Risk of MERS importation and onward transmission: a systematic review and analysis of cases reported to WHO

Background The continuing circulation of MERS in the Middle East makes the international dissemination of the disease a permanent threat. To inform risk assessment, we investigated the spatiotemporal pattern of MERS global dissemination and looked for factors explaining the heterogeneity observed in transmission events following importation. Methods We reviewed imported MERS cases worldwide up to July 2015. We modelled importations in time based on air travel combined with incidence in Middle East. We used the detailed history of MERS case management after importation (time to hospitalization and isolation, number of hospitals visited,…) in logistic regression to identify risk factors for secondary transmission. We assessed changes in time to hospitalization and isolation in relation to collective and public health attention to the epidemic, measured by three indicators (Google Trends, ProMED-mail, Disease Outbreak News). Results Modelled importation events were found to reproduce both the temporal and geographical structure of those observed – the Pearson correlation coefficient between predicted and observed monthly time series was large (r = 0.78, p < 10−4). The risk of secondary transmission following importation increased with the time to case isolation or death (OR = 1.7 p = 0.04) and more precisely with the duration of hospitalization (OR = 1.7, p = 0.02). The average daily number of secondary cases was 0.02 [0.0,0.12] in the community and 0.20 [0.03,9.0] in the hospital. Time from hospitalisation to isolation decreased in periods of high public health attention (2.33 ± 0.34 vs. 6.44 ± 0.97 days during baseline attention). Conclusions Countries at risk of importation should focus their resources on strict infection control measures for the management of potential cases in healthcare settings and on prompt MERS cases identification. Individual and collective awareness are key to substantially improve such preparedness. Electronic supplementary material The online version of this article (doi:10.1186/s12879-016-1787-5) contains supplementary material, which is available to authorized users.


Model-predicted number of imported cases
We computed the number e t, d of MERS-CoV infections exported during week t to destination country d outside the Middle-East region as: where o t, c was the frequency of MERS-CoV cases with onset in week t in region c with population pop(c), f t, c, d was the number of passengers flying from region c to country d on week t, W !" and W !" were matrices described below and ρ was the ratio of reporting (i.e. reported cases to actual cases) in the Middle East. Summation was over 20 source regions in the Middle East.
We first reconstructed the incidence curves i t, c for each Middle East region. For some cases, the date of onset was missing and was imputed using date of hospitalization or date of report (1,2). We averaged all results over 20 imputed time series. The variability due to imputation was however very small in the results (<1%). Then, the number of cases with onset in week t, o t, c , was linked to the incidence of infection in the preceding weeks using the incubation period distribution w !" ( ) (i.e. the fraction of infected cases with disease onset between − 1 and weeks) according to: For example, we have o 2 = i 1 w !" (2) + i(2) w !" (1), i.e. cases with onset in week 2 are those infected in week 1 with 2 weeks incubation period and cases infected in week 2 with 1 week incubation period. We took W !" from a study of South Korean Outbreak (3) (average 6.7 days - Figure S1). The corresponding values of W !" were (w !" (1) = 0.6, w !" (2) = 0.38, w !" (3) = 0.02).
Figure S1: Distributions of Incubation period (infection to onset), onset to hospitalization, and infection to hospitalization for MERS-CoV infection.
We then hypothesized that infected cases could travel outside the Middle-East area until hospitalization. Indeed, cases may travel before onset of the disease, but also after onset, as illustrated by 10 of the 22 exportation cases (45%) who travelled when already ill. Time from infection to hospitalization can be split in two parts: infection to onset, i.e. incubation, and onset to hospitalization. The distribution can then be computed as the convolution of the incubation period distribution and that of onset to hospitalization (W !" ). For onset to hospitalization, we analyzed 521 cases from the Middle East reported to WHO (out of 1291) for whom both onset and hospitalization dates were available. Onset to hospitalization took 4.4 days on average, so that the average duration from infection to hospitalization was 11.1 days ( Figure S1). We obtained w !" (1) = 0.16, w !" (2) = 0.57, w !" 3 = 0.22, w !" 4 = 0.05 for the portion of cases hospitalized the 1st, 2nd, 3rd and 4th week after infection. From the distribution we computed the prevalence of infected cases not already hospitalized in week t p t by: applying an actuarial survival method to take into account hospitalization (cases hospitalized in a given week only contributed half a week to exposure during this week). In matrix form this reads: The latter combined with the matrix equation above yield the prevalence of cases in travelers from the series of number of disease onset with time.
Last, we calibrated ρ by setting the predicted number of imported cases in Europe and North America (United States and Canada) equal to the registered importations in these countries over the period (10 importations). Europe was defined as the 32 countries participating in the ECDC surveillance. We obtained ρ = 82% [47% -164%] -confidence interval based on the likelihood ratio test.
Predicting future risk of importation We computed the predictive probability of the weekly number of importation cases worldwide depending on how many cases were reported in the Middle East in the past month. For these computations, the number of air passengers to each destination was fixed at the annual average, disregarding seasonal variation.
We computed P(E|O), where E is the number of importation cases and O the number of reported cases in the Middle East.

The distributions of observed cases is
where π is the probability of report, fixed here at 0.82 and p(x) the distribution of monthly incidence. We described p(x) as a Gamma distribution with mean 31 and standard deviation 40, to reflect typical values between 2012 and 2015.

The joint distribution of E and O is
The coefficient r summarized the link between incident cases and exported cases. It was computed as where d sums over all destination countries in a continent, c sums over provinces in the Middle East, f(c,d) is the average weekly number of air passengers from c to d, γ c is the fraction of all cases occurring in province c and pop c its population, and α the expected portion of cases from the last month not hospitalized. Values of were 3.18e-3 for Africa, 6.34e-4 for the Americas, 8.43e-3 for Asia, 1.63e-3 for Europe and 4.27e-3 for Oceania.

Describing secondary cases after importation
For a typical importation case, information is summarized as (n, d C , d H ), where n is the number of secondary cases and d X the number of days spent in setting X (Community or Hospital). We modelled n as a Poisson random variable with mean = ( ! , ! ), as described below.

Model formulations
Models included: -dependence on time before isolation: We used where m X corresponded to the average number of secondary cases before isolation (D-) or average number of secondary cases per day before isolation (D+).
-dependence on the setting: We assumed that the number of secondary cases in the community and the hospital were the same ( ! = ! ; model S-) or not ( ! ≠ ! ; model S+).
-overdispersion: We accounted for over-dispersion by using random parameters. More precisely, we adopted gamma-distributed random patient-level parameters ( ! , ! ) with mean and standard deviation ( ! , ! ) and ( ! , ! ). Overdispersion could be present in the hospital, in the community or in both settings.
We studied the set of models obtained by combining these characteristics listed in Table S1. In all cases, parameters of interest were the (daily) number of secondary cases (µ) and, depending on the model, the standard deviation of the random parameters. Models were fit with Mathematica. For the best-fitting model (model 7; P+/D+/S+), we computed the predicted probabilities of more than k secondary cases after importation as E(P(n >= k)) for k=0, 1 and 2.
The probabilities were: where α H = (µ H /σ H ) 2 and β H = σ H 2 / µ H are the shape and scale of the gamma distribution.

Collective attention and awareness and relation with imported case history
For each of the three indicators, being ( ) the attention during the week preceding the date t, we built the timeseries formed by the moving average , with T explored between 1 and 4. A time-series of duration of hospitalization H,i ( ! ) was built by associating for each patient i the number of days of hospitalization H,i to the date of hospitalization !,! . Analogously, a time-series of duration of the period in the community C,i ( ! ) is built by associating for each patient i the number of days in the community C,i to the latest between the dates of arrival and symptoms onset !,! . In order to compare H,i ( !,! ) and C,i ( !,! ) with ( ) we computed the linear interpolation of ( ) to build a daily time-series daily ( ) and then we extracted the sub-series daily ( !,! ) (with = , ) formed by the values of the daily attention time-series corresponding to the dates of hospitalization and arrival/symptoms onset for each case. The correlation between H,i and daily is then measured by the Pearson correlation coefficient. Results reported in the main paper are obtained with T=3.
In order to quantify the impact of attention on the duration of hospitalization and period in the community we identified periods of high attention and compare average length of stay in the community/hospital conditioned to high and baseline attention, [ X,i ] !" and [ X,i ] !" (with = , ). Periods of high attention are defined as the ones for which a t > a !" , with a !" equal to the 75% percentile of a t . We tested alternative definitions of high attention periods by considering 60% and 90% percentile, a !" = a t + a t , with a t average and a t standard deviation of the whole time-series a t , and by considering also an annually varying threshold obtained by computing the 75% percentile for each year separately.  Figure S2: Predicted vs. observed number of cases in Europe and North America. Number of cases are integrated over the whole study period. Bars indicate the 95% prediction interval. Figure S3: Expected number of cases within a week as a function of the observed epidemic activity at the source in the preceding month. Different panels correspond to different continents. Shaded areas indicate 95% and 99.9% prediction intervals.

Risk of transmission following importation
Predictions for the second best fitting model In the Italian importation event, there was uncertainty regarding 2 secondary transmission cases occurring in the community. These 2 cases were finally not confirmed by WHO and were not considered in the main analysis.