as statistical purposes (gender, ethnicity, admission age), which are presented briefly in Tables I and II.…”

“…Additionally, we extract patients' vital signs and lab values, collected during their ICU stays for modeling. We obtain 23 ICU mortality clinically relevant variables, which are in the form of time series [13], 7 vital signs (heartrate, systolic blood pressure, diastolic blood pressure, mean blood pressure, respiratory rate, temperature, peripheral oxygen saturation), and 16 lab values (albumin, blood urea nitrogen, bilirubin, lactate, bicarbonate, band neutrophi, chloride, creatinine, glucose, hemoglobin, hematocrit, platelet count (platelet), potassium, partial thromboplastin time, sodium, white blood cells). Additionally, we prune data outliers and perform grouping of similar clinical variables (using the item ids, see [22]).…”

“…For example, Johnson and Mark [11] focus on real-time ICU mortality prediction using logistic regression and gradient boosting, while Pattalung and Chaichulee [12] compare multiple ML algorithms for ICU mortality prediction. Additionally, Pattalung et al [13] focus on predicting the risk of ICU mortality by combining Recurrent Neural Networks (RNNs) with interpretable explanations. Finally, Rinta-Koski et al [14] propose Gaussian process classification for mortality prediction in a neonatal ICU.…”