Phenotypic features outperform raw geolocation for community level cognitive difficulty prediction using BRFSS SMART data

Abstract Public health surveillance datasets provide population level signals that may complement clinical data for understanding dementia risk and allocating resources, offering scalable community level coverage that individual clinical records cannot match on their own. In this study, we investigate whether geolocation features (latitude/longitude of metropolitan and micropolitan statistical areas) add predictive value beyond phenotypic (health behavior and condition prevalence) signals when modeling a dementia relevant outcome at the community level. Using CDC’s BRFSS SMART MMSA prevalence dataset, we construct a modeling table (4,289 subgroup location observations across 190 MMSAs in 2023) with three disability status targets, including serious difficulty concentrating, remembering, or making decisions. We formulate a question conditioned regression problem: given the textual description of the target question, phenotypic prevalence features, and optionally geolocation, we predict the observed prevalence (). We evaluate TF-IDF and Sentence-BERT encoders combined with linear models, elastic net, random forests, XGBoost, and LightGBM, using fold wise (train only) imputation to avoid cross fold leakage. Under random 10-fold cross validation, adding geolocation improves performance over text only (R $$^2$$ : 0.598 $$\\rightarrow$$ 0.862), but the additional gain beyond text+phenotype is not statistically significant (0.915 $$\\rightarrow$$ 0.916; paired Wilcoxon $$p=0.70$$ ). Under spatial cross validation, used to test generalization to entirely unseen geographic regions, text+phenotype and text+geo+phenotype remain statistically indistinguishable (0.873 vs. 0.873; $$p=0.63$$ ), a result that is robust to the choice of spatial fold definition. A repeated permutation test (30 shuffles of latitude/longitude) likewise found no significant drop in performance ( $$p=0.39$$ ), indicating no reliable evidence that the model uses geolocation beyond noise. SHAP analysis highlights mobility limitation and mental health prevalence as dominant phenotypic contributors; latitude enters the top features when geolocation is included but does not dominate.