Introduction

The present study describes methods to inform operational decision making and response planning in the areas of syndromic surveillance and outbreak detection, as well as daily situational intelligence relating to demand and hospital capacity. Syndromic surveillance has a role in the early detection of disease outbreaks, which then enables management strategies such as prompt isolation of patients, strict enforcement of quarantine of all contacts and top-down enforcement of community quarantine.

Knowledge of the timing of any disease outbreak is vital in public health service delivery and early warnings of outbreaks are paramount.1, 2 Monitoring the fluctuations and trends of syndromic surveillance data supports policy-related decision making.3, 4 Syndromic surveillance can also assist with day-to-day hospital capacity management and operations.

In the present study, we use the time between ED presentations to develop novel outbreak detection models to complement wider response efforts of government agencies.

Methods Design

The study involved secondary data analysis of ED presentations to major public hospitals in Queensland and South Australia. Exponentially Weighted Moving Averages (EWMAs) of the time between ED presentations were used to construct adaptive monitoring plans to detect outbreaks.

Sample and setting

For Queensland, the data spanned 1 January 2017–31 December 2020 and covered 27 major public hospitals. For South Australia, the data spanned from 1 January 2017 to 12 June 2020 and covered all 14 public metropolitan and major country hospitals.

Measures

Outcome measures of the study were outbreak detection models for Queensland and South Australian EDs.

Although globally topical, coronavirus counts across the study sites decreased after an initial first wave because of social distancing policy and other interventions, and ‘influenza-like illness’ was used as an initial focus for developing outbreak detection models to pick up future aberrances and geographic outbreaks quickly. Influenza-like illness surveillance data can be used to estimate the prevalence of COVID-19 and to inform about the scale of upcoming surges in hospital demand.5

Internationally the World Health Organization defines influenza-like illness as measured fever of ≥38°C and cough with onset within the last 10 days,6 both of which cannot be identifiable from ED clinical information systems. However, using the primary diagnosis field within ED datasets is a reasonable measure for surveillance7 and often includes subsets of diseases of the respiratory system defined by ICD10 chapters J00-J99 as well as some generic infection codes.

The list of influenza-like illness diagnoses codes used for the present study was based on previous use8, 9 but broadened to the codes shown in Table 1. Although these diagnoses cover a broad set of conditions (low specificity), the list was arrived at via consensus from a clinician-led 16-person Syndromic Surveillance Working Group comprising health disaster management staff, emergency physicians (FACEMs), Nursing Directors, ambulance personnel and the authors. Counts and seasonal patterns of the chosen ED diagnosis codes were checked against laboratory-confirmed cases of influenza and found to be consistent. For the Queensland analysis, the definition was expanded to include six additional codes specifically related to COVID-19 as noted in the table. Updated ICD coding guidelines make reference to a new diagnosis code established by the World Health Organization for COVID-19 (‘U07.1’), which is likely to be implemented in the next ICD update.10 Use of this code is growing and its inclusion in syndromic surveillance and future epidemiological studies is acknowledged.

TABLE 1. Primary diagnoses codes used to define influenza-like illness and develop outbreak detection models ICD10 primary diagnosis code Diagnosis description A08.4 Viral intestinal infection, unspecified B34.2† Coronavirus infection, unspecified site B34.9 Viral infection, unspecified J02.9 Acute pharyngitis, unspecified J06.9 Acute upper respiratory infection, unspecified J10.8 Influenza with other manifestations, other influenza virus identified J11.1 Influenza with other respiratory manifestations, virus not identified J18.0 Bronchopneumonia, unspecified J18.1 Lobar pneumonia, unspecified J18.8 Other pneumonia, organism unspecified J18.9 Pneumonia, unspecified J22 Unspecified acute lower respiratory infection J44.9 Chronic obstructive pulmonary disease, unspecified U04.9† Severe acute respiratory syndrome (SARS), unspecified U06.0† Emergency use of U06.0 U06.9† Emergency use of U06.9 U07.1† Emergency use of U07.1 (WHO: “COVID-19, virus identified”) U07.2† Emergency use of U07.1 (WHO: “COVID-19, virus not identified”) Data analysis

The traditional approach of monitoring disease outbreaks is based on daily or weekly counts of disease and flagging an outbreak when these counts are significantly higher than expected.11-15 However, recently it has been demonstrated that for early detection of disease outbreaks, it is much more efficient to monitor the time between events (TBEs) than to use daily counts.16, 17 The efficiency gain with monitoring TBE is that every individual instance (ED presentation) offers a decision point about a disease outbreak, and we do not have to wait for the end of the day or week to count the overall disease counts to make that decision. When the TBE values for a particular disease get smaller and smaller, then the likelihood of an unusual outbreak for that disease increases.

TBE are computed for each hospital for influenza-like illness ED presentations and EWMAs of the TBE16, 17 with different levels of temporal memory are calculated. These levels of memory relate to adopting different model coefficients resulting in differences in the ability of the model to correctly signal an outbreak and can be optimised depending on whether daily counts of the monitored signal are low (e.g. daily counts less than 25 per day) or high. This allows the plan to be robust at detecting outbreaks of different sizes.

We fit a Weibull regression model for expected behaviour of TBE for any time of the day during the year, using time within day (0–24 h), time (day number) and day of the week. A comparison of actual TBE values to expected values enables assessment of whether the actual values are persistently lower than expected to flag an outbreak. This is fitted using the gamlss library in the statistical programming language R. These models are used to predict the scale and shape parameters of the Weibull distribution that is appropriate for the data. For TBE, an outbreak is flagged whenever the EWMA smoothed TBEs value is less than a threshold.

Further detail on the TBE surveillance approach is presented within the Appendix S1.

An exemption from ethical review for this analysis was granted by the SA Department for Health and Wellbeing Human Research Ethics Committee (ref: REC/20/SAH/35) and the Royal Brisbane and Women's Hospital Human Research Ethics Committee (ref: LNR/2020/QRBW/66608). Exemption was on the basis that the analysis was negligible risk, involved the use of existing collections of data records that contained only non-identifiable data about human beings, there were no participant risks, burdens, inconveniences or the potential for breach of privacy, and the output supported preparedness and service delivery.

Results

Coronavirus counts across the study sites decreased after a first wave of the outbreak and as shown in Figure 1, the COVID-19-related codes including the specific diagnosis code for coronavirus (B34.2) only appeared from February 2020 within the study data and formed a tiny blip against the broader list of influenza-like illness diagnosis codes. When assessing influenza-like illness diagnoses across the study period, unspecified viral infection (B34.9) was the most prevalent diagnosis (27% of total counts in South Australia, 41% of total counts in Queensland). The specific codes for influenza (J11.1) and COVID-19 related codes accounted for only a small fraction of influenza-like illness cases over the study period (influenza: 5% in South Australia, 3% in Queensland; COVID-19: 0.2% in South Australia, 2% in Queensland). Including the broader list of codes for influenza-like illness gives a larger signal to monitor against compared to the specific diagnosis codes for influenza or coronavirus (Fig.  1).

Influenza-like illness (ILI) diagnosis codes compared to the specific codes for influenza (J11.1) and coronavirus codes across the study period. Top: South Australia hospitals: (), ILI grouping of diagnosis codes; (), ‘J11.1’ specific code for influenza; (), ‘B34.2’ specific code for coronavirus. Bottom: Queensland hospitals: (), ILI grouping of diagnosis codes; (), ‘J11.1’ specific code for influenza; (), COVID-19 grouping of diagnosis codes (B34.2, U04.9, U06.0, U06.9, U07.1, U07.2).

The Appendix S1 presents adaptive charts for monitoring TBE of influenza-like illness ED presentations for most of the major public hospitals in Queensland and South Australia across the study period (excluding two facilities at the request of SA Health). The charts indicate variation in outbreak detection from one hospital to the next.

The adaptive chart for influenza-like illness for the South Australian public hospitals pooled together (covering 12 of the 14 public metropolitan and major country hospitals representing close to a state-wide view) is presented at the top of Figure 2. An outbreak is flagged when the EWMA statistic drops below the red threshold. It can be seen that 2017 and 2019 had high counts of outbreaks, 2018 was quiet, and 2019 started much earlier in the year rather than traditional winter. After each flagged outbreak, the TBE is reset to its initial value – so each consecutive signal indicates it remains an unusually high outbreak. The large outbreak of influenza-like illness around March/April 2020 is evident, whereas the following months that usually mark the onset of the southern hemisphere flu season were quiet.

Top: Monitoring time between events of influenza-like illness ED presentations pooled across major South Australia public hospitals. Values of the plotted test statistic less than the red threshold indicate an outbreak. Bottom: Extract from weekly influenza notification chart in South Australia. Reproduced with permission from Communicable Disease Control Branch, South Australia. (), Influenza B; (), influenza A; (), sentinel general practitioner influenza diagnoses (ASPREN); (), ED influenza diagnoses (HASSED).

There is high consistency between the adaptive influenza-like illness chart when hospitals are pooled together and confirmed influenza notifications sourced from the Communicable Disease Control Branch of SA Health,18 reproduced with permission at the bottom of Figure 2. Both information sources reflect the large outbreak years of 2017 and 2019 and the fact that the flu season in 2019 started earlier in the year.

Discussion

Numerous methods have been suggested to detect an aberrance/departure from usual levels of diseases such as influenza-like illness. The US CDC have published their method for detecting ‘higher than expected’ activity:11 they first calculate a seasonal baseline using robust regression from the previous 5 years, and an increase of 1.645 standard deviations above the seasonal baseline (90% confidence interval) is considered the epidemic threshold. Baselines and thresholds are calculated at the national and regional levels and by age group. CDC also produce web interactives of influenza-like illness and make these data available at a weekly state level for US states. These weekly count data have recently been used to estimate the prevalence of SARS-CoV-2, which highlights the potential to use syndromic surveillance for early detection and understanding of emerging infectious diseases.5, 12

New Zealand's Institute of Environmental Science and Research (ESR) has developed a near real-time flu surveillance system, tracking acute respiratory illness ED presentations via ICD10 codes.13 In an approach similar to the above, they similarly define baseline influenza activity in historical data, then establish an epidemic threshold above which weekly rates are considered to be in the epidemic period.

NSW Health's Public Health Rapid Emergency Disease and Syndromic Surveillance System uses a statistic called the ‘index of increase’ to indicate when influenza-like illness presentations are increasing at a statistically significant rate.14 It accumulates the difference between the previous day's count of presentations and the average for that weekday over the previous 12 months. An index of increase value of 15 is considered an important indicator for the start of the influenza season.

The World Health Organization has also published surveillance standards for influenza, which contains suggestions for determining seasonal thresholds including Shewhart charts, CUSUM charts and EWMA charts.15

The above methods of monitoring disease outbreaks follow the traditional approach of using daily or weekly counts of disease and flagging an outbreak when these counts are significantly higher than expected. However, in the context of providing alerts as quickly as possible in continuous surveillance programmes, monitoring the time between events has been shown to be more responsive than using daily counts as we do not have to wait for the end of the day or week to make a decision about an outbreak.16, 17

The TBE approach to statistical process control is a very new concept, and we are one of the few groups working on its application to nonhomogeneous processes worldwide. The monitoring plans use an adaptation of EWMAs, which is a process control method with prior use in detecting anomalies typically for industrial production processes. Such manufacturing contexts for modelling failure rates are homogenous processes where the scale and shape parameters of distributions fitting the observed data are always the same. In healthcare, seasonality and day of the week influences are a variation source that leads to nonhomogeneous processes, and during disease outbreaks, there are generally stronger seasonal trends and within-day influences. These aspects make designing a monitoring plan for disease outbreaks a challenging task in practice. The solution applicable to nonhomogeneous processes presented in the present paper is monitoring based on Weibull-distributed TBE values and incorporating differing levels of temporal memory to cover outbreaks of different sizes.

We believe the approach offers superior advantages to existing attempts relating to real-time decision support because each event (ED presentation) provides an opportunity to judge a disease outbreak, and there is no need to wait for a counting period to end. We now know how to design these approaches to be robust at detecting any outbreak.17

In the present study, outbreak detection models were developed based on presentations of patients with influenza-like illness as coronavirus counts at the study sites decreased after an initial first wave. The approach is applicable to other hospitals but calculation of the test statistic that is compared to the threshold requires TBE >0. Most ED information systems do not record presentation time down to the nearest second, and particularly for large hospitals, there could be times when patients are recorded as having arrived within the same minute. This was not a problem in the present study, and monitoring plans that were developed for all hospitals within the ED dataset show outbreaks for individual hospitals coinciding with the COVID-19 outbreak and other times since 2017. The technique is relevant to monitoring other syndrome groupings such as abdominal pain, respiratory conditions, meningitis, gastrointestinal illness, and poisoning) and more broadly to other non-homogenous monitoring applications such as traffic accidents where other factors such as school holidays and day of the week influence the timing of when they are reported.

Syndromic surveillance using ED presentations is becoming an attractive data source for disease surveillance as the ED has become a highly utilised venue for acute care.19 However, ED presentation data is only one approach to detecting outbreaks based on (generally acute) symptoms, and social media (specifically twitter) has also been proposed as a potentially valuable signal to monitor.20 Twitter also has potential value in assessing sentiment dynamics during pandemics (positive versus negative sentimental polarity) particularly in response to government policies such as social distancing and lock downs.21 Others have observed the overwhelming quantity of information (‘infodemics’) associated with pandemics, coupled with the potential of political biases/propaganda, and the absence of expectations that user-generated content is subject to fact-checking, raises concerns about the credibility of information about pandemics on social media.22

Other related surveillance includes online questionnaires,23 temperature monitoring,4 ambulance dispatch calls24 and hybrid systems that cover a range of data sources including Facebook surveys and Google search trends.3, 25 An advantage of having technology companies deploy surveys is the high sample sizes that result. An advantage of our approach of using hospital primary diagnoses as the signal to monitor is that it is assigned by a clinician as opposed to self-reported diagnoses.

Limitations

Influenza-like illness was used for developing outbreak detection models with application to syndromic surveillance. A limitation of this is that COVID-19 has been reported to present with non-respiratory symptoms such as myalgia and anosmia.26, 27 Automatic swab analysis to confirm influenza is not routine ED practice at the study sites, and ‘influenza-like illness’ was defined based on a list of ICD diagnosis codes.

Disease outbreaks were shown to be detectable from ED presentation data. All current syndromic datasets suffer from some selection biases. ED data exclude visits that a patient may have made to a general practitioner and several studies have reported reduced numbers of ED visits during the early stages of the COVID-19 pandemic suggesting the public may have concerns about visiting EDs during a pandemic.28, 29 Changes in patterns of care seeking have been observed during the COVID-19 pandemic (e.g. declines in paediatric and older patient presentations and increases in ED visits for upper respiratory infections, shortness of breath, and chest pain29), and surveillance based on ED presentations can detect departures from expected levels quickly. However, an ideal monitoring strategy would be the amalgamation of numerous health-related datasets into a single surveillance system that reduces the biases and leverages off their collective power.

Conclusion

The present study has shown that disease outbreaks are detectable from ED presentation data. In relation to the current COVID-19 pandemic, as restrictions across nations are further eased or made more stringent, an implemented syndromic surveillance approach can pick up future aberrances and geographic outbreaks quickly so they can be contained.

Acknowledgements

The authors acknowledge funding support of this work from CSIRO, Clinical Excellence Queensland, and SA Health's Commission for Excellence and Innovation in Health and thank members of their data teams and SA Pathology for facilitating extracts of data used in the present study.

Competing interests

None declared.