Background: Principal components analysis (PCA) is often used to find characteristic patterns associated with certain diseases by reducing variable numbers before a predictive model is built, particularly when some variables are correlated. Usually, the first two or three components from PCA are used to determine whether individuals can be clustered into two classification groups based on pre-determined criteria: control and disease group. However, a combination of other components may exist which better distinguish diseased individuals from healthy controls. Genetic algorithms (GAs) can be useful and efficient for searching the best combination of variables to build a prediction model. This study aimed to develop a prediction model that combines PCA and a genetic algorithm (GA) for identifying sets of bacterial species associated with obesity and metabolic syndrome (Mets).
Results: The prediction models built using the combination of principal components (PCs) selected by GA were compared to the models built using the top PCs that explained the most variance in the sample and to models built with selected original variables. The advantages of combining PCA with GA were demonstrated.
Conclusions: The proposed algorithm overcomes the limitation of PCA for data analysis. It offers a new way to build prediction models that may improve the prediction accuracy. The variables included in the PCs that were selected by GA can be combined with flexibility for potential clinical applications. The algorithm can be useful for many biological studies where high dimensional data are collected with highly correlated variables.
Keywords: Biomarker; Genetic algorithm; Obesity; PCA.