Source-linked AI summary
In-Context Learning to Assess Built Environment Impacts on Perceived Neighborhood Walkability Among Mobility-impaired Older Adults
Houhao Liang, Kresimir Friganovic, Joanne Kua, Noor Hafizah Ismail, Su Su, Bryan Yijia Tan, Navrag B. Singh, Panos Mavros
TL;DR
Small datasets limit conventional models for studying how built environments shape perceived walkability among mobility-impaired older adults. This paper applies TabPFN-based in-context learning with SHAP-IQ, finding stronger classification performance and predominantly higher-order interaction logic.
Problem
Conventional models’ performance and interpretive fidelity often depend on large datasets, limiting evidence for small-sample built-environment walkability analysis.
Method
The study uses TabPFN for in-context classification of NEWS-A walkability perceptions and SHAP-IQ to interpret hierarchical feature interactions.
Results
TabPFN achieved a 54.89% Macro F1 score, outperforming tuned Random Forest and XGBoost, while circuity–drivable-road interactions primarily discriminated perceived walkability.
Takeaways & Limitations
Perceived walkability predictions were driven mainly by synergistic feature effects, with greenery gaining importance when combined with fear of falling or age-friendliness perceptions.
Takeaways & Limitations
The cross-sectional, self-reported NEWS-A measure captures perceived walkability rather than objective mobility.
Abstract
from arXiv · showhide
As global populations age, enhancing neighborhood walkability through inclusive urban design is important for mitigating built environment (BE) barriers that discourage physical activity and social participation among older adults. This study investigates the utility of in-context learning (ICL), using the transformer-based foundation model TabPFN, to determine how BE features influence perceived walkability, as measured by the Neighborhood Environment Walkability Scale (NEWS-A) survey. Using a small-scale dataset (N = 257) comprising a unique demographic of older adults with knee osteoarthritis or a history of falls, TabPFN achieved a macro F1 score of 54.89% for walkability perceptions categorized as Low, Neutral, and High using equal-width binning. This result outperformed optimized, grid-searched baseline models, including Random Forest (45.85%) and XGBoost (50.56%). To interpret these results, we employed Shapley Interaction Quantification (SHAP-IQ) to identify the hierarchical importance of feature interactions. Preliminary results revealed that the model's predictive logic was primarily driven by higher-order interactions. For example, the interaction between average street circuity and the ratio of drivable roads emerged as the primary discriminator of perceived walkability. Neighborhood greenery was found to have substantial predictive importance only when combined with an individual's fear of falling or perception of age-friendliness. Overall, ICL using TabPFN demonstrates superior performance on small-scale datasets, enhancing the fidelity of the resulting interpretive insights. Furthermore, SHAP-IQ provides a synergistic perspective on how higher-order feature interactions drive the model's predictions.
1 Introduction
The introduction frames neighborhood walkability as a consequential built-environment influence on mobility-impaired older adults while identifying limits in conventional modeling. It motivates investigating transformer-based in-context learning with TabPFN to capture complex interactions in small datasets and interpret perceived walkability measured by NEWS-A.
- Motivation: Built-environment configurations can either facilitate mobility or impose substantial physical constraints on older adults with knee osteoarthritis or a history of falls.The introduction positions the built environment as a dynamic landscape of barriers and supports rather than a neutral backdrop.
- Modeling gap: Linear regression is limited by its linearity assumptions, while Random Forest and XGBoost are increasingly used to model complex walkability interactions.These inductive machine-learning approaches are also commonly paired with SHAP for interpretation.
- Modeling gap: Conventional partition-based models depend on large datasets to form stable decision boundaries without overfitting.This dependence can reduce their performance and the fidelity of their interpretations when datasets are small.
- Proposed approach: In-context learning uses a small dataset as a context set and a frozen transformer foundation model pretrained on large tabular data for approximate inference.This contrasts with inductive models that learn patterns through iterative gradient updates from scratch.
- Study focus: The study explores TabPFN for decoding how the built environment affects perceived neighborhood walkability among older adults with knee osteoarthritis or a history of falls, using the NEWS-A survey.The target demographic is described as uniquely vulnerable, and walkability is operationalized through subjective perception.
2 Methodology
The study used a clinically characterized cohort of 257 mobility-impaired older adults and integrated NEWS-A perceptions with geocoded built-environment and individual health features. It classified relative walkability perceptions with TabPFN in an in-context learning workflow, benchmarked against optimized baselines, and interpreted predictions using SHAP-IQ.
- Study cohort: 257 participants remained after data cleaning and exclusion of incomplete responses, with a mean age of 70.76 years and 68.87% female.The cohort included participants with knee osteoarthritis, a history of falls, or both, and the study received institutional review board approval.
- Measures: The NEWS-A sum score served as the target variable representing subjective neighborhood-walkability perception.Participants also completed LSA, FES, AFEAT, and PHQ-9 surveys.
- Built-environment features: Residential postcodes were geocoded to calculate built-environment metrics relative to each participant’s neighborhood.A 400m radius buffer represented the immediate walkable neighborhood, consistent with the 5D walkability framework.
- Feature integration: The final feature set combined objective built-environment measures with individual clinical and health indicators after multicollinearity filtering.This integration was intended to capture interactions between physiological frailty and environmental characteristics.
- Modeling and interpretation: TabPFN classified NEWS-A scores into Low, Neutral, and High categories using an 80/20 context-set workflow with frozen model weights.Random Forest and XGBoost were extensively optimized through grid-search cross-validation as baseline models.
- Modeling and interpretation: SHAP-IQ quantified synergistic feature interactions alongside traditional SHAP analysis of first-order contributions.Its compatibility with TabPFN made complex interaction values computationally inexpensive compared with inductive machine-learning models.
3 Preliminary Results
TabPFN outperformed tuned Random Forest and XGBoost in the three-class NEWS-A classification task. SHAP-IQ further indicated that perceived walkability was driven more by synergistic feature pairs than isolated predictors, especially walkway circuity and drivable-road ratio.
- Predictive performance: TabPFN was evaluated against Random Forest and XGBoost as inductive models for interpreting built-environment impacts on perceived walkability.The comparison focused on predictive performance before discussing built-environment effects on perceived neighborhood walkability.
- Predictive performance: 55.77% accuracy was achieved by TabPFN, compared with 46.15% for tuned Random Forest and 53.85% for tuned XGBoost.The NEWS-A scores were discretized into Low, Neutral, and High categories, framing the task as three-class classification.
- Feature importance: When assessed independently, AFEAT and FES scores and the steps ratio were among the top predictors of perceived walkability.AFEAT and FES represent individual clinical contexts, while the steps ratio represents a localized morphological barrier.
- Feature importance: SHAP-IQ ranked the interaction between average walkway circuity and the ratio of drivable roads as the leading synergistic effect.This interaction analysis shifted emphasis from isolated factors toward complex variable pairs.
4 Discussion and Conclusion
The study supports TabPFN-based in-context learning as a promising approach for examining walkability perceptions among mobility-impaired older adults, with model decisions primarily driven by feature synergies. The exploratory findings remain limited by the analysis design and subjective, cross-sectional measurement of walkability.
- Contributions: TabPFN-based in-context learning showed potential for understanding how built-environment features interact with walkability perceptions among older adults with knee osteoarthritis or fall histories.The study describes this demographic as unique and reports superior classification performance compared with fine-tuned inductive machine-learning models.
- Interpretation: Model feature-importance results indicated that its decisions were primarily driven by synergistic effects among features.SHAP-IQ was used to examine the hierarchical importance of these feature synergies.
- Limitations and future work: The exploratory analysis lacks cross-validation through conventional statistical methods, motivating future integration of extracted interactions into standard regression models.Such integration is proposed to improve explainability and practical utility.
- Limitations and future work: The cross-sectional design and self-reported NEWS-A scale measure perceived walkability rather than objective mobility.Future studies should add longitudinal GPS tracking to examine how environmental perceptions translate into actual spatial trajectories.