Multivariate evidence-based pediatric dengue severity prediction at hospital arrival

Background: For individuals with dengue-like symptoms arriving at hospitals, early detection of those likely to progress to-or not progress to-severe dengue can be of great use. Methods: We studied 237 Cambodian children hospitalised in Kampong Cham hospital with dengue-like symptoms. Using dengue severity as primary endpoint, we ran univariate analyses and built multivariate random forest classifiers to predict this endpoint using early clinical and laboratory data. Findings: In a random forest analysis using 56 available variables we obtained AUC = 0•94, and for a sensitivity of 90%: specificity = 89%, positive predictive value (PPV) = 74%, and negative predictive value (NPV) = 96%. Platelet count, HDL cholesterol, and ultrasound pleural effusion and ascites were the four variables most associated with severe dengue outcomes. A random forest on only these four variables gave AUC = 0·88, and for a sensitivity of 90%: specificity = 82%, PPV = 64%, and NPV = 96%. Interpretation. Future severe dengue with significant vascular leakage can be correctly predicted at hospital arrival in a large majority of cases using multivariate random forests. In addition to platelet count and ultrasound pleural effusion and ascites, HDL cholesterol level on the day of admission is also a strong predictor of severe dengue.