(RS28) Comparison of Models for the Prediction of Death Among Severe Dengue Cases by Statistical and Machine Learning Algorithms

Introduction:

Many studies built predictive models to predict dengue outbreaks, identify dengue and stratify dengue, but none has predicted death in severe dengue cases.

Results and discussion:

The basic logistic regression model had a c-statistic (also area under the receiver operating characteristic curve) of 85.3%, sensitivity of 28.7% and rescaled Brier score of 0.32. A stepwise logistic regression model had a c-statistic of 86.2%, sensitivity of 24.5% and rescaled Brier 0.29. The two best performing models with the highest c-statistic were a rule based C5.0 decision tree, 89.5% and a naive Bayes model, 89.1%. The two most sensitive models were the naive Bayes model, 50.3%, and the logistic regression model. The two best performing models with the highest rescaled Brier were a random forest model, 0.88, and a C5.0 rule-based decision tree, 0.65.

Conclusion:

Quantifying the risk of death and its prediction justify intensifying clinical management and aggressive treatment.