Spatial Suitability Analysis for Site Selection of Healthcare Facilities Using a Multimodal Machine Learning Approach.

This study proposes a novel end-to-end, geospatial and computer vision machine learning pipeline designed to estimate the suitability of geographic areas to affect community health through newly constructed healthcare facilities. The proposed multimodal model integrates several geographic features such as population density, landcover, elevation, and road networks along with the coordinates of current healthcare facilities to identify suitable locations through multi-scalar geospatial features. The proposed approach introduces novel transformations where features are passed as configurable images tiles to overcome the limitations of traditional geospatial analysis. The study utilizes an enhanced version of the OpenStreetMap humanitarian data layer for healthcare facilities, which was filtered based on building footprint, facility name, and online presence to focus on larger facilities and manually verified against government databases to ensure comprehensive coverage. Although the multimodal approach is model-agnostic, the proposed implementation utilizes a Light Gradient Boosted Trees Classifier with Early Stopping for the learning algorithm. This architecture combines the effectiveness of gradient boosting, and features that understand spatial proximity to achieve better accuracy in the classification of multimodal data. The outcome of this study indicates that the proposed model can predict suitable areas for facility construction, thereby facilitating decision-making and future planning in the healthcare sector.