Correcting for day of the week and public holiday effects: improving a national daily syndromic surveillance service for detecting public health threats

Syndromic surveillance is the near real-time collection, analysis, interpretation, and dissemination of health related data to enable the early identification of the impact of potential public health threats [1]. The real-time syndromic surveillance team at Public Health England (PHE) co-ordinates a suite of syndromic surveillance systems in order to provide early warning of outbreaks of infectious disease, situational awareness during a public health incident, and reassurance of lack of impact [2‐5]. These syndromic surveillance systems are used to complement and support existing public health surveillance programmes.

Line graphs of time series data offer a simple and effective way to review data and undertake exploratory analysis [6, 7]. They are used, in addition to automated statistical alarms, by the real-time syndromic surveillance team to investigate, interpret, and present the current trends in syndromic data and for comparisons of the current data with previous years to identify changes from the norm. Regular, large fluctuations at small time-scales can, however, make it difficult to identify longer time-period trends in time series graphs. These difficulties can be overcome by adding to the graph a smooth trend curve which takes into account these known day-to-day fluctuations [8].

The GP in-hours syndromic surveillance system (GP in-hours SSS) monitors the number of in-hours family doctor (known as general practitioner, or GP, in the UK) consultations [9]. Daily data on the number of GP consultations are analysed, and are aggregated into syndromic indicators based on symptoms and clinical diagnoses (e.g. influenza-like illness, diarrhoea, chickenpox) [9]. Although much of the GP in-hours SSS is automated, statistical alarms are created that require manual, in-depth investigation [10]. Effective data visualisations must be used in order for the manual investigation stage not to become the bottleneck of the real-time data analysis process [11].

Graphs of the syndromic indicators from the GP in-hours SSS are presented to the public and wider audiences in weekly bulletins published by PHE [12]. This is an additional reason to ensure that the current trend in illness levels can be clearly interpreted from the graph without additional data or expert knowledge.

Regular fluctuations at a weekly time-scale, known as day of the week effects, have been observed in the number of patient consultations with GP services [10]. The number of consultations is also observed to regularly change on a public holiday and on the days immediately after [10]. We refer to this as a public holiday effect.

The purpose of syndromic surveillance is to identify abnormally elevated disease levels as early as possible so that action can be taken to minimise the problem [13, 14]. However, if the systematic changes in the number of consultations with GPs due to day of the week and public holidays are not accounted for, they could mask real increases in disease levels, create false alarms, and delay decision making over public holiday periods as more data are required to understand the current trend. It is important to try to distinguish the expected changes in consultation numbers due to day of the week or public holiday effects from unexpected changes due to potential public health threats.

The purpose of this work is to develop and explore an appropriate smoothing method that takes the expected day of the week and public holiday effects into account simultaneously and displays no trend due to these predictable variations. This method will be applied to time series graphs to enhance visual analysis of daily GP consultation data for syndromic surveillance. This will improve daily risk assessments by epidemiological investigators.

Data from healthcare services reflect the time at which patients sought healthcare advice. This does not necessarily correspond with date of symptom onset. In particular, patients with milder illnesses may not present unless they become more severe or complications develop [15, 16]. Therefore, the number of healthcare consultations is not a simple measure of illness in the population but rather a combination of illness levels, severity of the illness, availability of healthcare services, and ability or willingness to seek healthcare [17]. Based on this, we develop a data-driven smoothing method, the extended working day moving average, using scaling factors to take both day of the week and public holiday effects into account.

The rest of this paper is organised as follows. The Background will conclude with a short discussion of the existing literature of smoothing methods to account for day of the week and public holiday effects in healthcare data, a description of the specific calendar effects observed in the GP in-hours SSS, and a description of the seven-day and working day moving average. The limitations of these methods justify the development of the extended working day moving average to take day of the week and public holiday effects into account simultaneously, which will be described in the Methods section. This will be followed by a description of the data from the GP in-hours SSS to which the smoothing methods will be applied. An evaluation of the extended working day moving average, with comparison to the seven-day and working day moving averages will be presented in the Results section. Finally, the strengths and limitations of the smoothing methods and the impact of using the extended working day moving average on public health practice will be discussed.

As previously described, the extended working day moving average was applied to synthetic data and the severe asthma and gastroenteritis syndromic indicators from the GP in-hours SSS. The 7-day and working day moving averages were also applied for comparison.

Using the percentages 2%, 18%, and 24% described in the Data section, the scaling factors for the extended working day moving average applied to the synthetic data were calculated as 0.1 for weekends and public holidays, 1.2 for typical Mondays and the first working day after a public holiday, and 0.9 for all other typical weekdays. The scaling factors calculated from the severe asthma and gastroenteritis indicator data are given in Table 1.

The extended working day moving average showed a no-trend line when applied to the synthetic data, as the combination of day of the week and public holiday effects were taken into account (Fig. 2). The extended working day moving average also continued to display the trends in the syndromic data throughout public holiday periods (Fig. 3).

In the absence of public holidays, the seven-day moving average applied to the synthetic data smoothed the regular day of the week effect to highlight the current trend. However, there is a dip in the smoothing trend curve for 7 days around the public holiday (Fig. 2). These synthetic data followed the expected behaviour of no-trend syndromic data around a public holiday. With real data, this dip in the smoothing curve could mask an actual increase in disease levels over this time period. However, this change is entirely expected due to the change in healthcare service provision on public holidays. Additionally, the 7-day moving average was lower than the average number of consultations on a working day. It is more useful that the smooth trend curve gives an indication of the number of healthcare contacts on a typical working day.

These same results were also observed when the 7-day moving average was applied to surveillance data for the severe asthma and gastroenteritis indicators (Fig. 3).

The working day moving average applied to synthetic data gave a better smooth curve than the 7-day moving average (Fig. 2). However, a drop 3 days before and a peak 4 days after public holidays were still present in the smoothing curve when applied to both synthetic and real data (Figs. 2 and 3). These were due to the combination of the day of the week and public holiday effects. The drop was caused by that 7-day sum not including a typical Monday, and the peak was caused by that 7-day sum including both a typical Monday and the elevated Tuesday directly after the public holiday.

In the absence of big day of the week effects, the working day moving average would smooth a simple public holiday effect. However, the interaction between day of the week and public holiday effects, and extended holiday effects such as a change in activity on the first working day after a public holiday, are not accounted for.

Smoothing trend curves are used to help investigators visually identify current unusual activity during daily surveillance of syndromic disease data. It is easy to retrospectively look at the smoothing curve given by the working day moving average and identify the spikes as clearly spurious due to their short duration. However, in order to emphasise how misleading the 7-day and working day moving averages can be we applied all the smoothing methods to the dataset that would be available a week after a Monday public holiday. This graph would be used to assess the current trend in the number of severe asthma consultations (Fig. 4). The trend 1 week after a public holiday would be noted as increasing if either the 7-day or working day moving averages were used. This could lead to unnecessary alarm. The extended working day moving average did not show an increasing trend and, more importantly, neither did the data. The extended working day moving average would make it easier for investigators to identify unusual activity during this period.

It is widely acknowledged that day of the week and public holiday effects exist in healthcare data used for syndromic surveillance and that this can disguise anomalies in the data when visually inspecting it [10, 17‐23]. In this study, we described the previous smoothing method used by PHE to smooth data from the GP in-hours SSS. We also developed a smoothing method where both day of the week and public holiday effects are taken into account simultaneously. We demonstrated how the extended working day moving average can be used to aid interpretation of the trends in real-time syndromic surveillance data from GP services, thereby improving the public health action resulting from the analysis. The extended working day moving average method retains the ability to display unusual changes in the trends of syndromic indicators from the GP in-hours SSS during public holiday periods, and it removes the potentially misleading spikes observed in the working day moving average. This reduces the potential for delays in the detection of public health threats during this time.

The inter-quartile ranges of the proportion of consultations on each day of the week are quite narrow (Fig. 1). This indicates that the day of the week effect is consistent throughout the year. However, day of the week and public holiday effects are just one cause of noise in these complex data sets. The number of GP consultations fluctuates and contains regular trends due to other factors that we do not discuss or control for here. These include, for example, seasonal disease outbreaks and changes in the data collection systems.

In this study only relatively simple data-driven smoothing methods were considered. Syndromic surveillance uses large, varied data sets, and it is desirable for syndromic surveillance reporting systems to be as automated as possible. A simple data-driven smoothing approach ensures sufficient flexibility so that smoothing methods can be applied to a wide range of indicators in an automated way [25]. As discussed in the Background, data-driven smoothing methods have previously been used to remove day of the week and/or public holiday effects from daily syndromic data as part of more complex detection algorithms [17, 20, 26, 27]. However, this study shows that both day of the week and public holiday effects must be considered simultaneously to create adequately smooth daily healthcare data. We have addressed this problem in the context of GP in-hours consultation data used for daily syndromic surveillance in England, and we have focused on methods to improve time series graphs used for daily risk assessments by investigators.

The extended working day moving average was developed for the GP in-hours SSS coordinated by PHE. We demonstrated the method applied to the gastroenteritis and severe asthma indicators as examples. However, the day of the week and public holiday effects observed in these two indicators are also observed across the GP in-hours SSS in a consistent way (see, for example, the plots of data for a large number of indicators within the PHE weekly bulletin [12]). It is therefore appropriate and straightforward to apply the method to other syndromic indicators from the GP in-hours SSS, and we see the same results as discussed here. As a result of this, the extended working day moving average is now in use across the GP in-hours SSS.

Day of the week or public holiday effects are also seen in attendance data from many other healthcare services. This includes emergency departments [28], walk-in clinics [29], military treatment facilities [15], sexual health clinics [30], telehealth services [5], and internet based symptom-checker services [31]. It is also seen in the other syndromic surveillance systems operated by PHE. This work has demonstrated the importance of being aware of day of the week and public holiday effects in analysis and interpretation of this type of data, including the effect on days near to the public holiday itself. We have shown how an inadequate treatment of these effects can lead to potential confusion in the current trend and delay decision making.

However, the extended working day moving average described here was developed for use with just one particular syndromic surveillance system. Further work is needed to investigate whether the extended working day moving average could be applied to other surveillance systems. In particular, whether it is valid for those which monitor attendances at 7-day healthcare services. Additionally, if the day of the week and public holiday effects are not as large as those observed in the GP in-hours SSS a simpler method could be sufficient. Further work in this area will describe the extent of the day of the week and public holiday effects across different syndromic surveillance systems. This will also involve an investigation of the public health aspects of these effects, rather than purely the statistical approaches considered during this analysis.

The main limitation of the extended working day moving average is that historical data are needed to compute the scaling factors. In particular, sufficient data are required to learn how the number of consultations changes around each public holiday. On the other hand, the working day moving average and 7-day moving average do not require historical data and therefore can be used immediately with new syndromic surveillance systems.

Our results show that basic smoothing techniques are not able to account fully for the public holiday effects observed in the GP in-hours SSS. We have developed and demonstrated an improved smoothing technique that can make it easier for investigators to identify unusual activity during daily surveillance of syndromic GP data. This method is now in use in the GP in-hours SSS at PHE. It has led to enhanced visualisations of this data during the analysis phase and in weekly public health bulletins [12].

Based on this study, it is recommended that analysis and visualisation methods for syndromic data carefully take both day of the week and public holiday effects into account.