Fuzzy Soft Sets for Stroke Risk Feature Selection: Age-Adjusted Ranking, Clinical Validity and Robustness to Severe Class Imbalance

Feature relevance for stroke risk is usually judged with crisp statistics or post-hoc model explanations, neither of which describes a risk factor in the graded, linguistic terms clinicians use. This study ranks the candidate features of a public, de-identified stroke dataset (5,109 patients, 4.87% with stroke) with a fuzzy soft set discrimination score (FSDS), which sums the absolute differences between the mean Low, Medium and High memberships of stroke and non-stroke patients. The score is shown to be twice the total variation distance between the two classes' term distributions, which puts continuous and binary features on one scale; for a binary feature it is twice the difference in prevalence. Because crude scores mix risk factors with proxies for age, an age-adjusted score is introduced, computed within age strata and tested by permutation within strata. Without adjustment, "ever married" ranked second and "work type: children" sixth; after adjustment both lost significance (q = 0.16 and 0.84), and only average glucose level, hypertension, BMI and heart disease remained significant beside age (q ≤ 0.048). In ten repetitions of five-fold cross-validation, with ranking repeated inside every training fold, the top eight FSDS features matched the full feature set and ANOVA, chi-square and mutual-information selection (logistic regression ROC-AUC 0.840 against 0.838–0.841; corrected resampled t-test, all p > 0.4). With class weighting, balanced random forests or SMOTE, recall was 42%–80%. FSDS selections were less stable than ANOVA selections in small samples. The contribution is interpretability and confound control, not higher accuracy.