Comparison of feature selection approaches in youth depression determination based on handwriting kinematics

Abstract

Depressive disorder (DD) in youth is a significant, yet underrecognized mental health issue, often accompanied by psychomotor retardation. Handwriting analysis provides a non-invasive and measurable method for detecting such symptoms. This study explores feature selection approaches to improve machine learning-based classification of DD using kinematic features from a task—repetitively writing the lowercase cursive letter “l”. From 177 extracted features, 40 were retained through statistical filtering and further refined using five selection approaches. Logistic regression models were trained and evaluated using subject-wise leave-one-out cross-validation. In this paper, a comparison of different feature selection approaches (Recursive Feature Elimination, Sequential Forward Selection, SHapley Additive exPlanations, Minimum Redundancy Maximum Relevance, and Feature Importance) is presented, considering the occurrence of optimal feature sets as well as the binary classification accuracy of subjects into the DD and control groups.