Authors (including presenting author) :
SHUM HP (1), CHAN KC (2), CHAN CCY (1), MAN MY (1), YAN WW (1)
Affiliation :
(1) Department of Intensive Care, Pamela Youde Nethersole Eastern Hospital, (2) Department of Intensive Care, Tuen Mun Hospital
Introduction :
The Acute Physiology and Chronic Health Evaluation (APACHE) II and IV logistic regression model are the international standards for hospital mortality prediction in critically ill patients. Artificial neural network (ANN) has been proposed as an alternative.
Objectives :
This retrospective study aimed to compare the performance of ANN, APACHE II and APACHE IV risk of death (ROD) to predict hospital mortality.
Methodology :
All admissions to the PYNEH ICU from Jan 2010 to Dec 2019 were included. 10% was randomly selected from the included samples to form the validation set. The remaining 90% was used for ANN model development and among them, 70% was assigned for training while 30% was assigned for testing purpose. We trained ANNs with one hidden layer with 12 units. Hyperbolic tangent and Softmax activation functions were employed for the hidden and output layers respectively. The ANN was constructed using the same parameters as in the APACHE IV score model. The performance was assessed with the area under the receiver operating characteristic curve (AUROC) and Brier score.
Result & Outcome :
14503 admissions were included. 4.9% of the recruited cases had at least one missing data and was handled by multiple imputation method. The hospital mortality rate was 19.3%. Using 53 parameters, the ANN model “Development set” (AUROC 0.886, 95% CI 0.879-0.92, Brier score 0.097, Hosmer Lemeshow test p=0.3363) was found to be superior to the APACHE II ROD (AUROC 0.800, 95% CI 0.790-0.810, Brier score 0.1519, p< 0.001) and APACHE IV ROD (AUROC 0.841, 95% CI 0.833-0.850, Brier score 0.1232, p< 0.001) for predicting hospital mortality. The validation set contains 1400 patients. The ANN model “validation set” (AUROC p=0.878, 95% CI 0.857-0.900, Brier score 0.099, Hosmer Lemeshow test p=0.3712) was once again superior than APACHE II ROD (AUROC 0.824, 95% CI 0.796-853, Brier score 0.1436, p< 0.001) but similar to APACHE IV ROD (AUROC 0.869, 95% CI 0.848-0.891, Brier score 0.1141, p=0.3432). ANN model has better calibration than APACHE II and APACHE IV ROD.
In conclusion, the ANN-based model outperforms APACHE II ROD, but similar to APACHE IV ROD for prediction of hospital mortality in adult critically ill patients.
