We developed and tested a novel algorithm for the detection and inference of space-time clusters for data sets, the Voronoi Based Space-Time Scan (VBScan). The concept of Minimum Spanning Tree (MST) is adapted with the novel Voronoi distance, which is used to compute the set of potential clusters. This set is then evaluated using the spatial scan statistic, producing the most likely cluster of cases.

The class of problems considered here assumes a point data set to represent the location of individuals in a population, classified either as controls or disease cases, within a limited domain in space-time. The cluster is modeled in space coordinates as a connected graph with tree structure, joining a subset of the disease cases, and in space-time coordinates as a sequence of such trees with space projections that have non-null intersection. A distance measure, named Voronoi distance, is proposed here in order to define a meaningful distance for the construction of a minimum spanning tree (MST) that represents the more likely connections between individuals, in a given graph. This structure allows the direct application of the scan statistics, with the calculation of the likelihood ratio of the estimated cluster.

The Voronoi distance between any two points may also be interpreted as an approximation to the line integral of the population density function over the segment joining those two points. For this reason, the Voronoi MST is the natural extension of the Euclidean MST, taking into account the heterogeneity of the population density. On the other hand, the Euclidean distance is an approximation to the corresponding line integral only when the map is cartogram transformed, in such a way that the population density becomes homogeneous. The Voronoi distance concept is employed once again in our method, after the collection of potential clusters is extracted from the Voronoi MST: it is used to estimate the number of control individuals under the region of influence of each one of the case individuals. This allows the definition of the population associated to each potential cluster, which may be evaluated through the spatial scan statistic.

By proposition 1, we attached a ball of radius *ω*_{
i
} /2 to each case *c*_{
i
} belonging to the cluster *S*. The value *ω*_{
i
} was chosen as the minimum weight of the edges that are incident to *c*_{
i
} in the Voronoi MST. An alternative definition may use the average (or even the median) of the weights of the edges that are incident to *c*_{
i
} , instead of the minimum value of the weights. We have conducted numerical simulations suggesting that there are negligible differences of performance using these alternative definitions, compared with the original definition using the minimum value of the weights. This is a good indication that proposed definition of local population of the cluster is stable.

The results of numerical simulations show that the proposed algorithm, space-time VBScan, has higher power of detection, positive predictive value, sensitivity and computational speed than the space-time Elliptic Scan. The flexibility verified of VBScan allows an enhanced ability to deal with the variation of the disease spread along the time dimension.

An application was presented for Dengue fever incidence, with data available at individual level, in the municipality of Lassance, Brazil. Because we make use of an already existing team of community health agents, originally employed for health monitoring in general, Dengue fever surveillance is very cost effective in our setting, and we can focus our effort on mapping, data collection, data integrity issues and analysis. In a future work, we will use additional zoonosis and environmental data, and apply covariate analysis. This will allow better monitoring and forecasting of outbreaks.

VBScan also includes topological information from the point neighborhood structure, in addition to the usual geometric information. For this reason, it is more robust than purely geometric methods such as the elliptic scan. Those advantages were illustrated in a real setting for dengue fever space-time clusters, where the population spreads along a grid of straight lines according to the street mapping. It is worthy to notice that this kind of geometry of population distribution appears very often in urban environments. In those cases, the employment of VBScan should be recommended.

In the examples that we have analyzed, we observed that the Voronoi distance is very reliable to approximate the population heterogeneity, even for some unusual population distribution patterns, like a city block with zero individuals living in its interior and many individuals living on its borders.

One potential limitation of our analysis is the spatial mobility of individuals from their residences to workplace, which could impair the geographic delineation of the detected clusters. In a future work we will address this issue, using tools such as the workflow scan statistic [2].

The ability for the early detection of space-time clusters of disease outbreaks, when the number of points in the dataset is large, was shown to be feasible, due to the reduced computational load of the proposed methodology compared with classical methods. The proposed methodology is shown to present an enhanced power for the detection of space-time disease clusters.