Variable selection methods for developing a biomarker panel for prediction of dengue hemorrhagic fever

Area of Science:

• Medical Informatics
• Computational Biology
• Epidemiology

Background:

• Accurate prediction models for severe dengue (DHF) require careful variable selection.
• Clinical and laboratory data from 51 acute dengue infection patients were analyzed.
• The study addresses the challenge of bias and data dependency in feature reduction methods.

Purpose of the Study:

• To compare the performance of various variable selection methods for DHF prediction.
• To identify the most critical clinical and laboratory features for predicting DHF.
• To propose an exhaustive feature space search approach to overcome limitations of existing methods.

Main Methods:

• Comparison of Multivariate Adaptive Regression Splines, Learning Ensemble, Random Forest, Bayesian Moving Averaging, Stochastic Search Variable Selection, and Generalized Regularized Logistics Regression.
• Application of model averaging techniques (bagging, boosting, ensemble learners).
• Utilized deviance chi-square testing for linearity assumptions and bootstrapping for regression coefficient evaluation.

Main Results:

• Ensemble methods demonstrated high accuracy, but Generalized Regularized Regression offered superior predictive power.
• Deviance chi-square testing confirmed linearity assumptions for predictors in the best-performing model.
• Bootstrapping was employed to validate predictive regression coefficients.

Conclusions:

• Feature reduction methods can introduce bias; an exhaustive search is proposed as a solution.
• Interleukin-10 (IL-10), platelet count, and lymphocyte count were identified as key predictors for DHF.
• These findings are based on blood chemistry and cytokine measurements from the study cohort.