Updating mortality risk estimation in intensive care units from high-dimensional electronic health records with incomplete data

Abstract

Background: The risk of mortality in intensive care units (ICUs) is currently addressed by the implementation of scores using admission data. Their performances are satisfactory when complications occur early after admission; however, they may become irrelevant in the case of long hospital stays. In this study, we developed predictive models of short-term mortality in the ICU from longitudinal data.

Methods: Using data collected throughout patients' stays of at least 48 h from the MIMIC-III database, several statistical learning approaches were compared, including deep neural networks and penalized regression. Missing data were handled using complete-case analysis or multiple imputation.

Results: Complete-case analyses from 19 predictors showed good discrimination (AUC > 0.77 for several approaches) to predict death between 12 and 24 h onward, yet excluded 75% of patients from the initial target cohort, as data was missing for some of the predictors. Multiple imputation allowed us to include 70 predictors and keep 95% of patients, with similar performances.

Conclusion: This proof-of-concept study supports that automated analysis of electronic health records can be of great interest throughout patients' stays as a surveillance tool. Although this framework relies on a large set of predictors, it is robust to data imputation and may be effective early after admission, when data are still scarce.

Keywords: Clinical decision support systems; Electronic health records; Machine learning; Multiple imputation; Neural network.

Fig. 1

Time-slot formatting of data. For variables with values evolving within a predictive time slot, the latest values were used. Durations of 6 and 12 h were compared for predictive time slots

Fig. 2

Flowchart for the definition of the three cohorts from the MIMIC-III database. Patients were selected according to age, length of stay ≥ 48 h and available data among the selected predictors

Fig. 3

Predictive performances of the elastic net, CNN, LSTM and LSTM-CNN models. Discrimination is represented by the ROC curve (upper figures), and calibration is represented by a smoothed calibration plot showing the observed probabilities (and 95% confidence intervals) according to predicted probabilities (lower figures). The thick gray line shows values expected for a perfect calibration, with observed probabilities equal to predicted probabilities. All estimates are averaged over the 10 repeated 5-fold cross-validation datasets and over the imputed datasets for the Imputed-19 and Imputed-70 cohorts