Bio-Inspired Wrapper Feature Selection for Predicting Extended Length of Stay After Geriatric Hip Fracture Surgery
Background: Extended length of stay (eLOS) after geriatric hip fracture surgery strains scarce orthopaedic resources and inflates cost under diagnosis-related-group and diagnosis-intervention-packet reimbursement schemes. Machine-learning predictors of eLOS exist but frequently carry many predictors, raising data-collection burden, overfitting risk at low events-per-variable, and poor bedside uptake. Parsimonious models obtained through principled feature selection offer a more deployable alternative for enhanced-recovery pathways. Methods: We analysed a retrospective single-centre cohort (n approximately 1,137) with the outcome defined as postoperative eLOS of at least 14 days. Five binary bio-inspired metaheuristics -- binary grey wolf optimizer, binary particle swarm optimization, binary genetic algorithm, binary whale optimization, and binary Harris hawks optimization -- were benchmarked as wrapper feature selectors around a logistic-regression base learner with random-forest and XGBoost re-evaluations of the selected subsets. We benchmarked the five swarm/evolutionary operators against four filter or embedded baselines (LASSO, recursive feature elimination, mutual-information, and full-feature) under leakage-controlled nested stratified cross-validation across 30 estimator fits (6 independent seeded random-initialisations under a 5-fold outer cross-validation harness), with SMOTE resampling confined to training folds. Results: Discrimination was statistically equivalent across all nine strategies (mean AUROC 0.748 to 0.760; 95 % confidence intervals overlapping for every pair). Chance-corrected selection stability was not. Because raw overlap measures are inflated by large subsets, and the Kuncheva index by between-run variation in subset size, stability was quantified primarily with a chance-corrected index defined for variable subset sizes. Under this measure the swarm-based wrappers were indistinguishable from random selection of equivalent size (bGWO 0.082, 95 % CI 0.044 to 0.120; bWOA 0.047, 0.016 to 0.078), despite raw Jaccard values of 0.568 and 0.539 that exceeded their chance baselines by only +0.023 and +0.008. The binary genetic algorithm combined the greatest parsimony among the bio-inspired wrappers (10.6 +/- 2.3 of 27 predictors) with significantly higher chance-corrected stability (0.225, 0.156 to 0.294; paired t(4) = 4.57 versus bGWO and 5.03 versus bWOA, both p < 0.05). No strategy approached the value obtained for a genuinely fixed feature core (0.61 in simulation). Decision-curve analysis showed positive net benefit against treat-all and treat-none across threshold probabilities 0.10 to 0.50. Conclusions: Bio-inspired wrapper feature selection achieved parsimony without loss of discrimination on a 1,137-patient geriatric hip-fracture cohort, but the two families of operator behaved differently. Among five binary metaheuristics only the evolutionary operator (bGA) produced a reproducible feature subset, at roughly half the size selected by the swarm-based operators; bGWO and bWOA were not distinguishable from random selection once chance-corrected. Where the objective surface is flat, the choice of operator family determines whether a bio-inspired search converges on a reproducible subset or merely on a large one. Stability metrics that are not chance-corrected should not be used to rank feature-selection strategies.