Rapid detection of pandemic influenza in the presence of seasonal influenza

Abstract

Background: Key to the control of pandemic influenza are surveillance systems that raise alarms rapidly and sensitively. In addition, they must minimise false alarms during a normal influenza season. We develop a method that uses historical syndromic influenza data from the existing surveillance system 'SERVIS' (Scottish Enhanced Respiratory Virus Infection Surveillance) for influenza-like illness (ILI) in Scotland.

Methods: We develop an algorithm based on the weekly case ratio (WCR) of reported ILI cases to generate an alarm for pandemic influenza. From the seasonal influenza data from 13 Scottish health boards, we estimate the joint probability distribution of the country-level WCR and the number of health boards showing synchronous increases in reported influenza cases over the previous week. Pandemic cases are sampled with various case reporting rates from simulated pandemic influenza infections and overlaid with seasonal SERVIS data from 2001 to 2007. Using this combined time series we test our method for speed of detection, sensitivity and specificity. Also, the 2008-09 SERVIS ILI cases are used for testing detection performances of the three methods with a real pandemic data.

Results: We compare our method, based on our simulation study, to the moving-average Cumulative Sums (Mov-Avg Cusum) and ILI rate threshold methods and find it to be more sensitive and rapid. For 1% case reporting and detection specificity of 95%, our method is 100% sensitive and has median detection time (MDT) of 4 weeks while the Mov-Avg Cusum and ILI rate threshold methods are, respectively, 97% and 100% sensitive with MDT of 5 weeks. At 99% specificity, our method remains 100% sensitive with MDT of 5 weeks. Although the threshold method maintains its sensitivity of 100% with MDT of 5 weeks, sensitivity of Mov-Avg Cusum declines to 92% with increased MDT of 6 weeks. For a two-fold decrease in the case reporting rate (0.5%) and 99% specificity, the WCR and threshold methods, respectively, have MDT of 5 and 6 weeks with both having sensitivity close to 100% while the Mov-Avg Cusum method can only manage sensitivity of 77% with MDT of 6 weeks. However, the WCR and Mov-Avg Cusum methods outperform the ILI threshold method by 1 week in retrospective detection of the 2009 pandemic in Scotland.

Conclusions: While computationally and statistically simple to implement, the WCR algorithm is capable of raising alarms, rapidly and sensitively, for influenza pandemics against a background of seasonal influenza. Although the algorithm was developed using the SERVIS data, it has the capacity to be used at other geographic scales and for different disease systems where buying some early extra time is critical.

Figure 1

Historical SERVIS ILI data. (a) Weekly ILI cases reported by the SERVIS sentinel GPs aggregated at the national level. Week 1 is the first week of October, and week 33 is the third week of May. In the 2002-03 season, the influenza surveillance was continued for another 15 weeks because of the SARS epidemic. (b) The weekly reported ILI cases of the influenza season 2004-05 aggregated at the health board level. (Health boards are represented by the following letter codes. AA: Ayrshire & Arran; BDS: Borders; DG: Dumfries & Galloway; FF: Fife; FV: Forth Valley; GMP: Grampian; GGC: Greater Glasgow & Clyde; HG: Highland; LN: Lanarkshire; LO: Lothian; OR: Orkney; SH: Shetland; and TA: Tayside). In that season, only 10 out of 13 participating HBs had one or more sentinel GPs that reported ILI cases.

Figure 2

The joint probability distribution of (WCR, N_HB). The joint probability distribution of weekly case ratio (WCR), calculated from seasonal ILI cases aggregated over Scotland, and the number of health boards (N_HB) reporting increases in the ILI cases over the previous week. This probability diagram was constructed using SERVIS data from six seasons, 2001-02 to 2006-07. WCR is binned with bin size 0.1. The inset plot zooms in on the left corner of the main plot.

Figure 3

Detection specificity of the WCR method. An illustration of how we calculate the specificity (Sp) of the WCR algorithm as a function of the detection threshold δ. The straight-line plots show two values of δ: 0.000059 (dashed line) and 0.0008 (dot-dash line) which give specificity of 99% and 95% respectively. The other plots are probability values of (WCR, N_HB) in any given week for different seasons. If the probability in any given week dips below one of the threshold probabilities, then a false positive alert for a pandemic outbreak is generated.

Figure 4

Detection of a pandemic by the WCR method. Probability values of weekly (WCR, N_HB) pairs for the model pandemic cases are obtained from the distribution shown in Figure 2. Two examples, one each for pandemic starting weeks 1 (top) and 15 (bottom), are for a single run of the pandemic model. Pandemic cases were sampled at the reporting rate of 0.5% from the weekly infections and added to the 2004-05 seasonal data and WCR and N_HB were calculated from the resultant time series. The dashed lines represent the two values of the detection threshold δ: 0.000059 (Sp = 99%) and 0.0008 (Sp = 95%). The pandemic starting on week 1 was detected on week 7, while the one starting on week 15 was detected on week 21.

Figure 5

Distribution of detection times. A comparison between the distributions of detection times of (a) WCR algorithm, (b) ILI rate threshold, and (c) Mov-Avg Cusum method. Pandemic detection times are in terms of within n weeks of the first pandemic infections occurring; different colours of the colour bar code different values of n. Individual stacked bars represent the distribution of detection times calculated for a set of 1800 overlaid time series for different pandemic start weeks, as shown on the x-axis, from week 1 to week 33 of a typical influenza season.

Figure 6

speed of detection in the early weeks of pandemics. Cumulative probability of detection as a function of times taken in pandemic detection (i.e., weeks within pandemic starting). These plots are for different specificity (Sp) and pandemic case reporting rate (α): (a) Sp = 99%, α = 0.5%; (b) Sp = 95%, α = 0.5%; (c) Sp = 99%, α = 1%; (d) Sp = 95%, α = 1%, (e) Sp = 99%, α = 5%; and (f) Sp = 95%, α = 5%. The dashed line represents the 50%-level of pandemic detections. (For each method in all subplots, the detection data were pooled together from a set of 1800 overlaid time series times 33 weeks.)

Figure 7

Retrospective detection of the current pandemic using the 2009 SERVIS data. The retrospective detection of the 2009 influenza A(H1N1)v pandemic using SERVIS ILI data. The first cases of the 2009 pandemic were reported in Scotland in the 29^th week of the season and our algorithm as well as the Mov-Avg Cusum method detects the pandemic 12 weeks later in week 41. The ILI threshold method detects it 1 week later in week 42.