Mortality prediction in intensive care units with the Super ICU Learner Algorithm (SICULA): a population-based study

Abstract

Background: Improved mortality prediction for patients in intensive care units is a big challenge. Many severity scores have been proposed, but findings of validation studies have shown that they are not adequately calibrated. The Super ICU Learner Algorithm (SICULA), an ensemble machine learning technique that uses multiple learning algorithms to obtain better prediction performance, does at least as well as the best member of its library. We aimed to assess whether the Super Learner could provide a new mortality prediction algorithm for patients in intensive care units, and to assess its performance compared with other scoring systems.

Methods: From January, 2001, to December, 2008, we used the Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC-II) database (version 26) including all patients admitted to an intensive care unit at the Beth Israel Deaconess Medical Centre, Boston, MA, USA. We assessed the calibration, discrimination, and risk classification of predicted hospital mortality based on Super Learner compared with SAPS-II, APACHE-II, and SOFA. We calculated performance measures with cross-validation to avoid making biased assessments. Our proposed score was then externally validated on a dataset of 200 randomly selected patients admitted at the intensive care unit of Hôpital Européen Georges-Pompidou, Paris, France, between Sept 1, 2013, and June, 30, 2014. The primary outcome was hospital mortality. The explanatory variables were the same as those included in the SAPS II score.

Findings: 24,508 patients were included, with median SAPS-II of 38 (IQR 27-51) and median SOFA of 5 (IQR 2-8). 3002 of 24,508 (12%) patients died in the Beth Israel Deaconess Medical Centre. We produced two sets of predictions based on the Super Learner; the first based on the 17 variables as they appear in the SAPS-II score (SL1), and the second, on the original, untransformed variables (SL2). The two versions yielded average predicted probabilities of death of 0·12 (IQR 0·02-0·16) and 0·13 (0·01-0·19), whereas the corresponding value for SOFA was 0·12 (0·05-0·15) and for SAPS-II 0·30 (0·08-0·48). The cross-validated area under the receiver operating characteristic curve (AUROC) for SAPS-II was 0·78 (95% CI 0·77-0·78) and 0·71 (0·70-0·72) for SOFA. Super Learner had an AUROC of 0·85 (0·84-0·85) when the explanatory variables were categorised as in SAPS-II, and of 0·88 (0·87-0·89) when the same explanatory variables were included without any transformation. Additionally, Super Learner showed better calibration properties than previous score systems. On the external validation dataset, the AUROC was 0·94 (0·90-0·98) and calibration properties were good.

Interpretation: Compared with conventional severity scores, Super Learner offers improved performance for predicting hospital mortality in patients in intensive care units. A user-friendly implementation is available online and should be useful for clinicians seeking to validate our score.

Funding: Fulbright Foundation, Assistance Publique-Hôpitaux de Paris, Doris Duke Clinical Scientist Development Award, and the NIH.

Figure 1. Receiver-Operating Characteristics curves

Super Learner 1: Super Learner with categorized variables; Super Learner 2: Super Learner with non-transformed variables. These results were obtained using 10-fold cross-validation. We also implemented 50-fold cross-validation and found no material change in the estimated performance of the SICULA algorithm (cvAUC for the SICULA: 0.91 [0.90–0.92]).

Figure 2. Distribution of the predicted probability of death in the survivors and the non-survivors

Upper panel: SAPS II on the left, SOFA on the right; Medium panel: new fit of the SAPS II on the left, new fit of the APACHE II on the right; Lower panel: on the left, Super Learner using categorized variables (Super Learner 1), on the right Super Learner with non-transformed variables (Super Learner 2).

Figure 3. Cross-validated mean-squared error for the Super Learner and the 12 candidate algorithms included in the library

Upper panel concerns the Super Learner with categorized variables (Super Learner 1): Mean Squared Error (MSE) associated with each candidate algorithm (top figure) – Receiver Operating Curves (ROC) for each candidate algorithm (bottom figure); Lower panel concerns the Super Learner with non-transformed variables (Super Learner 2): Mean Squared Error (MSE) associated with each candidate algorithm (top figure) – Receiver Operating Curves (ROC) for each candidate algorithm (bottom figure).