Earlier Prediction of Influenza Epidemic by Hospital-based Data in Taiwan

2017 2 nd International Conference on Education, Management and Systems Engineering (EMSE 2017) ISBN: 978-1-60595-466-0 Earlier Prediction of Influenza Epidemic by Hospital-based Data in Taiwan Chia-liang FU 1, Ray-jade CHEN 2 and Yu-sheng LO 1,* 1 Graduate Institute of Biomedical Informatics, College of Medical Science and Technology, Taipei Medical University, Taipei, Taiwan 2 Department of Surgery, School of Medicine, College of Medicine, Taipei Medical University, Taipei, Taiwan *Corresponding author Keywords: Influenza, ILI, Clinical data, Surveillance information system, Electronic medical records. Abstract. In Taiwan, a pandemic with human-transmissible influenza severely affected the allocation of acute care resources. Using hospitals clinical data may offer a unique opportunity to predict of influenza epidemic. The aim of this study was to make a use of Taipei Medical University (TMU) clinical database, which can routinely collect clinical data from the three TMU affiliated hospitals, for early predicting of influenza epidemic. We first identified the influenza-related factors through literature validation. Then we extracted the outpatient medical records (2015.06 ~ 2016.10), and through the Pearson correlation coefficient calculated the best combination of influenza detection factors. In addition, through the CUSUM control chart to establish influenza detection model. Finally, we compare the influenza surveillance model with the Taiwan Centers for Disease Control (TWCDC) to monitor the consistency of the model. During the study period, the optimal combination of influenza detection factors had the highest correlation with TWCDC ILI data and virologic surveillance data (r = 0.95086 and 0.87312, respectively; p, 0.0001). The results showed that TMU EDMF and TWCDC epidemic information trends consistent. We automatically collect clinical data to ensure the timeliness of information and 1 to 2 weeks earlier than traditional monitoring methods. Introduction Influenza (flu) is a contagious respiratory illness caused by flu viruses. The flu is transmitted easily [1-3] from person to person in seasonal epidemics, and may lead to severe illness, hospitalization, or even death [5]. Besides, the flu and pneumonia was the 10 leading causes of death in US and Taiwan [6,7]. The most important thing is that serious not only affects the allocation of acute care resources, reflecting the combination of Taiwan's medical resources and epidemic prevention [4]. In order to strengthen the monitoring and resilience of influenza epidemic, all countries in the world are looking to assist in the establishment of early detection systems through information technology, such as Global Influenza Surveillance Network and FluNet System established by WHO [8-11]. In addition, health facility in the world, such as the Taiwan Center for Disease Control and Prevention (TWCDC), has also established a syndromic surveillance system to collect epidemic data for hospitals, medical insurance and contract laboratories [12-18]. The traditional influenza detection method has high accuracy, but it has a problem of timeliness. On the other hand, new technologies that use big data (social media data) can detect trends in real time, but epidemic information is often overestimated, most notably Google Flu Trends (GFT) [19]. This project was first launched in 2008 by Google.org to gather Google search queries that can be used to identify trends and calculate forecasts. GFT was able to predict regional outbreaks of flu 7-10 days earlier than US CDC [20]. However, the "Science" study in 2013 noted that the GFT's actual comparative analysis of influenza data trends in US CDC in 2013, GFT's predicted deviation of more than 200%, significantly higher than the actual data values and closed in 2015 [19]. 403

There are no real-time influenza detection tools in Taiwan, so we use clinical databases to replace social media data as a data source for detecting influenza outbreaks. The purpose of this study was to establish an earlier detection model of flu (EDMF) based on the TMU clinical database routinely collected by three affiliated hospitals. We believe that the timeliness of EDMF information is one to two weeks earlier than that of traditional influenza surveillance. Methods We purposed a method of analyzing the related clinical data, which can be used for depicting influenza data trends in real-time and was to give an early forecasting one to two week ahead of TWCDC surveillance report. Based on the TMU clinical database, firstly we extracted a total of 5,809,089 outpatient data from the historical medical records (2015.06 ~ 2016.10), of which 1,869,205 were Taipei Medical University Hospital, 1,873,212 Wanfang Hospital and 2,066,672 for Shuang Ho Hospital. The influenza-like illness (ILI) epidemic and virologic surveillance data from the TWCDC were used to perform this analysis. We downloaded the publicly available data from the TWCDC website [13]. In addition, the use of literature validation and reference to the definition of influenza-related, by the TMU clinical database to establish a meaningful correlation factors. We have completed the definition of influenza-related factors and studied the contents of the clinical database to determine the data content and collection range of each detection factor (Table 1). Pearson correlation analysis using SAS EG, version 7.4 was performed to examine the correlation of the data from TMU clinical database with the TWCDC ILI and virologic surveillance data. Strong correlation is defined as the correlation coefficient of 0.7. And then, we choose the strongest correlation factors to apply CUSUM procedure. The purpose of this study is to apply the cumulative sum (CUSUM) control chart of the quality improvement tools to detect the outbreak of the influenza Epidemic. In this study, we created a CUSUM chart in Excel using the QI Macros SPC software. Table 1. Data source of variables and the corresponding definition. Variables Diagnosis 1. Flu-like diagnosis (A1) 2. Flu diagnosis (A2*) Signs and symptoms (B) Medication (C) Flu test (D**) 1. Flu test (positive) (D1) 2. Flu test (negative) (D2) Data source of variables and the corresponding definition Data source:outpatient medical records 1. Flu-like diagnosis (1) ICD 9:480.xx~487.xx, 465.9 (2) ICD 10 (Taiwan implemented at 2016.01):A22.1, A37.91, A48.1, B25.0, B44.0, J09.X1, J09.X2, J09.X3, J09.X9, J10.00, J10.01, J10.08, J10.1, J10.2, J10.81, J10.82, J10.83, J10.89, J11.00, J11.08, J11.1, J11.2, J11.81, J11.82, J11.83, J11.89, J12.1, J12.2, J12.3, J12.81, J12.89, J12.9, J13, J14, J15.0, J15.1, J15.20, J15.211, J15.29, J15.3, J15.4, J15.5, J15.6, J15.7, J15.8, J15.9, J16.0, J16.8, J17, J18.0, J18.1, J18.8, J18.9, J06.9 2. Flu diagnosis (1) ICD 9:487.xx (2) ICD 10 (Taiwan implemented at 2016.01):J09.X1, J09.X2, J09.X3, J09.X9, J10.00, J10.01, J10.08, J10.1, J10.2, J10.81, J10.82, J10.83, J10.89, J11.00, J11.08, J11.1, J11.2, J11.81, J11.82, J11.83, J11.89 Data source:outpatient medical records Fever, BT > 38 C, Lethargy, Weariness, Fatigue, Drowsy, Headache, Muascle soreness, Muascle ache, Weakness, Sore throat, Throat hurts, Contact with influenza/flu, Contact influenza/flu, Influenza, Flu, Cough Data source:outpatient medication records Tamiflu, Zanamivir Relenza Data source:laboratory results records 1. Flu test (positive):positive, + 2. Flu test (negative):negative, - * A2 is a subset of A1;** D = D1 D2 404

Result The study was January 4, 2015 (week 26) through October 29, 2016 (week 43). Our analyses used 70 weeks of data obtained from the TWCDC ILI and virologic surveillance systems used to monitor national and regional influenza trends. According to the definition of epidemic molecules, there are 27 classification results of correlation factors. We use these correlation factors as a database query to generate a total of 27 data sets. We then examined the correlation between data from 27 datasets and TWCDC ILI and virological monitoring data at the same period (Table 2). During the entire period of our study, the highest correlation coefficients were 0.87312 (p, 0.0001) with virologic surveillance and 0.95086 (p, 0.001) with TWCDC ILI. The lowest correlation coefficients were -0.1358 (p, 0.2622) with virologic surveillance and -0.004 (p, 0.9735) with TWCDC ILI. Table 2. Pearson s correlation coefficients between TMU s clinical data and TWCDC ILI surveillance data from 2015/06 through 2016/10. Pearson correlation coefficient, N = 70 (weeks) Prob > r (under H0 ): Rho=0 No Variables Taiwan CDC Influenza-like illness (ILI) Taiwan CDC Positive rates of influenza virus r p r p 1 A1 0.78199 p < 0.0001 0.66415 p < 0.0001 3 A2 0.94308 p < 0.0001 0.83024 p < 0.0001 2 B -0.004 p = 0.9735-0.1358 p = 0.2622 4 C 0.92358 p < 0.0001 0.80386 p < 0.0001 5 D 0.95086 p < 0.0001 0.8385 p < 0.0001 6 D1 0.90856 p < 0.0001 0.87312 p < 0.0001 -- -- -- -- -- 19 A1 D 0.93715 p < 0.0001 0.8427 p < 0.0001 20 A1 C D 0.86261 p < 0.0001 0.77747 p < 0.0001 21 A1 D1 0.89846 p < 0.0001 0.87093 p < 0.0001 22 A1 (D1 (C D2)) 0.91675 p < 0.0001 0.86225 p < 0.0001 -- -- -- -- -- Figure 1. TWCDC ILI Trend (January February 2016). Figure 2. TMU EDMF (January February 2016). Based on the previous analysis, we found that RIDT data had the highest correlation with TWCDC ILI, and that data was easy to collect, so we chose to construct an earlier detection model of flu (EDMF) using RIDT data. We first collected RIDT data from the TMU clinical database (January-February 2016) and downloaded synchronic ILI data from the TWCDC website. We then set up the ILI trend graph and EDMF using these data, respectively (Fig. 1, Fig. 2). In Figure 1, that the flu outbreak occurred from the 6th week of 2016 to the end of the 8hth week. At the same time, EDMF control chart trend increased significantly, consistent with the ILI monitoring information. We proved that Figure 1 shows a high degree of consistency with Figure 2. 405

Discussion In this study, we found that the correlation between influenza virus detection and TWCDC ILI data was highest, the correlation coefficient was 0.95086 (p, 0.0001), the correlation between influenza detection positive and TWCDC virological monitoring data was the highest, the correlation coefficient was 0.87312 (p, 0.0001). The results are highly correlated with TWCDC ILI epidemiological information than similar studies [18,20]. Furthermore, we solve the timeliness problem through automatically collect clinical data. TMU affiliated hospitals daily 2:00 am automatically the day before the patient treatment data uploaded to the TMU clinical database. Automatically collect clinical data to ensure the information timeliness and 1 to 2 weeks earlier than the traditional influenza surveillance. In the 7th week of Taiwan in 2016, influenza infection was the highest weekly growth in the past five years [21]. This event gave us a good chance to verify the detection model. In our model (Fig. 2), we found that CUMSUM value of the figure exceeded the threshold for the 7th week of 2016. We had successfully detected the outbreak of the flu epidemic through this model. Our EDMF is based on outpatient medical data, so our model will be affected by the holidays, underestimating the epidemic information. In addition, the accuracy of the data will be affected by some environmental factors (such as temperature, season, week and time-dependent) [18]. Therefore, how to reduce the occurrence of false alarms, to ensure that the sensitivity of this prediction model will be our next major challenge. Conclusion In Taiwan, a pandemic influenza severely affected the allocation of acute care resources. Using hospitals clinical data may offer a unique opportunity to predict of influenza epidemic. The use of this detection model when the flu outbreak occurs hospital infection control personnel can start the investigation. When influenza outbreaks are detected, the infection control team members, health care workers, and the hospital management take appropriate control measures and response assistance (such as vaccination, special outpatient). Of course, our detection model also needs to be improved, such as false alarms, environmental factors. But the biggest contribution of this study is that we have successfully used regional clinical data to predict the national epidemic trends, and more importantly, we make information more immediate and easier to use. Acknowledgement This research was partially supported by Ministry of Science and Technology (MOST) of Taiwan under the grant numbers MOST 105-2218-E-038-004. We thank all of the infection control team members at TMUH for their support. References [1] WHO. Influenza (Seasonal). (Accessed 2016 at http://www.who.int/mediacentre/factsheets/ fs211/en/) [2] Centers for Disease Control and Prevention. Key Facts about Influenza (Flu) & Flu Vaccine. (Accessed 2016 at http://www.cdc.gov/flu/keyfacts.htm) [3] Brankston G, Gitterman L, Hirji Z, Lemieux C, Gardam M. Transmission of influenza A in human beings. Lancet Infect Dis. April 2007, 7 (4): 257 65. [4] Centers for Disease Control, Taiwan (2016). Severe Complicated Influenza. (Accessed 2016 at http://www.cdc.gov.tw/home/influenza) [5] WHO. Influenza fact sheet. (Accessed 2015 at http://www.who.int/mediacentre/factsheets/ 2003/fs211/en/) 406

[6] Centers for Disease Control and Prevention. National Center for Health Statistics. (Accessed 2016 at, http://www.cdc.gov/nchs/products/databriefs/db229.htm) [7] Centers for Disease Control, Taiwan (2012). The number of influenza deaths in our country is equivalent to the ninth largest cause of death. (Accessed 2016 at http://www.cdc.gov.tw/professional/info.aspx?treeid=45da8e73a81d495d&nowtreeid=1bd193ed6da baee6&tid=ad2d4abb1a3c1a51) [8] WHO. The World Health Report 2007: a safer future: global public health security in the 21st century. (Accessed 2015, at http://www.who.int/whr/2007/whr07_en.pdf) [9] WHO. WHO strategic action plan for pandemic influenza. (Accessed 2015, at http://www.who.int/entity/csr/resources/publications/influenza /StregPlanEPR_GIP_2006_2.pdf) [10] WHO. WHO global influenza surveillance network. (Accessed 2015 at http://www.who.int/csr/ disease/influenza/surveillance/en/) [11] WHO. Global pandemic influenza action plan to increase vaccine supply. (Accessed 2015, at http://www.who.int/csr/resources/publications/ influenza/cds_epr_gip_2006_1.pdf) [12] Centers for Disease Control, Taiwan (2013). Strategy Plan for Executive of Influenza pandemic Response (Third Edition). (Accessed 2016 at http://www.cdc.gov.tw/infectionreportinfo. aspx?treeid=075874dc882a5bfd&nowtreeid=8dba723ff186fac0&tid=a3139624d57fd9f9) [13] Centers for Disease Control, Taiwan. Infectious disease notification system. (Accessed 2016 at http://www.cdc.gov.tw/professional/submenu.aspx?treeid=4c19a0252bbef869&nowtreeid=1ddfe9c 52acce91a) [14] Irvin CB, Nouhan PP, Rice K. Syndromic analysis of computerized emergency department patients chief complaints: an opportunity for bioterrorism and influenza surveillance. Ann Emerg Med. 2003;41(4):447 52. [15] Ritzwoller DP, Kleinman K, Palen T, Abrams A, Kaferly J, Yih W, et al. Comparison of syndromic surveillance and a sentinel provider system in detecting an influenza outbreak Denver, Colorado, 2003. Morb Mortal Wkly Rep. 2005; 54(Suppl):151 6. [16] Moore K, Black J, Rowe S, Franklin L. Syndromic surveillance for influenza in two hospital emergency departments. Relationships between ICD-10 codes and notified cases, before and during a pandemic. BMC Public Health. 2011; 11:338. [17] Shimoni Z, Rodrig J, Dusseldorp N, Niven M, Froom P. Increased emergency department chief complaints of fever identified the influenza (H1N1) pandemic before outpatient symptom surveillance. Environ Health Prev Med. 2012;17(1):69 72. [18] Chan, T. C., Teng, Y. C., & Hwang, J. S. (2015). Detection of influenza-like illness aberrations by directly monitoring Pearson residuals of fitted negative binomial regression models. BMC public health, 15(1), 1. [19] Lazer, D., Kennedy, R., King, G., & Vespignani, A. (2014). The parable of Google flu: traps in big data analysis. Science, 343(6176), 1203-1205. [20] Cho, S., Sohn, C. H., Jo, M. W., Shin, S. Y., Lee, J. H., Ryoo, S. M.,... & Seo, D. W. (2013). Correlation between national influenza surveillance data and google trends in South Korea. PLoS One, 8(12), e81422. [21] Taiwan records 320 severe flu cases in a week, highest in 5 years. (Accessed 2017, at http://focustaiwan.tw/news/asoc/201602230019.aspx) 407