A real-time spatio-temporal syndromic surveillance system with application to small companion animals

Abstract

Lack of disease surveillance in small companion animals worldwide has contributed to a deficit in our ability to detect and respond to outbreaks. In this paper we describe the first real-time syndromic surveillance system that conducts integrated spatio-temporal analysis of data from a national network of veterinary premises for the early detection of disease outbreaks in small animals. We illustrate the system's performance using data relating to gastrointestinal disease in dogs and cats. The data consist of approximately one million electronic health records for dogs and cats, collected from 458 UK veterinary premises between March 2014 and 2016. For this illustration, the system predicts the relative reporting rate of gastrointestinal disease amongst all presentations, and updates its predictions as new data accrue. The system was able to detect simulated outbreaks of varying spatial geometry, extent and severity. The system is flexible: it generates outcomes that are easily interpretable; the user can set their own outbreak detection thresholds. The system provides the foundation for prompt detection and control of health threats in companion animals.

Figure 1

The results from our outbreak simulation study when using Scheme 1. In this scheme a single premise i at the centre of each region experiences an outbreak. Here we choose an exceedence level of l = 0 (see supplementary material for other levels). This figure shows the results of 9 simulations plus the baseline level. The top row of timeseries plots is the ‘baseline’, that is the actual SAVSNET data without any simulated outbreak i.e. γ = 0. The subsequent rows from top to bottom depict increasing severities of simulated outbreak labelled according the probability of a case at premise i e.g. p = 0 and so on. The columns, from left to right, relate to the density of the region; ‘sparse’, ‘medium’ and ‘dense’ respectively. For each simulation we plot the timeseries of the predicted distribution of S_i,t for premise i. In each time timeseries the solid black line is the predicted value of S_i,t, shaded areas are pointwise 50%, 90% and 95% predictive intervals. As an aid to rapid interpretation, we use a traffic-light system: if the predictive probability, q, is above 0.99 (defined as ‘very high’) the light shows red, if above 0.9 (‘high’) orange, if above 0.8 (medium) yellow, otherwise (‘low’) green (no outbreak). The outbreak commences on 15^th February. The more intense the outbreak is the more the traffic light system tends towards red.

Figure 2

The results from our outbreak simulation study under Scheme 2. The overall layout and format of timeseries plots is the same as Fig. 1, for details see its caption. The simulated outbreak begins on 15^th February and the timeseries plots are for premise i at the centre of each region. Here we depict results using Scheme 2, that is premise i and its neighbours, within an 8 km radius, experience an outbreak. Again we choose an exceedence level l = 0 (see Supplementary Material for other levels).

Figure 3

Maps of regions in which we simulated outbreaks where a premise is located at a coloured dot. These premises were selected for illustrative purposes, the actual SAVSNET data shows no indication that they are atypical or that they experienced a genuine outbreak during February 2016. As the base layer we use map tiles by Stamen Design, under CC BY 3.0: data by OpenStreetMap, under ODbL. The premise at the centre of each outbreak region is in the middle of the large light grey circle (8km radius). This figure shows the results of 4 simulations for 17^th February 2016 when we use an exceedence level of l = 0; n.b. the corresponding temporal results are given in Fig. 1. and 2. The top and bottom rows relate to the density of the region, ‘sparse’ and ‘dense’, respectively, and the left and right columns relate to simulation Schemes 1 and 2 respectively. The simulated probability of a case at the premise in the centre of each region is $p=0.15$. To aid interpretation, we use the traffic-light system described in Fig. 1 caption, as such each coloured circle on the map is derived from the predicted distribution of S_i,t at each corresponding premise. Panels (a,c) show when the central premise has neighbours who are not experiencing an outbreak it is less able to detect the outbreak, panel (c), when compared to a premise without neighbours, panel (a). If the neighbours also experience an outbreak the system is then better able to detect this outbreak at central premise, panel (d), compared with when the neighbours did not experience an outbreak, panel (c).