Development and validation of a machine learning based suitability assessment model for primary healthcare facilities in metropolitan China

Primary healthcare (PHC) facilities play a critical role in preventive care and general medical services. Current urban PHC site selection, typically a manual process at the community level, suffers from insufficient consideration of surrounding neighborhoods, resulting in poorly coordinated service networks and low operational efficiency of facility selection. Machine learning-based modeling approaches offer a promising solution by enabling multi-factor integration for site suitability assessment, thereby assisting governments in rapid decision-making. However, no comprehensive machine learning framework has yet been established specifically for spatial suitability evaluation of PHC facilities. A model framework was developed to evaluate site selection suitability for PHC facilities. We collected geographical data of PHC facilities in Beijing, Shanghai, Guangzhou, and Shenzhen, along with three categories of explanatory variables: physical geographical factors, socioeconomic factors, and urban built environment factors. All spatial data were converted or resampled into 100 m×100 m grid units as standardized analytical units. Machine learning models, including Random Forest (RF), Gradient Boosting Decision Trees (GBDT), and Extreme Gradient Boosting (XGBoost), were subsequently constructed and trained. The best-performing model was then selected, with SHapley Additive exPlanations (SHAP) and variable importance metrics applied to interpret factor contributions. The selected model was ultimately employed to generate predictive site suitability maps for PHC facility placement across the four metropolises. Among the tested models, XGBoost demonstrated the strongest predictive performance, achieving a mean AUC of 0.964 across the four cities. Socio-economic factors consistently emerged as the dominant determinants of PHC facility siting, with population density, urban land use classification, and population age structure exerting the strongest influences. Physical geographical factors showed moderate importance, while urban built environment variables contributed relatively less to model predictions. Spatial suitability maps revealed that highly suitable locations were concentrated in dense urban cores and established residential areas, aligning well with observed facility distributions. This study proposes a robust and interpretable machine learning framework for assessing PHC facility site suitability at the metropolitan scale. The findings highlight the central role of socio-economic demand factors in shaping PHC facility locations, while underscoring the secondary influence of physical geography and limited contribution of built environment variables. The generated suitability maps provide a rapid, city-wide preliminary screening tool that can support government agencies in scientifically planning and optimizing urban PHC facility networks.