DMS Statistics and Data Science Seminar

Speaker: Dr. Fekadu Bayisa (local)

Title: Semiparametric Lasso-like Elastic-net Regularized Spatial Point Process Modelling of Ambulance Call Risk

Abstract: This work analyses the spatial dynamics of ambulance/emergency alarm call events, with the aim of identifying spatial covariates which have association with the events and discerning hotspot regions of the events. The study has been motivated by the problem of designing optimal dispatching strategies for prehospital resources such as ambulances. The dataset at hand is a large-scale marked spatial point pattern of call events. For each event, we have recordings of the spatial location of the call events as well as marks containing the associated priority level, given by 1 (highest priority) or 2, and gender labels, given by female or male. To achieve our objectives, our starting point is to model the spatially varying call occurrence risk as an intensity function of a spatial inhomogeneous Poisson process that we assume is a log-linear function of some underlying spatial covariates. The spatial covariates are related to road network coverage, population density, and socio-economic status of the population in the study area. Since mobility is clearly a factor which has a large impact on where people are in need of an ambulance, and since none of our spatial covariates quantify human mobility patterns, we here take on a pragmatic approach where, in addition to other spatial covariates, we include a non-parametric intensity estimate of the events as covariate in the intensity function. A new heuristic algorithm is developed to select an optimal estimate of the kernel bandwidth to obtain the non-parametric intensity estimate of the events and to create other covariates. Since we consider a large number of spatial covariates, as well as their products (the second-order interaction terms), and since some of them may be strongly correlated, lasso-like elastic-net regularisation has been exploited in the log-likelihood intensity modelling in order to carry out variable selection and to reduce variance inflation from overfitting and bias from underfitting. As an effect of the variable selection, the fitted model structure contains individual covariates, which are of both road network and demographic kinds. In particular, we find that hotspot regions of the calls have been observed along dense parts of the road network in the study area. Moreover, a mean absolute error evaluation of the proposed model to generate the intensity of emergency alarm/ambulance call events indicates that the estimated model is stable and can be utilised to generate a reliable intensity estimate over the region that can be used as input in the problem of designing the dispatching strategies for prehospital resources.