Human temperatures for syndromic surveillance in the emergency department: data from the autumn wave of the 2009 swine flu (H1N1) pandemic and a seasonal influenza outbreak

Background The emergency department (ED) increasingly acts as a gateway to the evaluation and treatment of acute illnesses. Consequently, it has also become a key testing ground for systems that monitor and identify outbreaks of disease. Here, we describe a new technology that automatically collects body temperatures during triage. The technology was tested in an ED as an approach to monitoring diseases that cause fever, such as seasonal flu and some pandemics. Methods Temporal artery thermometers that log temperature measurements were placed in a Boston ED and used for initial triage vital signs. Time-stamped measurements were collected from the thermometers to investigate the performance a real-time system would offer. The data were summarized in terms of rates of fever (temperatures ≥100.4 °F [≥38.0 °C]) and were qualitatively compared with regional disease surveillance programs in Massachusetts. Results From September 2009 through August 2011, 71,865 body temperatures were collected and included in our analysis, 2073 (2.6 %) of which were fevers. The period of study included the autumn–winter wave of the 2009–2010 H1N1 (swine flu) pandemic, during which the weekly incidence of fever reached a maximum of 5.6 %, as well as the 2010–2011 seasonal flu outbreak, during which the maximum weekly incidence of fever was 6.6 %. The periods of peak fever rates corresponded with the periods of regionally elevated flu activity. Conclusions Temperature measurements were monitored at triage in the ED over a period of 2 years. The resulting data showed promise as a potential surveillance tool for febrile disease that could complement current disease surveillance systems. Because temperature can easily be measured by non-experts, it might also be suitable for monitoring febrile disease activity in schools, workplaces, and transportation hubs, where many traditional syndromic indicators are impractical. However, the system’s validity and generalizability should be evaluated in additional years and settings. Electronic supplementary material The online version of this article (doi:10.1186/s12873-016-0080-7) contains supplementary material, which is available to authorized users.


Introduction
Judged based on the appearance of the data alone, it was not clear whether fever rates from a single ED would be sufficient to detect significant increases in febrile disease, or whether the detected increases would correspond with the periods of elevated influenza activity during the period of the study. To address these points, we retrospectively applied Manitz and Hohle's Bayesian outbreak detection algorithm (BODA) [1,2]. This algorithm generates alarms signalling that the underlying process is out-of-control relative to the recent past. Alarms generated by the algorithm were compared with two outbreaks that occurred during the study period-the autumnwinter wave of the 2009 swine flu (H1N1) pandemic and the seasonal influenza outbreak of 2011.

Technical details of BODA
Conceptually, BODA is a Bayesian extension of the well-known Farrington algorithm of outbreak detection [3]. In essence, BODA works by learning the expected distribution of fevers in a non-epidemic period, and then checks whether the current rate of fevers is statistically consistent (in control) or inconsistent (out of control) with this expected distribution, while using Bayesian updating to incorporate information on the fever rates during previous days and account for both estimation and prediction uncertainty.
BODA was selected for several reasons: First, because it allowed the false alert criterion to be specified in the same manner as in the design of our original validation analyses. Second, because it accommodates overdispersion, which was observed in the data. Third, because it allowed us to incorporate the number of in-use thermometers within the algorithm. Fourth, because it is widely available in the R surveillance package [4]. Finally, because it has few adjustable parameters in the surveillance package, which decreases experimenter degrees of freedom. We viewed it as especially important to reduce experimenter degrees of freedom given the limited number of outbreaks that were analyzed in our study (n = 2). The parameters were chosen a priori-they were not tuned.
The details of each BODA implementation were as follows: A negative binomial generalized additive model was used to predict the count of expected fevers and learn non-epidemic (incontrol) parameters for use by the algorithm. These non-epidemic parameters were learned from the last 12 weeks of the study. Although it would be more common to learn parameters from the first weeks of data, this was not possible because the autumn wave of the H1N1 epidemic coincided with that period. The number of thermometers used during each time segment (week or day) was included in BODA as a covariate. The probability of a false alert was set at ≤3 per year for the main analysis and at 1-5 per year for secondary analyses.

True alerts and false alerts
True alerts were defined as alerts triggered during weeks for which ILI activity in New England exceeds the CDC's threshold [5]. False alerts were defined as alerts triggered during any other period. The algorithm's parameters were set to allow ≤3 false alarms per year.
Since fever rates may be elevated by diseases other than influenza, the definitions of true and false alerts are practical approximations, rather than gold standard ideals. However, it was necessary to develop some criteria for true and false alerts, even if imperfect, because these definitions are required to evaluate the performance of an aberrant event detection analysis.

Robustness and timeliness of outbreak detection
Two analyses were performed to examine the robustness and timeliness of outbreak detection using real-time fever data. To examine robustness, we re-ran the outbreak detection analysis with various acceptable rates of false alerts, ranging from ≤1 to ≤5 per year. We then checked whether outbreaks continued to be detected across the range of allowed false alert rates. To examine timeliness, we ran the outbreak detection analysis twice. First, we ran it on fever data that had been binned into single week intervals, simulating the case of having data available on a weekly basis, which is currently the most common situation in US biosurveillance of influenza and some other diseases. Second, we ran the BODA on fever data that had been binned into 1-day intervals as a proxy for real-time availability. We then compared the earliest dates at which increased fever rates were detected during the periods of the influenza outbreaks, using the weekly and daily data with various false alert criteria.

Results
Additional Figure 1 presents the results of applying a Bayesian outbreak detection algorithm (BODA) [1] to the number of fevers measured in the ED per thermometer. The light blue and dark blue areas show the weekly and daily rates of fevers, respectively. For example, data from the first week of August 2011 show that around 5 fevers were measured per thermometer. The outbreak detection algorithm produces an epidemic threshold (red lines) that is used to judge whether an outbreak is detected. When the observed fever rates exceed this threshold, an alarm is declared (red triangles). Using fever rates from the Boston ED, the algorithm successfully detected both the autumn-winter wave of the 2009-2010 swine flu pandemic and the seasonal flu epidemic of 2011. Outbreak detection was successful for both the weekly data and the daily data, even though the daily data appear quite noisy to the eye.
False alarms are an important concern in outbreak detection, since they lead to wasted effort and potential "boy who cried wolf" situations. Here, we defined a false alarm as any alarm issued outside of the epidemic periods shown in the figure (see the Methods section for more details). A key advantage of the Bayesian outbreak detection algorithm is that it allows the expected number of false alarms to be established prospectively. In the figure, we arbitrarily employed the criterion that an average of ≤3 false alarms should be issued per year. Both the weekly and daily analyses satisfied this criterion (weekly, 0 false alarms; daily, 2 apparent false alarms in 2 years). In additional analyses, we employed false alarm criteria ranging from ≤1 to ≤5 per year. While the weekly outbreak detection analysis showed diminishing ability to detect the outbreaks for low false alarm criteria, the daily analysis detected both outbreaks for all of the false alarm criteria, suggesting that daily fever data provided robust evidence of these outbreaks (Additional Table 1).
To examine the timeliness that could be offered by using real-time data instead of weekly data, we compared the dates at which the algorithm was able to detect periods of elevated febrile activity based on daily and weekly rates of fever. As compared with the weekly data, the daily data provided 18 days of advanced notice for the H1N1 pandemic and 5 days of advanced notice for the seasonal flu outbreak with ≤3 false alarms per year. These indications are promising, especially because weekly CDC surveillance on influenza-like illness and laboratory-confirmed influenza is usually subject to several additional days of delay in practice, as the data are compiled. However, our analysis was importantly limited by the fact that our analysis only investigated two epidemic periods, both of which were of influenza.

Summary
In brief, the analysis successfully detected significant increases in fever rates, which corresponded with both the autumn-winter wave of the H1N1 pandemic and the seasonal influenza outbreak. The analysis suggests that temperature data could have potential as a surveillance tool for febrile illness, especially if it were employed in conjunction with other syndromic indicators. However, temperature data from additional settings and years would be needed to establish fever rates as a means of rapidly detecting outbreaks of influenza or other specific diseases.