Prediction of hospital mortality among critically ill patients in a single centre in Asia: comparison of artificial neural networks and logistic regression-based model.

INTRODUCTION: This study compared the performance of the artificial neural network (ANN) model with the Acute Physiologic and Chronic Health Evaluation (APACHE) II and IV models for predicting hospital mortality among critically ill patients in Hong Kong.

METHODS: This retrospective analysis included all patients admitted to the intensive care unit of Pamela Youde Nethersole Eastern Hospital from January 2010 to December 2019. The ANN model was constructed using parameters identical to the APACHE IV model. Discrimination performance was assessed using area under the receiver operating characteristic curve (AUROC); calibration performance was evaluated using the Brier score and Hosmer-Lemeshow statistic.

RESULTS: In total, 14 503 patients were included, with 10% in the validation set and 90% in the ANN model development set. The ANN model (AUROC=0.88, 95% confidence interval [CI]=0.86-0.90, Brier score=0.10; P in Hosmer-Lemeshow test=0.37) outperformed the APACHE II model (AUROC=0.85, 95% CI=0.80-0.85, Brier score=0.14; P<0.001 for both comparisons of AUROCs and Brier scores) but showed performance similar to the APACHE IV model (AUROC=0.87, 95% CI=0.85-0.89, Brier score=0.11; P=0.34 for comparison of AUROCs, and P=0.05 for comparison of Brier scores). The ANN model demonstrated better calibration than the APACHE II and APACHE IV models.

CONCLUSION: Our ANN model outperformed the APACHE II model but was similar to the APACHE IV model in terms of predicting hospital mortality in Hong Kong. Artificial neural networks are valuable tools that can enhance real-time prognostic prediction.