Applying the moving epidemic method to determine influenza epidemic and intensity thresholds using influenza‐like illness surveillance data 2009‐2018 in Tunisia

Abstract Background Defining the start and assessing the intensity of influenza seasons are essential to ensure timely preventive and control measures and to contribute to the pandemic preparedness. The present study aimed to determine the epidemic and intensity thresholds of influenza season in Tunisia using the moving epidemic method. Methods We applied the moving epidemic method (MEM) using the R Language implementation (package “mem”). We have calculated the epidemic and the different intensity thresholds from historical data of the past nine influenza seasons (2009‐2010 to 2017‐2018) and assessed the impact of the 2009‐2010 pandemic year. Data used were the weekly influenza‐like illness (ILI) proportions compared with all outpatient acute consultations. The goodness of the model was assessed using a cross validation procedure. Results The average duration of influenza epidemic during a typical season was 20 weeks and ranged from 11 weeks (2009‐2010 season) to 23 weeks (2015‐2016 season). The epidemic threshold with the exclusion of the pandemic season was 6.25%. It had a very high sensitivity of 85% and a high specificity of 69%. The different levels of intensity were established as follows: low, if ILI proportion is below 9.74%, medium below 12.05%; high below 13.27%; and very high above this last rate. Conclusions This is the first mathematically based study of seasonal threshold of influenza in Tunisia. As in other studies in different countries, the model has shown both good specificity and sensitivity, which allows timely and accurate detection of the start of influenza seasons. The findings will contribute to the development of more efficient measures for influenza prevention and control.


| INTRODUC TI ON
Seasonal influenza continues to be a public health problem worldwide. Although in most cases, it leads to an increased number of consultations, it may cause severe illness and death especially among high-risk groups. In fact, the World Health Organization (WHO) has recently updated global estimates to more than 3 million severe cases and from 290 000 to 650 000 respiratory deaths due to influenza each year. 1 These annual epidemics mobilize considerable resources from health services and even a small-scale epidemic can have a significant socio-economic burden.
Ongoing monitoring and assessment of seasonal influenza are therefore essential to ensure early warning of epidemics and tailored preventive and control measures in real-time. The last 2009 pandemic revealed many deficiencies in most countries' influenza surveillance systems, especially the capacity to estimate the severity of the season in a timely manner. For this reason, the WHO has progressively developed a framework on pandemic influenza severity assessment (PISA) and recommended member states to apply the proposed tools and measures. 2 The framework is based on different steps, including setting thresholds for selected parameters and applying them in the routine surveillance of seasonal epidemics.
Various mathematical and statistical models have been developed to establish thresholds for influenza activity and study the dynamics of the disease. [3][4][5] This mathematical modeling provides valuable information and a strong support to the preparedness and response plan. Of the popular methods currently in use, the moving epidemic method (MEM) is one of the most recommended and so far had provided a robust signal to detect influenza epidemics in many countries. 6,7 First developed in Spain in 2001, the MEM was adopted by the same authors to determine influenza thresholds in many European countries. 8 One of its strengths is its ability to also define different intensity levels in a given region or country and the possibility to compare them between countries and/or seasons. 9,10 In Tunisia, influenza surveillance was first based on the virological surveillance ensured by the National Influenza Centre (NIC) recognized by the WHO since 1980 and supported by the Primary Health Care Direction of Ministry of Health. Starting from the late 1990s, the epidemiological surveillance was established through the network of Influenza-like illness (ILI) sentinel sites at primary healthcare centers in the 24 governorates of the country. This network was progressively improved mainly by reducing the number from 268 in 1999 to 113 ILI sites in 2014, better representativeness and training of all staff involved in the surveillance. 11,12 Each year, the proportion of ILI among the total number of consultations at ILI sentinel sites determines the intensity of influenza season. The epidemic threshold of 10% adopted since then was based on combination of criteria and a national approach. 13,14 Given the importance of seasonality and intensity levels in influenza severity assessment, our study aimed to determine the epidemic and intensity thresholds of influenza season in Tunisia by applying the moving epidemic method (MEM) based on ILI historical surveillance data of the last 9 years (2009-2018).

| Available data
Influenza surveillance in Tunisia is carried out each year from 1st October (week 40) to 30th April (week 18) over a period of 30 weeks.
Data collection is based on standardized forms of weekly aggregated data of ILI cases. These paper forms are sent from ILI sites at the local level to the regional directions in each governorate then to the Primary Health Care Direction at the national level. Aggregated data forms consist of general information including ILI site, governorate, the number of ILI cases and the total number of outpatients by gender and age groups (0-5 years; 6-16 years and ≥16 years). In Tunisia, case definition of ILI was an outpatient with fever (≥38°C) and cough or sore throat with onset less than 5 days prior to presentation in the absence of a specific diagnosis. 11 Since 2014, the case definition recommended by WHO has been used instead: acute respiratory illness, and measured fever ≥38°C, and cough, and onset in previous 10 days. 12 Collected data are analyzed to compute weekly ILI proportions compared with all outpatient acute consultations at both the national and regional levels. We analyzed data from up to nine influenza seasons (2009-10 to 2017-18).

| Moving epidemic method
We applied the moving epidemic method (MEM) to establish epidemic and intensity thresholds, based on previous publications and the WHO's interim guidance for influenza severity assessment. 2,8,9 For that, we used the R Language implementation of MEM (package "mem") which is available online for free. 8,15 This method, based on a complex mathematical algorithm, can be summarized in three steps. First, determine the start, scope and end of the influenza epidemics by dividing the season in three periods

| Cross-validation procedure of the model
The goodness of the model was assessed using a cross validation procedure. This procedure is based on the extraction of each season from the historical series and using it as "a target season", for which we calculate the beginning and end of the epidemic period.
Subsequently, the pre-and post-epidemic thresholds are calculated on the basis of the remaining seasons and excluding the target season. These steps were repeated for all the available seasons.
Values of the target season inside and outside of the defined epidemic period were compared to the thresholds calculated using all historical information but the target season.
Aiming to evaluate the performance of the epidemic threshold to detect epidemics, we studied the sensitivity (Se), specificity To optimize the goodness of the model, we also looked at the optimum slope parameter to find the value that maximizes the sensitivity and specificity. It is an inner parameter ranging from 2% to 4%. 8

| Descriptive analysis of the epidemic movement of influenza in Tunisia
The beginning, the end and the extent of the epidemic seasons differed from one season to another (Table 1).  Except the pandemic season which was considerate of a very high intensity, most of the seasons were described as low. This was useful to characterize the dynamics of influenza over time.

| Epidemic and intensity thresholds
The different thresholds and intensity levels were determined using two models; one including and the other excluding the 2009-10 pandemic season.
When including the pandemic season, the average duration of influenza epidemic during a typical season was 20 weeks (Figures 2A   and 3A). This optimal duration of 20 weeks covers 76.61% of total sum of proportions.  Table 3  This validation also allowed us to characterize the overall intensity of each season. Out of the nine seasons, six were low, two as medium, and one very high intensity ( Table 3).

| Cross validation of the model
The MEM provided a sensitivity of 85% in detecting the epidemic period. This sensitivity during the overall seasons and for   This choice was largely motivated by the type of data available by the Tunisian influenza surveillance system. In fact, the basic requirements of the MEM are simple and reliable epidemiological data for a time period between 5 and 10 years, preferably ILI data. 8 The MEM is a tool developed to better understand annual influenza epidemics and assess the epidemic status and intensity on a weekly basis. 8,9 The method was progressively improved and implemented in European documents by the ECDC and WHO. 15 Later on, it became widely used in many countries outside Europe such as USA, Australia, New Zealand, and Canada. 6,10,16,17 Other countries have opted for the method proposed by WHO and based on the peak mean values of influenza activity. 5,18,19 The determined epidemic threshold with the exclusion of the pandemic season was 6.25%. It showed a very high sensitivity (85%) and a high specificity (69%). However, when including 2009-10, the threshold increased to 8.99% with a sensitivity and specificity of 39% and 87%, respectively. The different levels of intensity were also affected with a considerable increase. This is understandable since there was higher ILI rates registered during this year and thus It would therefore be useful to establish one common method for ILI data analysis and interpretation in our region as was done in Europe. 9,22 Applying the MEM to define the thresholds also allowed us to vi- countries and most regions of the Northern Hemisphere sharing the same winter timing. [24][25][26] The specificity of the determined epidemic threshold was lower than its sensitivity. Sensitivity is important for detecting epidemics but specificity is crucial to avoid false alerts. In fact, once an epidemic is declared, the media's interest increases and prevention and control measures are implemented, especially vaccination campaigns and antiviral use. 8 That is why it is important to avoid false alerts and to use these attributes wisely trying to find the good balance between specificity and sensitivity.

| D ISCUSS I ON
Besides, it is important to underline that the specificity of the model is related to the case definition used. The lower is the specificity of the case definition, the lower is the specificity of the model.
Although a new case definition was used since 2014, the changes enhanced sensitivity without greatly compromising the specificity. 27 We therefore consider that the specificity of the model did In these situations, additional virological data are necessary to confirm the start of the epidemic period, especially that our results showed a better performance of the MEM model excluding the pandemic season than the one including this season.

| CON CLUS ION
In summary, the moving epidemic method is a simple method offering a flexible procedure to calculate epidemic thresholds based on historical epidemiological data. Its strength lies in its ability to also determine different intensity thresholds useful to the weekly monitoring of the season's intensity.
Our study is the first mathematically based study of seasonal threshold of influenza in Tunisia using historical ILI weekly data.
The determined epidemic threshold was 6.25%, differing from the threshold of 10% adopted until now. The high sensitivity and specificity of this threshold in the detection of epidemics make it robust and reliable.
Indicating the start and assessing the intensity of influenza seasons remain a high priority for Ministries of Health, not only at the national level for timely preventive and control measures but also at the international level by contributing to the pandemic preparedness.
The next step is therefore to implement the use of the determined epidemic threshold for public health purposes with monitoring the next seasons.