Assessment of suitable hospital location using GIS and machine learning

Choosing the suitable site for a new hospital is a difficult aspect of decision-making for the decision-makers. Large data availability with challenging features and the proliferation of different methodologies have made it extremely difficult to select the best models that perform for a particular site selection problem. This research provided a comprehensive study for hospital site suitability and introduced machine learning models. The experiment was conducted in two study areas which are Gaza Strip, Palestine and Melaka, Malaysia. First, the conditioning factors were optimized and ranked to identify and select the most correlated factors to predict the suitability of a hospital site by applying the correlation feature selection (CFS) algorithm and the greedy-stepwise search method. Second, to assess the hospital site suitability, three machine learning (ML) models, namely, support vector machine (SVM), multilayer perceptron (MLP) and linear regression (LR) were introduced to predict the suitability of the hospital site. In addition, two multi-criteria decision-making models (MCDM), namely, analytical hierarchy process (AHP) and fuzzy overlay were used to compare the models and verify the results. The ML models Performance were verified using the receiver operating characteristics (ROC) curves and cross-validation with other evaluation metrics; correlation coefficient, root mean square error (RMSE), mean absolute error (MAE), relative absolute error (RAE), as well as root relative squared error (RRSE). The comparison of the model shows that in Melaka and Gaza Strip, MLP (AUC: 92.20%, 84.90%) and AHP (AUC: 91.40% and 83.20%) are more reliable and acceptable for hospital suitability mapping in both locations with consistent and realistic results. The high-level performance and accuracy of the model outcomes supported the conclusion that the proposed methodology in this research can successfully produce a site suitability map for locating new hospitals. Third, an insight into the machine learning models utilized and how their predicted weights affect hospital site suitability mapping was provided. A clear dissimilarity between the ML and MCDM models in terms of the predicted weights characteristics of the conditioning factors in both study areas are discovered. The study has revealed that some conditioning factors are more significant than others because of inherent traits associated with the spatial characteristics of each case study that results in the differences in the weights. Fourth, five location-allocation models were implemented based on the calculus of coverage, mainly implemented in the search for poor coverage to propose new hospital sites in both study areas. In this research, site suitability from one study area to another was verified using all the implemented methods. Thus, the proposed approaches would be effectively and easily replicated in other regions. Moreover, the results of the proposed approaches provided detailed information that would be useful to decision makers to locate the hospital for effective health delivery planning and implementation.

Text KHALED Y. M. ALMANSI - IR3.pdf Download (1MB)

Actions (login required)

View Item