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

Abstract

Feature selection plays a fundamental role in identifying clinically meaningful features from high-dimensional mental health data. However, conventional feature selection (CFS) methods typically evaluate features globally, without explicitly considering relationships among depressive symptoms or the heterogeneity of clinically related feature subsets. To address this limitation, we introduce Partitioned Feature Selection (PFS), a clinically informed framework designed to preserve domain-specific information while identifying relevant features for depression severity estimation. PFS first partitions survey items into clinically meaningful domains based on semantic consistency and clinical relevance. Within each domain, four complementary feature-selection methods—Recursive Feature Elimination (RFE), SelectKBest (SKB), Univariate Feature Selection (UFS), and Mutual Information Feature Selection (MIFS)—are independently applied to determine group-wise individual feature importance. Next, PFS normalizes and aggregates the individual feature importance scores and weights them according to each domain's association with global depression severity to compute Integrated Group Feature Importance (IGFI). It then uses cross-validation to determine the optimal number of top-ranked features for the final feature subset. The framework was evaluated using two independent multiclass depression datasets, in which global depression severity labels were determined according to established BDC questionnaire scoring criteria. Sixteen machine learning classifiers were used to evaluate the predictive performance of the selected feature subsets. Experimental results demonstrated competitive predictive performance of PFS across both datasets. The best-performing SVC achieved 94.57% accuracy, 88.91% Macro-F1, and 99.66% AUC-ROC on Dataset 1, and 86.40% accuracy, 89.22% Macro-F1, and 98.61% AUC-ROC on Dataset 2. Compared with CFS-RFE, PFS achieved higher accuracy on Dataset 1 (94.57% vs. 90.22%) and comparable accuracy on Dataset 2 (86.40% vs. 86.13%). Random-regrouping ablation indicated that the principal benefit of clinically informed partitioning lies in preserving semantic and clinical interpretability rather than consistently improving classification accuracy over arbitrary partitions. An exploratory Wilcoxon signed-rank comparison across classifiers further indicated that PFS maintains predictive performance comparable to CFS. Furthermore, comparison between IGFI and SHAP demonstrated agreement in the identification of important features across clinical domains, although their relative rankings differed. Overall, PFS integrates clinical domain knowledge with group-wise feature selection to provide a structured and interpretable approach to questionnaire-based depression severity analysis.