Parsimonious Sequential Feature Group Selection Methods for Predicting Treatment Outcome in Chronic Disorders

Abstract In the era of burgeoning high-dimensional data, particularly in fields likemedicine, where data collection is increasing, the need for effective feature selectionin machine learning is paramount. Feature selection helps mitigate challengessuch as noise and irrelevant information, leading to a deeper understanding ofresults and improved predictive capabilities. While traditional methods focus onindividual features, some data naturally groups into cohesive “feature groups”,such as those collected from sensors or assessment batteries. This paper aimsto rank and select these feature groups to maximize predictive performance. Itintroduces tailored feature selection methods targeting entire feature sets ratherthan individual features. By applying these methods to data from a randomizedclinical trial, which utilizes multiple assessment batteries for patient diagnosisand treatment, the study demonstrates improved predictions and insights intothe importance of specific assessment batteries. The results show that utilizingfeature group selection methods allows for the identification of crucial assessmentbatteries for prediction while excluding less significant ones, potentiallysaving costs, time, and computational resources. However, the study is limited to predefined feature groups.