Authors: (including presenting author): :
SHUM HP (1), CHAN CCY (1), MAN MY (1), YAN WW (1)
Affiliation: :
(1) Department of Intensive Care, Pamela Youde Nethersole Eastern Hospital
Introduction: :
Streptococcus pneumoniae is a common cause of community-acquired pneumonia which results in significant morbidity and mortality. The logistic regression model is routinely used for mortality prediction in this clinical condition. The performance of other prediction models is not clear.
Objectives: :
This study aimed to compare the performance of the artificial neural network (ANN), classification tree (CT) and logistic regression (LR) model for prediction of 30-day mortality in patients suffered from pneumococcal pneumonia.
Methodology: :
This retrospective analysis included adults with the pneumococcal disease who were admitted to Pamela Youde Nethersole Eastern Hospital from 1 January 2011 to 31 December 2018. Demographics, microbiological characteristics, and outcomes were compared between 30-day survivors and non-survivors. ANN, CT and LR model were developed using the same covariates. For the ANN model (1 hidden layer with 10 neurons), 70% recruited samples was assigned for training while 30% was assigned for testing purpose. Hyperbolic tangent and Softmax activation functions were employed for the hidden and output layers respectively. For the CT model, exhaustive Chi-Square Automatic Interaction Detector (CHAID) was the splitting method. For LR mode, the forward stepwise approach was adopted. The performance was assessed with the area under the receiver operating characteristic curve (AUROC) and Brier score.
Result & Outcome: :
In total, 792 patients had pneumococcal disease; 701 survived and 91 (11.5%) died within 30 days. Notably, 106 (13.4%) patients had invasive pneumococcal disease and 170 (21.5%) patients received intensive care. Age, presence of septic shock and chronic kidney disease were identified as the key predictors for 30-day mortality by all models. ANN offered the best discrimination and calibration (AUROC 0.849, 0.839 and 0.814; Brier score 0.077, 0.080 and 0.084 for ANN, LR and CT models respectively) for 30-day mortality prediction.
In conclusion, age, presence of septic shock and underlying chronic kidney disease are important predictors for 30-day mortality among patients with pneumococcal pneumonia. ANN model surpassed LR and CT model for mortality prediction.
