A A Parsimonious and Authentically Validated Feature Subgroup for Questionnaire-Based Insomnia Classification: A Multi-Method Reduction Benchmarked Across Machine Learning and Deep Learning Models

Abstract

Questionnaire batteries for insomnia screening are long, and most machine-learning studies use every item, leaving unclear which items are necessary. Using a large clinically annotated survey, we examined whether a small, principled subset of features can reproduce the classification performance of the full battery, and whether such a reduction can be validated rather than assumed. Methods: We analysed the Onidra dataset (10,008 participants; four clinician-assigned insomnia grades) with 51 predictor variables. Five complementary feature-selection methods, mutual information, ANOVA F-test, L1-penalised logistic regression, random-forest importance, and recursive feature elimination, were applied, and the resulting subgroups were compared with the full battery across ten classifiers (seven machine-learning and three feed-forward deep neural networks) using stratified five-fold cross-validation, multiclass Brier score, expected calibration error, and permutation importance. To avoid optimistic bias, feature selection was repeated inside every training fold, and differences were assessed with repeated cross-validation, the Friedman test, and Wilcoxon signed-rank tests. Results: Three of the five methods returned an identical ten-variable subgroup, and seven variables were selected by all five. Each consensus subgroup reproduced full-battery accuracy across all model families; for the strongest model the ten-feature subgroup reached 0.966 accuracy versus 0.965 for all 51 features, and was statistically non-inferior to the full battery (Wilcoxon p < 0.05) with a lower expected calibration error (0.013 versus 0.028). A nested, select-within-fold analysis reproduced the estimate exactly, and a parsimony analysis indicated that most of the model-usable signal occupied roughly five items. Conclusions: On this cohort, a ten-variable subgroup reproduced the performance of the full 51-variable battery with comparable or better calibration, and the reduction was unbiased under nested validation. The compact subgroup is a candidate efficient and transparent alternative to the full questionnaire; external validation on independent cohorts is the natural next step.