AI Assesses Walkability for Mobility-Impaired Seniors.

Researchers explored the application of in-context learning (ICL) using the TabPFN foundation model to understand how urban design elements affect perceived neighborhood walkability, specifically for older adults with mobility challenges. The study utilized a relatively small dataset of 257 individuals with knee osteoarthritis or a history of falls, analyzing their responses to the Neighborhood Environment Walkability Scale (NEWS-A) survey. TabPFN achieved a macro F1 score of 54.89% for classifying walkability perceptions (Low, Neutral, High), surpassing optimized baseline models like Random Forest and XGBoost. To interpret these results, Shapley Interaction Quantification (SHAP-IQ) was employed, revealing that the model's predictions were largely driven by complex, higher-order interactions between features. For instance, the interplay between average street circuity and the ratio of drivable roads was a primary determinant of perceived walkability. Interestingly, neighborhood greenery's importance only became substantial when combined with an individual's fear of falling or perception of age-friendliness. This demonstrates ICL's superior performance on small datasets and the value of SHAP-IQ for uncovering nuanced predictive logic.