Discussion
The time period selected for the operation of the complex EDI trigger was from Days 11 – 35 of each response, with the response beginning once the first IP (index case) was discovered. Day 11 was both the earliest day that a 5-day EDR could be calculated (excluding Day 1 of the response) and the earliest possible date by which a vaccination programme could be initiated, given the requirement to confirm the strain type of FMD virus, receive sufficient vaccine doses from the international vaccine bank and arrange teams of personnel to administer the vaccine. Day 35 represented 5 weeks into the response, and by this time there should be a reasonably clear picture of the spatio-temporal progression of the disease, but still be at an early enough stage that vaccination could assist with outbreak control.
The thresholds for the numbers of IPs used in the trigger were based on the third quartiles of cumulative IPs detected as at days 14, 21, 28 and 35 within the 10,000 outbreaks simulated for the study by Garner and colleagues (2016). Our premise was that should these values be exceeded prior to or equal to those time points, then it would signify a large outbreak was developing (i.e. within the largest 25% of outbreak sizes). Whilst this was somewhat arbitrary, it was felt that it was not unreasonable.
The trigger mostly fired between Days 11 – 21, with the largest proportion of firings on Day 11 (see Figure 1). Overall, the trigger had very high sensitivities for predicting large and long epidemics. Conversely, the trigger had poor specificity and PPV, but very high NPV. What this means is that the trigger missed very few outbreaks that subsequently turned out to be large, but not all of the outbreaks predicted to be large or long turned out to be so. Putting it another way, if the trigger did not fire during Days 11 – 35, the subsequent outbreak size was most likely to be small or short. Sensitivity analysis of the thresholds for classifying outbreaks as large or long showed that the performance characteristics of the trigger did not change much, with slight increases in sensitivities as the threshold was raised from the 70th percentile to the 80thpercentile, with corresponding small decreases in specificities. The ability to identify most of the large and long epidemics during the early stage of simulated responses showed the utility of the trigger in providing inputs that could assist the decision-makers responsible for making response decisions based on epidemic situation, including the need for urgent implementation of vaccination to strengthen the eradication efforts.
An assessment of the relative performances of the IP and EDR components showed that they were both of similar high sensitivity, but the EDR component had lower specificity and PPV measures (see Tables 8 & 11). Note that the IP and EDR components were not strictly independent of each other, as they operated in parallel and the threshold that fired first was the one that was recorded. Further, the trigger could only fire once for each outbreak, so this did not mean that thresholds were not exceeded later, it’s just that there were relatively few first firings beyond Day 21. It appears that the lower specificities and PPVs demonstrated by the EDR component were because the 5-day EDR values were less stable and could exhibit short-term spikes. Nevertheless, the overall results suggest that the complex EDI used could identify large and long outbreaks during the early stages of simulated outbreaks.
There are several options for trying to improve the overall performance of the EDI trigger. Given the high number of firings that occurred on the first day that the trigger started operating (Day 11), the trigger could be adapted to operate from Day 8 onwards by using a 4-day EDR together with cumulative IP number thresholds appropriate to the time frame, then switch to a 5-day EDR on Day 10 and then to a 6-day EDR on Day 12 of the response. It must be acknowledged however, that it is very difficult to get accurate assessments of disease dynamics very early in the response phase, as new IPs would comprise a combination of cases that were infected during the ‘Silent Phase’ but still be in the process of being discovered, as well as newly infected cases. To try and improve the specificity of the trigger, particularly of the EDR component, longer baseline time periods for the EDR calculations would reduce the likelihood of short-term spikes that are not truly indicative of sustained transmission. Further, EDR thresholds could be raised slightly, or it could be made a requirement that the EDR thresholds had to be exceeded on two separate occasions for the trigger to fire.
Extrinsic factors that influenced the trigger firing were farm type of the primary case, time to first detection and farm density in the vicinity of the primary case. Time to first detection has been associated with larger outbreaks as reported by other researchers (McLaws & Ribble, 2007; East et al. 2015). Type of primary case farm would affect movement frequencies, and increasing farm densities would allow for higher levels of local spread.
The two logistic regression models that explored factors associated with large and long outbreaks (see Tables 12 and 13 respectively) showed that trigger firing was by far the most influential factor: the odds of a large or long outbreak were many times higher when the trigger fired. The other significant factors were related to resourcing issues – affecting active surveillance and depopulation capabilities. Veterinarians (vet variable) have roles in surveillance within the 3 km patrol zones around IPs and in traced farms as well as in depopulation, and Field Technicians (ft variable) have a role in depopulation. The survfpd variable indicates the number of farms that a veterinarian can visit per day, so increasing numbers of farms that could be visited per day by veterinarians had an effect of reducing the odds of a large and long outbreak.
In conclusion, the study showed that an EDI using a combination of cumulative IP numbers and EDR values indicating sustained spread are highly predictive of the eventual size of the outbreak. Further evaluation and improvement of the complex EDI may lead to a valuable tool to predict the eventual epidemic size based on the observed IP numbers and rates of spread during the early stages of outbreaks. Linking temporal changes in observed dissemination rates to the implementation of response measures can provide insights into the effectiveness of the controls already in place and, if required, highlight the need for additional response measures to deliver more effective control of an FMD outbreak.