Comparative evaluation of feature selection methods for identifying proteomic predictors of incident major depressive disorder using UK Biobank proteomic data

Machine learning captures non-linear patterns in high-dimensional data; however, challenges such as feature redundancy, overfitting, and limited interpretability highlight the need for feature selection. We compared four feature-selection methods—F-test, LASSO, PowerSHAP, and LLpowerSHAP—for identifying proteomic predictors of incident major depressive disorder (MDD) from 2,920 plasma proteins in 48,378 UK Biobank participants, including 2,384 incident cases. Methods were evaluated using repeated five-fold outer cross-validation (20 repeats; 100 evaluations), and selected features were evaluated using logistic regression, XGBoost, LightGBM, and CatBoost. We compared predictive performance, feature count, selection stability and concordance, and computational cost. Predictive performance was modest and comparable across methods, with small classifier-dependent differences (mean ROC-AUC = 0.614–0.629; PR-AUC = 0.071–0.081). The F-test retained the largest protein sets (357–579 proteins), whereas PowerSHAP yielded the fewest (16–147). The F-test and LLpowerSHAP showed the greatest selection stability (adjusted stability measure = 0.75 and 0.73, respectively), while LLpowerSHAP and PowerSHAP had the highest cross-method concordance (Jaccard = 0.52). Computational demand ranged from 1.25 s per evaluation for the F-test to 3.46 h for LLpowerSHAP. Moderate correlation filtering and covariate inclusion largely preserved predictive performance, whereas stringent filtering reduced performance and altered protein selection; covariate inclusion mainly affected LASSO. Thirty-six proteins were recurrently selected in ≥ 50% of evaluations by all methods, of which 30 had nominal association with incident MDD (OR = 0.88–0.95 for inverse and 1.06–1.23 for positive associations). The fixed 36-protein set showed modest discrimination (ROC-AUC = 0.622–0.641; PR-AUC = 0.078–0.083). Overall, broadly comparable predictive performance masked differences in parsimony, stability, concordance, and computational demand, highlighting trade-offs in feature-selection method choice.