Monitoring sick leave data for early detection of influenza outbreaks

Early outbreak detection is crucial for preparedness and timely public health and medical responses. It provides useful information to physicians, companies, and the public, ensuring proper drug prescription, health service planning, workplace preparedness, and continuity of operations in case of high absenteeism [1], among many other uses.

Most countries face periodic influenza (or “flu”) epidemics that vary in size and severity from year to year [2]. Seasonal flu can be highly virulent and, like many respiratory viruses, can spread rapidly through populations highlighting a need for a robust epidemiological surveillance system to detect emerging outbreaks. Surveillance system guidelines developed by the US Centers for Disease Control and Prevention (CDC) suggest that systems should be simple, reliable, flexible, timely, and readily accepted by diverse individuals and organizations to ensure participation [3].

Flu surveillance systems vary by country and rely on various types of data. National health agencies monitor flu epidemics using healthcare records, medical sentinel systems, pharmaceutical sales and other data sources. Most of these systems rely on data from healthcare settings that rely on patient healthcare seeking behavior or after results of clinical tests, which often reflects those with symptoms or relatively advanced stages of the disease. These systems fail to capture individuals who do not seek medical care, whether due to asymptomatic infection, perceived mildness of infection or a general reluctance to seek care [1, 4]. Given these gaps, alternative data streams from non-healthcare settings may provide a valuable complement to classical surveillance systems [5].

Human resources data collected at the workplace have received relatively little attention for outbreak detection, but present characteristics that are useful for infectious disease surveillance. In many settings, absenteeism data are routinely collected and centralized either for the use by companies themselves or for health insurance purposes. Due to legal purposes and to the implication of work absence on salaries, these data are comprehensive and reliable. Though a handful of studies have assessed the role of sick leave data for outbreak detection, they did not develop a comprehensive assessment of its performance and robustness. Bollaerts et al., Patterson et al. and Groenewold et al. [1, 6, 7] assessed the usefulness of work absenteeism surveillance as a tool for early warning systems for influenza. Work absenteeism study can also supplement more traditional medical data by providing information about an epidemic’s socioeconomic impact [1, 4]. As a consequence, the US National Institute for Occupational Safety and Health (NIOSH) has been monitoring health-related workplace absenteeism among full-time workers using data received monthly from the Current Population Survey since 2017 and making this data available online [8].

A great challenge in epidemiological surveillance lies in identifying data streams that allow for sensitive, specific and timely outbreak detection. In the context of outbreak detection, surveillance sensitivity can refer to both (i) the proportion of true cases detected, and (ii) the probability of detecting an outbreak, including the changes in the number of cases over time [3]. Surveillance specificity refers to the probability of correctly identifying when an outbreak is not occurring [9]. Lastly, surveillance timeliness generally refers to the time difference between an event and its standard reference [5]. Some studies have suggested that absenteeism data along with others such as over-the-counter pharmaceutical sales and emergency visits seem to be more timely than sentinel Influenza-Like Illness (ILI) surveillance [5, 6], other traditional flu data sources [7], physician diagnoses [5], and virological data [5].

In this study, we assess the sensitivity, positive predictive value and timeliness of workplace absenteeism data for detection of flu outbreaks in France. Our hypothesis is that monitoring sick-leave data at the workplace might help anticipate outbreaks in a timely manner using routinely collected data. We then compare the performance of a sick leave based monitoring system to the performance of the national standard surveillance system of influenza in France which is based on ILI data.

Incidence curves obtained from the Sentinelles surveillance networks from 2010 to 2017 reveal annual peaks of influenza-like illness (ILI), from 163 to 1290 per 100,000 per week, occurring approximately between December and February (Figs. 1 and S1). During the summers, the incidence approaches zero. By comparison, weekly sick leave incidence varied about an annual average of 1021 to 1335 per 100,000 per week, depending on the region (Figs. 1 and S1). They exhibit greater variability and more peaks per year than ILI incidence. However, based on visual inspection, the highest seasonal peaks tend to coincide with ILI incidence peaks. Moreover, most of the seasonal sick-leave troughs coincide with Christmas and summer school holidays periods, which can be explained by a decrease of the at-risk population, i.e. an increase in the number of workers on paid leave. Finally, there is no apparent change between 2014 and 2015 despite the strong variation in the volume of workers in the database.

For three French regions, Fig. 2 presents the incidence of sick leaves and ILI for the 2015–2017 time period. In each region, the Sentinelles surveillance system identified exactly one alert per year, triggered a few weeks before or during the peak of ILI incidence (Figs. 2 and S2). The exception was Bretagne, where no alert was identified during winter 2016–2017 (Figure S2). By comparison, the sick leave surveillance system triggered one to three alert episodes per year.

We assessed the sick leave surveillance system on its ability to detect and anticipate ILI incidence peaks. Table 1 summarizes the indicators of its performance regarding ILI peak detection and anticipation in each region, averaged over the 2 years of the model test. The sensitivity per episode (probability of detection of the ILI outbreak) had a mean of 0.92 (range 0.5–1) across regions, while the positive predictive value per episode had an average of 0.58 (range 0.2–1). The sick leave alert generally occurred prior to the peak ILI incidence, on average 5.88 weeks (range 2.5–11) before the peak.

We also compared the performance of our sick leave model with the Sentinelles surveillance system. The results can be read in Table S1 (Supplementary Material). The main result is that the sick leave model alert was always triggered earlier than the Sentinelles model when both alerts match, with an average lead time of 2.5 weeks (range: 0.5–4).

Workplace absenteeism data can be used by public health surveillance systems to detect emerging infectious disease epidemics [16]. Despite this, to this date, few health authorities worldwide use absenteeism data to inform outbreak surveillance. Here, we assessed the potential of workplace absenteeism data to monitor influenza and detect epidemics in France, using an adapted statistical method to analyze this data. We applied this method to a comprehensive national database of workplace absenteeism and validated it against the French national surveillance system based on sentinel GPs. Our results suggest that a system based on workplace absenteeism could be highly sensitive and detect influenza epidemics earlier than the current French surveillance system.

We found that the surveillance system we propose would be able to detect outbreaks 5.9 weeks before the peak and about 2.5 weeks before the Sentinelles system. This suggests that sick-leave data could be almost as timely as emergency visits data that has a timeliness of 3 weeks compared to ILI data systems [17]. Our findings in this regard are in line with previously published studies in several contexts worldwide. In a French study from 1994, sick-leave data from a large company allowed to detect flu epidemics with up to 2 weeks of advance [9]. In a more recent Belgian study, worker absenteeism data from the Belgian Medical Expertise and from the Belgian railway system was shown to start rising 2–3 weeks in advance and to peak 2 weeks in advance, as compared with ILI data from the Belgian sentinel GP surveillance system [6]. In the UK, data on workplace absenteeism among employees of Transport for London peaked up to 2 weeks before the NHS ILI surveillance data [7]; and monitoring workplace absence due to “cold”, “cough” or “influenza” among the staff of a large hospital organization was shown to allow the detection of flu epidemics with a significant advance of up to 9 weeks [18].

Very few of these previously published studies included an assessment of the sensitivity, specificity or positive predictive value of an absenteeism-based surveillance system. However, in the French study from 1994, the sensitivity and specificity of surveillance based on sick-leave data from a large company were estimated at 74 and 67% for the identification of epidemic weeks (Youden index: 0.42) and 67 and 94% for the detection of epidemics (Youden index: 0.61), with an 80% positive predictive value [9]. The UK study based on hospital staff absenteeism also noted that the resulting system did not lead to more false positives than the NHS surveillance data in London.

The quality of the developed model depends strongly on the quality of the data collected within companies. Sick-leave data have the advantage of describing quasi-real individual behavior regarding sick-leave and presence at the workplace. These data are in fact used to enter employees’ pay and are subsequently fulfilled by obligation in the computer system. Our data may not be representative of French population because it describes data from a health-insurer. For instance, it insures few construction companies because they have their own specific insurer. The data also do not include unemployed people. However, representativeness is not necessarily required to build up an outbreak detection system. In fact, outbreak detection aims to detect any unusual, expected number of cases to generate signal alarms. Similarly, the GP Sentinelles network does not include all GPs but a small subset of the same practitioners over time.

Another downside of our data is related to the definition of the sick leave rate. The denominator of this rate is the number of workers and it includes workers that are on holiday. The sick leave rate therefore drops during school holidays and the systems do not detect any alerts during those holidays even if they are included in the statistical model as covariates. This is an issue if the epidemic occurs during the holidays and this is the case during the year 2016–2018 where the peak occurs during the Christmas break for two regions. The model could then be more sensitive if the denominator was the number of employees actually at work.

Furthermore, the model accuracy estimated by the false positive rate is strongly related to the definition of cases to detect. The time series of cases is based only on flu cases. However, other epidemics such as gastroenteritis can influence sick leaves data and could be considered for future works: the sick leaves outbreaks may correspond to other disease and may explain the poor positive predictive value in some cases. The detection algorithm could also be improved if the estimation of expected cases could adjust for any potential past exogenous environmental factors such as a terrorist attack, strikes or unexpected bad weather episodes. The model accuracy is moreover strongly related to the method chosen for the algorithm. The Serfling method may actually not be the more accurate model and was chosen to be consistent with the Sentinelles method. Some other regression-based models are known to be more specific, like the Farrington algorithm [19, 20].

Another limitation of our study is linked to the fact that we relied on data that were consolidated on an annual basis, and not in real-time. As such, the system is able to detect outbreak based on the dates of sick-leaves, but only retrospectively. A proper integration of sick-leave data into a health surveillance system would thus require an effort to ensure sick-leave data are consolidated and made available in real time. Our results suggest that the resulting increased timeliness of a surveillance system including this data stream may justify this effort. Moreover, these data sometimes already exist in near-real time because of local legislation. For instance, French companies must declare sick leave within 5 days.

Sick leave could then be an alternative source of data for the detection of sick leave. Many other data sources have already been used for near real-time prediction of influenza outbreaks: GP networks of course, but also Google queries with Google flu [21], social network database [22], large epidemiological databases [23] and, more recently, wastewater have been used for COVID-19 monitoring [24]. Sick leave data is however more robust than Google (and probably social network) data, which has shown its limitations. Google queries for relevant keywords are indeed not only correlated to the epidemic but is also correlated with Google’s practices: the data generation process changes over time and Google sometimes recommends certain keywords for commercial purposes [25]. Moreover, large epidemiological database may be more sensitive and specific but are much more complex to handle. Sick leave data can be simple to retrieve, easy to analyze and most of all very timely compared to the data usually used. These data could thus enable near real-time prediction, which would allow for reactive monitoring of influenza outbreak.

Many of the previously published studies assessing the potential of workplace absenteeism data for flu surveillance simply provided visual analyses comparing and correlating absenteeism data with ILI surveillance data [6, 7]. By contrast, in this work, we propose a statistical approach and algorithm to analyze the French workplace absenteeism data, and to raise alarms when outbreaks are detected. This allows us to both propose a complete surveillance system that could be used in practice provided the data is available, and to fully assess the performance of this surveillance system.