Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Improving the Performance of Outbreak Detection Algorithms by Classifying the Levels of Disease Incidence

  • Honglong Zhang ,

    Contributed equally to this work with: Honglong Zhang, Shengjie Lai

    Affiliation Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Diseases, Chinese Center for Disease Control and Prevention, Beijing, China

  • Shengjie Lai ,

    Contributed equally to this work with: Honglong Zhang, Shengjie Lai

    Affiliation Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Diseases, Chinese Center for Disease Control and Prevention, Beijing, China

  • Liping Wang,

    Affiliation Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Diseases, Chinese Center for Disease Control and Prevention, Beijing, China

  • Dan Zhao,

    Affiliation Institute of Immunization and Prevention, Beijing Center for Disease Control and Prevention, Beijing, China

  • Dinglun Zhou,

    Affiliation West China School of Public Health, Sichuan University, Chengdu, China

  • Yajia Lan,

    Affiliation West China School of Public Health, Sichuan University, Chengdu, China

  • David L. Buckeridge,

    Affiliation Surveillance Lab, McGill Clinical and Health Informatics, Department of Epidemiology and Biostatistics, McGill University, Montreal, Quebec, Canada

  • Zhongjie Li ,

    lizj@chinacdc.cn (ZL); yangwz@chinacdc.cn (WY)

    Affiliation Key Laboratory of Surveillance and Early-warning on Infectious Disease, Division of Infectious Diseases, Chinese Center for Disease Control and Prevention, Beijing, China

  • Weizhong Yang

    lizj@chinacdc.cn (ZL); yangwz@chinacdc.cn (WY)

    Affiliation Chinese Center for Disease Control and Prevention, Beijing, China

Improving the Performance of Outbreak Detection Algorithms by Classifying the Levels of Disease Incidence

  • Honglong Zhang, 
  • Shengjie Lai, 
  • Liping Wang, 
  • Dan Zhao, 
  • Dinglun Zhou, 
  • Yajia Lan, 
  • David L. Buckeridge, 
  • Zhongjie Li, 
  • Weizhong Yang
PLOS
x

Abstract

We evaluated a novel strategy to improve the performance of outbreak detection algorithms, namely setting the alerting threshold separately in each region according to the disease incidence in that region. By using data on hand, foot and mouth disease in Shandong province, China, we evaluated the impact of disease incidence on the performance of outbreak detection algorithms (EARS-C1, C2 and C3). Compared to applying the same algorithm and threshold to the whole region, setting the optimal threshold in each region according to the level of disease incidence (i.e., high, middle, and low) enhanced sensitivity (C1: from 94.4% to 99.1%, C2: from 93.5% to 95.4%, C3: from 91.7% to 95.4%) and reduced the number of alert signals (the percentage of reduction is C1∶4.3%, C2∶11.9%, C3∶10.3%). Our findings illustrate a general method for improving the accuracy of detection algorithms that is potentially applicable broadly to other diseases and regions.

Introduction

Detecting infectious diseases outbreaks at an early stage is crucial for timely implementation of control measures, which can minimize morbidity and mortality. A considerable amount of research has focused on developing statistical methods to identify aberrations in disease incidence data accurately and quickly [1][4]. In current public health practice, aberrancy-detection algorithms, including temporal (e.g., Exponentially Weighted Moving Average and cumulative sum), spatial (e.g., Spatial Scan Statistic), and spatio-temporal (e.g., Space-Time Scan Statistic) methods, can contribute important information to support outbreak detection and management [3][6]. Evaluations of surveillance systems have demonstrated that many factors affect the accuracy of aberration detection, including the mode of disease transmission, seasonal patterns in disease incidence, the detection algorithm used, and the alerting threshold of the algorithm [7], [8]. Disease incidence can vary greatly between regions under surveillance, but it is not known to what extent this variation in incidence affects the accuracy and timeliness of aberration detection if the same algorithm with a constant alerting threshold is applied to all regions.

Hand, foot and mouth disease (HFMD) is caused by serotypes of enterovirus and, usually leads to mild symptoms, but can result in serious complication or death [9][11]. In China, HFMD tends to infect infants and children younger than 5 years old [12], with more than one million cases reported and three hundred deaths nationwide in 2009 [13]. Within China, Shandong province is one of the most seriously affected regions with an annual incidence of 92.2 per 100,000 to 149.4 per 100,000 between 2009 and 2012. Early detection of outbreaks and prompt diagnosis and treatment of cases at high risk of severe disease are key principles in minimizing the impact of HFMD [14]. In this study, we used data from surveillance of HFMD in Shandong province to evaluate a novel strategy to improve the accuracy and timeliness of outbreak detection. More precisely, we examined whether setting the optimal alerting threshold separately in each region according to the disease incidence in that region would improve the accuracy and timeliness of outbreak detection as compared to using a constant alerting threshold across all regions.

Methods

In China, a probable HFMD case was defined as a patient with papular or vesicular rash on hands, feet, mouth or buttocks, with or without fever. A confirmed case was defined as a probable case with laboratory evidence of enterovirus infection (by EV71, CA16, or other enteroviruses) detected by reverse-transcriptase polymerase chain reaction (RT-PCR), real-time RT-PCR, or by virus isolation [15]. HFMD has been a notifiable infectious disease in China since May 2008. Clinicians are required to report both probable and confirmed HFMD cases through a web-based reporting system, the Nationwide Notifiable Infectious Diseases Reporting Information System. Using this system, clinicians in all health care institutes throughout the country can report cases of notifiable infectious diseases via the Internet to a data center located in the Chinese Center of Disease Control and Prevention (China CDC). Cases of HFMD reported between January and December 2009 in Shandong province of China were used in this study.

We obtained the population of the 142 counties in Shandong province in 2009 from Chinese State Statistics Bureau. The counties were categorized into 3 groups on the basis of annual HFMD incidence, ① low level: with the disease incidence rate ranging from 7 to 149 per 100,000, ② middle level: with the disease incidence rate ranging from 150 to 249 per 100,000, ③ high level: with the disease incidence rate ranging from 250 to 420 per 100,000. Following an examination of the distribution of disease incidence by county, we selected these categories to reflect natural groupings in the data.

We used the three Early Aberration Reporting System (EARS) algorithms (C1, C2, and C3) developed by the US Center for Disease Control and Prevention [3], [16]. These algorithms require few historical baseline data and are based on statistical process control methods. These algorithms estimate the expected value on any given day as the average of the observed values over 7 previous days. For the C1 algorithm, the baseline is the past 7 days (ie, t-1 to t-7), while for the C2 and C3 algorithms, the baseline incorporates a 2-day lag before the current day (ie, t-3 to t-9). The C3 algorithm also maintains a 3-day running sum, and the commonly used threshold for C1 and C2 is 2.0 [3], [16]. All three algorithms are described in detail elsewhere [3], [16], [17]. In our study, acknowledging that the algorithm threshold could impact outbreak detection performance, we tested 30 candidate threshold values (from 0.1 to 3.0, interval is 0.1) for C1, C2, and C3, so as to determine the optimal threshold for each algorithm when were applied to data for each incidence category. All algorithms and analyses were implemented with R software [18].

In China, the definition of reported HFMD outbreak was that ≥10 cases occurring in the same gathering settings (e.g., kindergarten, school), or ≥5 cases occurring in the same village or community within one week [19]. The HFMD outbreaks reported in 2009 were assumed to be the only true outbreaks in the data, as all of these outbreaks were verified through field investigation by local public health departments [15]. We defined the start and end of an outbreak as the first and last dates, respectively, of reported cases associated with the outbreak.

We evaluated algorithms in terms of their sensitivity, specificity, and time to detection (TTD). Sensitivity was defined as the number of outbreaks during which at least one alert was signaled, divided by the total number of reported outbreaks. Specificity was defined as the number of non-outbreak days with no alert, divided by the total number of non-outbreak days [20]. TTD was defined as the median number of days from the beginning of each outbreak to the first alert during the outbreak. If the algorithm alerted on the first day of an outbreak, detection time was zero. To enable the calculation of detection timeliness of all outbreaks, if an outbreak was undetected, TTD was assigned the total duration of the outbreak, so as to enable calculation of the median timeliness across all outbreaks. Therefore, TTD is an integrated index that reflects both the timeliness and sensitivity of an algorithm [7]. The optimal threshold for an algorithm was the one with the shortest TTD, or with the highest specificity when (a) the TTD was either same or (b) had a difference of less than half a day and the difference between the specificity was >5.0% [7]. We used the Student t test to examine whether the number of signals was significantly different by setting the optimal threshold in each region according to the level of disease incidence and using the same optimal thresholds to the whole region.

Results

In 2009, a total of 138,593 cases and 108 outbreaks of HFMD were reported from the 142 counties of Shandong province. The county incidence rate ranged from 7 per 100,000 to 420 per 100,000. According to our classification criteria of low, middle and high disease incidence, there were 85 (59.9%) counties with a low disease incidence, 39 (27.5%) counties with a middle disease incidence, and 18 (12.6%) counties with a high disease incidence (Table 1). The total number of outbreaks reported was 32, 47 and 29 in low, middle and high disease incidence regions, respectively. The median number of cases per outbreak was similar among the three groups with low, middle and high level of disease incidence.

thumbnail
Table 1. The cases and outbreaks in the regions with high, middle and low level of hand, foot and mouth (HFMD) incidence rate.

https://doi.org/10.1371/journal.pone.0071803.t001

Using data from all counties to determine the optimal alerting threshold for C1, C2 and C3, we found that the optimal thresholds for C1, C2 and C3 were 0.4, 0.4, and 0.5, respectively. When using these thresholds to apply the three algorithms to the regions with high disease incidence, C1 and C2 had the highest sensitivity (96.6%), C3 had the highest specificity (88.2%), and the three methods had the same TTD (3 days) (Table 2). For the regions with middle disease incidence, C1 and C2 had the highest sensitivity (91.5%), while C3 had the highest specificity (88.5%), and the three methods had the same TTD (2 days). For the regions with low disease incidence, C1 and C3 had the highest sensitivity (96.9%), C3 had the highest specificity (89.0%), and three methods had the same TTD (0.5 day).

thumbnail
Table 2. The number of signals, sensitivity, specificity and time to detection for C1, C2 and C3 algorithms, based on the optimal thresholds identified by employing the data for all counties.

https://doi.org/10.1371/journal.pone.0071803.t002

When using only data from counties within a single incidence class to determine the optimal alerting threshold for each algorithm, for regions with a high disease incidence, C1, C2 and C3 all had an optimal threshold of 0.3. Applying the algorithms to high-incidence counties with this threshold resulted in a TTD of 2 days, with C1 having the highest sensitivity (100%) and C2 having the highest specificity (86.8%) (Table 3). For the regions with a middle disease incidence, C1, C2 and C3 all had an optimal threshold of 0.3, which resulted in a TTD of 1 day, with C1 having the highest sensitivity (100%) and C2 having the highest specificity (87.2%). For the regions with low disease incidence, the optimal thresholds for C1, C2 and C3 were 0.7, 1 and 1.3, respectively. The three methods had the same TTD of 0.5 days, with C1 having the highest sensitivity (96.9%) and C3 having the highest specificity (92.6%).

thumbnail
Table 3. The number of signals, sensitivity, specificity and time to detection for C1, C2 and C3 algorithms, based on the optimal thresholds identified separately for the 3 categories of region with different level of HFMD incidence rate.

https://doi.org/10.1371/journal.pone.0071803.t003

Comparing algorithm performance when the optimal thresholds were determined from all counties, as opposed to within each incidence class, the sensitivity of outbreak detection was higher when using only data from one incidence class (C1: from 94.4% to 99.1%, C2: from 93.5% to 95.4%, C3: from 91.7% to 95.4%), as was specificity (C1: from 87.6% to 88.1%, C2: from 88.3% to 89.7%, C3: from 88.8% to 90.0%), The number of signals was statistically significantly decreased when using only data from one incidence class (the percentage of signals reduction for C1∶4.3% (p<0.001), for C2∶11.9% (p<0.001), for C3∶10.3% (p<0.001)), while maintaining a consistent TTD (1 day).

Discussion

The results of this study demonstrate that adopting optimizing surveillance alert thresholds by incidence category can improve aberration detection performance as compared to using the same alert threshold across all regions. In particular, for the EARS algorithms applied to HFMD data from counties in Shandong province we observed the same TTD, but higher sensitivity and specificity when alert thresholds were optimized within three incidence categories.

Our findings may be explained in part by the observation that the number of cases and the number and scale of outbreaks differed greatly among regions from different incidence categories. These factors help to explain why the optimal threshold of an algorithm may differ across regions with unequal disease incidence. In other words, the optimal alert threshold for an aberration detection method across all region is a compromise of sorts. A gain in detection accuracy can be realized by further optimizing the alert threshold for groups of sub-regions with similar disease incidence.

An important strength of this study is the use of a large amount of real surveillance data with validated case and outbreak reports. This study is the first to suggest a straightforward method for improving the accuracy of outbreak detection algorithm in a large area by optimizing alerting thresholds within incidence categories.

One limitation of our study is that we used only one disease as an example and it is possible that our results will not generalize to other diseases with a low incidence. In our study we divided counties in Shandong province into three incidence categories to reflect natural groupings in the data, but without considering other factors, such as the differences of population, case report timeliness and completeness, and the characteristics of seasonality and weekend effect of surveillance data [20][22], which could affect algorithm performance. The objective of this study, however, was to explore the influence of the variation in incidence rates on algorithm performance, conditional on the observed variation in other factors. Any attempt to simultaneously estimate the absolute effects (and possible interactions) of multiple determinants of outbreak detection would require more extensive adjustment for other factors, but that was not the objective of the study. Also, it is possible that using a greater number of incidence categories and taking into account more characteristics of disease occurring could further improve detection performance and we consider this to be a promising area for future research.

In conclusion, our study illustrates a general method for improving the accuracy of aberration detection algorithms that is potentially applicable broadly to other diseases and regions. Although not measured directly in this study, improvements in the accuracy and timeliness of outbreak detection can have an important impact of the effectiveness of measures to control epidemics and minimize the impact of diseases.

Acknowledgments

We thank Mr. Jizeng Liu and Mr. Wenhua Liu for assisting in the data collation.

Author Contributions

Conceived and designed the experiments: WZY ZJL SJL HLZ. Performed the experiments: HLZ SJL. Analyzed the data: SJL HLZ LPW DZ DLZ YJL. Contributed reagents/materials/analysis tools: HLZ SJL. Wrote the paper: HLZ SJL DB ZJL WZY.

References

  1. 1. Yang WZ, Xing HX, Wang HZ, Lan YJ, Hu SX, et al. (2004) Early warning technology study on seven infectious disease controls charts methods. Chin J Epidemiol 25(12): 1039–1041.
  2. 2. Burkom HS (2003) Development, Adaptation, and Assessment of Alerting Algorithms for Biosurveillance. J HOPKINS APL TECH D 24(4): 335–342.
  3. 3. Hutwagner LC, Thompson W, Seeman GM, Treadwell T (2003) The bioterrorism preparedness and response Early Aberration Reporting System (EARS). J Urban Health 80 (2 Suppl 1)i89–i96.
  4. 4. Buckeridge DL, Burkom H, Campbell M, Hogan WR, Moore AW (2005) Algorithms for rapid outbreak detection: a research synthesis. J Biomed Inform 38(2): 99–113.
  5. 5. Hjalmars U, Kulldorff M, Gustafsson G, Nagarwalla N (1996) Childhood leukaemia in Sweden: using GIS and a spatial scan statistic for cluster detection. Stat Med 15(7–9): 707–715.
  6. 6. Buckeridge DL, Musen MA, Switzer P, Crubezy M (2003) An analytic framework of space-time aberrancy detection in public health surveillance data. AMIA Annu Symp Proc: 120–124.
  7. 7. Wang XL, Zeng D, Seale H, Su L, Cheng H, et al. (2010) Comparing early outbreak detection algorithms based on their optimized parameter values. J Biomed Inform 43(1): 97–103.
  8. 8. Wang XL, Wang QY, Liu DL, Zeng DJ, Cheng H, et al. (2009) Comparison between early outbreak detection models and simulated outbreaks of measles in Beijing. Chin J Epidemiol 30(2): 159–162.
  9. 9. Chan KP, Goh KT, Chong CY, Teo ES, Lau G, et al. (2003) Epidemic hand, foot and mouth disease caused by human enterovirus 71, Singapore. Emerging Infectious Diseases 9: 78–85.
  10. 10. Zhang Y, Zhu Z, Yang W, Ren J, Tan X, et al. (2010) An emerging recombinant human enterovirus 71 responsible for the 2008 outbreak of Hand, Foot and Mouth Disease in Fuyang city of China. Virol J 7: 94.
  11. 11. Zhang Y, Tan XJ, Wang HY, Yan DM, Zhu SL, et al. (2009) An outbreak of hand, foot and mouth disease associated with subgenotype C4 of human enterovirus 71 in Shandong, China. Journal Of Clinical Virology 44: 262–7.
  12. 12. Heymann DL (2008) Control of communicable diseases manual.19th ed. Washington, DC: American Public Health Association p.151–154.
  13. 13. Chang ZR, Zhang J, Sun JL, Zhang WD, Wang ZJ (2011) Epidemiological features of hand, foot and mouth disease in China, 2008–2009. Chin J Epidemiol 32(7): 676–680.
  14. 14. World Health Organization Western Pacific Region. A guide to clinical management and public health response for hand, foot and mouth disease. Available: http://www.wpro.who.int/publications/docs/GuidancefortheclinicalmanagementofHFMD.pdf. Accessed 2013 Jun 1.
  15. 15. Ministry of Health of the People’s Republic of China. Guideline of hand, foot and mouth disease control and prevention (2009 edition) [in Chinese]. Available: http://www.moh.gov.cn/mohjbyfkzj/s3578/200906/41047.shtml. Accessed 2013 March 1.
  16. 16. Hutwagner CL, Barson JV (2005) Use of the early aberration reporting system (EARS) for detection of bioterrorism agent attacks. Aviat Space Environ Med 76(10): 1001–1002.
  17. 17. Zhang HL, Lai SJ, Li ZJ, Lan YJ, Yang WZ (2010) Application of cumulative sum control charts methods in infectious disease outbreak detection. Chin J Epidemiol 31(12): 1403–1406.
  18. 18. Watkins RE, Eagleson S, Veenendaal B, Wright G, Plant AJ (2008) Applying cusum-based methods for the detection of outbreaks of Ross River virus disease in Western Australia. BMC Med Inform Decis Mak 8(37).
  19. 19. Ministry of Health of the People’s Republic of China. Norms of hand, foot and mouth disease clustering and outbreaks response (2012 edition) [in Chinese]. Available: http://www.moh.gov.cn/mohjbyfkzj/s3577/201206/55215.shtml. Accessed 2013 Jun 1.
  20. 20. Li ZJ, Lai SJ, Buckeridge DL, Zhang HL, Lan YJ, et al. (2011) Adjusting outbreak detection algorithms for surveillance during epidemic and non-epidemic periods. J Am Med Inform Assoc 2012 19: e51ee53
  21. 21. Lai SJ, Li ZJ, Zhang HL, Lan YJ, Yang WZ (2011) The comparison of two different types of baseline data regarding the performance of aberration detection algorithm for infectious disease outbreaks. Chin J Epidemiol 32(6): 579–582.
  22. 22. Centers for Disease Control and Prevention. BioSense real-time hospital data user guide application version 2.11 (November 2007). Available: http://www.cdc.gov/biosense/files/CDC_BioSense_BioSense_Hospital_Data_User_Guide_V2.11.pdf. Accessed 2012 March 1.