Partitioned Feature Selection: Locally Uniform Feature Selection with Structured Feature Grouping for Explainable Predictor Identification

Abstract

Feature selection helps to identify important features in high-dimensional mental health data. However, conventional feature selection (CFS) methods commonly rank all features within a single feature space using one overall statistical or predictive criterion. This global evaluation method does not take into account the fact that symptoms of depression are interrelated and fall within various distinct clinical domains. To address this limitation, we propose a Partitioned Feature Selection (PFS) framework in which the symptoms are first organized into relevant domains, and subsequently, relevant features are identified within each domain. We applied four complementary feature selection methods to determine group-wise individual feature importance. PFS then normalizes and combines these importance scores and weights them according to each domain's association with global severity to compute Integrated Group Feature Importance (IGFI). The resulting IGFI scores provide a unified ranking of features while retaining their domain information. We evaluated PFS on two independent multiclass depression datasets using 16 classifiers. Validation-guided retention selected 21 features for Dataset 1 and 23 features for Dataset 2. Using the selected untuned SVC configuration without SMOTE, PFS achieved test accuracies of 95.65% and 89.60%, with Macro-F1 scores of 95.43% and 92.37%, respectively. The ten-trial random-grouping analysis showed that the predictive effect of clinical partitioning varied between the datasets, while the grouping preserved semantic and clinical interpretability. An exploratory Wilcoxon signed-rank analysis showed that the statistical comparison between PFS and CFS-RFE varied across the two datasets. Additionally, IGFI and SHAP identified similar important features across clinical domains. Overall, PFS integrates clinical domain knowledge with group-wise feature selection to provide a structured and interpretable approach to questionnaire-based depression severity analysis. The task represents questionnaire-based severity estimation rather than prediction of an independently established clinical diagnosis.