Adapting Syndromic Surveillance Systems to Increase Value to Local Health Departments

Objective

Our objective was to describe changes in use following syndromic surveillance system modifications and assess the effectiveness of these modifications.

Introduction

Syndromic surveillance systems offer richer understanding of population health. However, because of their complexity, they are less used at small public health agencies, such as many local health departments (LHDs). The evolution of these systems has included modifying user interfaces for more efficient and effective use at the local level. The North Carolina Preparedness and Emergency Response Research Center previously evaluated use of syndromic surveillance information at LHDs in North Carolina. Since this time, both the NC DETECT system and distribution of syndromic surveillance information by the state public health agency have changed. This work describes use following these changes.

Methods

Data from NC DETECT were used to assess the number of users and usage time. Staff from 14 NC LHDs in 2009 and from 39 LHDs in 2012 were surveyed (May–August of 2009 and June of 2012) to gather information on the mode of access to syndromic surveillance information and how this information was used. Data were analyzed to assess the link between the mode of access and use of syndromic surveillance data.

Results

System changes made between 2009 and 2012 included the creation of “dashboards” (Figure 1) which present users with LHD-specific charts and graphs upon login and increases in the distribution of syndromic surveillance information by the state public health agency. The number of LHD-based NC DETECT system users increased from 99 in 2009 to 175 in 2012. Sixty-two of 72 respondents completed the 2012 survey (86%). Syndromic surveillance information was used in 28/40 LHDs (70%) for key public health tasks. Among 20 NC EDSS leads reporting an outbreak in the past year, 25% reported using data from NC DETECT for outbreak response, compared to 23% in 2009 (Figure 2). Among 30 responding NC EDSS leads, 57% reported using data from NC DETECT to respond to seasonal events such as heat-related illness or influenza, compared to 46% in 2009. NC DETECT data were reported to have been used for program management by 30% (compared to 25% in 2009), and to have been used in reports by 33% (compared to 23% in 2009).Open in a separate windowFigure 1:NC DETECT dashboardsOpen in a separate windowFigure 2:Uses of syndromic surveillance information, communicable disease staff 2009 (13 LHDs) and 2012 (31 LHDs)

Conclusions

Changes in how syndromic surveillance information was distributed supported modest increases in use in LHDs. Because use of syndromic surveillance data at smaller LHDs is rare, these modest increases are important indicators of effective modification of the NC syndromic surveillance system.     

Using Information Entropy to Monitor Chief Complaint Characteristics and Quality

Objective

We describe how entropy, a key information measure, can be used to monitor the characteristics of chief complaints in an operational surveillance system.

Introduction

Health care processes consume increasing volumes of digital data. However, creating and leveraging high quality integrated health data is challenging because large-scale health data derives from systems where data is captured from varying workflows, yielding varying data quality, potentially limiting its utility for various uses, including population health. To ensure accurate results, it’s important to assess the data quality for the particular use. Examples of sub-optimal health data quality abound: accuracy varies for medication and diagnostic data in hospital discharge and claims data; electronic laboratory data used to identify notifiable public-health cases shows varying levels of completeness across data sources; data timeliness has been found to vary across different data sources. Given that there is clear increasing focus on large health data sources; there are known data quality issues that hinder the utility of such data; and there is a paucity of medical literature describing approaches for evaluating these issues across integrated health data sources, we hypothesize that novel methods for ongoing monitoring of data quality in rapidly growing large health data sets, including surveillance data, will improve the accuracy and overall utility of these data.

Methods

Our analysis used chief complaint data derived from the original real-time HL7 registration transactions for ED encounters over a 3-year study period between January 1, 2008 and December 30, 2010 from over 100 institutions participating in the Indiana Public Health Emergency Surveillance System (PHESS) [1]. We used the following syndrome categories based on various definitions: respiratory, influenza like illness, gastrointestinal, neurological, undifferentiated infection, skin, and lymphatic. We calculated entropy for chief complaint data [2]. Entropy measures uncertainty and characterizes the density of the information contained in a message, commonly measured in bits. We analyzed entropy stratified a) by syndrome category, b) by syndrome category and time, and c) by syndrome category, time, and source institution.

Results

Analysis of more than 7.4 million records revealed the following: First, overall information content varied by syndrome, with “neurological” showing greatest entropy (Figure 1). Second, entropy measures followed consistent intraorganizational trends: information content varied less within an organization than across organizations (Figure 2). Third, information entropy enables detection of otherwise unannounced changes in system behavior. Figure 3 illustrates the monthly entropy measures for the respiratory syndrome from 5 sources over 36 months. One source changed registration software. Their visit volume didn’t change, but the information content of the chief complaint changed, as indicated by a substantial shift in entropy.Open in a separate windowFigure 1:Entropy (bits) for chief complaints classified into specific syndrome categories.Open in a separate windowFigure 2:Entropy (bits) for chief complaints classified into specific syndrome categories stratified by source system for 10 high-volume emergency departments.Open in a separate windowFigure 3:Monthly entropy (bits) for chief complaints classified into specific syndrome categories, stratified by source system for 5 high-volume emergency departments. Note the shift in values for one source that changed registration software.

Conclusions

As we face greater data volumes, methods assessing the data quality for particular uses, including syndrome surveillance, are needed. This analysis shows the value of entropy as a metric to support monitoring of surveillance systems. Future work will refine these measures and further assess the inter-organizational variations of entropy.     

Usefulness of Syndromic Surveillance for Early Outbreak Detection in Small Islands: The Case of Mayotte

Objective

To present the usefulness of syndromic surveillance for the detection of infectious diseases outbreak in small islands, based on the experience of Mayotte.

Introduction

Mayotte Island, a French overseas department of around 374 km2 and 200 000 inhabitants is located in the North of Mozambique Channel in the Indian Ocean (Figure 1).Open in a separate windowFigure 1Map of the western Indian Ocean featuring Mayotte IslandIn response to the threat of the pandemic influenza A(H1N1)2009 virus emergence, a syndromic surveillance system has been implemented in order to monitor its spread and its impact on public health (1). This surveillance system which proved to be useful during the influenza pandemic, has been maintained in order to detect infection diseases outbreaks.

Methods

Data are collected daily directly from patients’ computerized medical files that are filled in during medical consultations at the emergency department (ED) of the hospital Center of Mayotte (2). Among the collected variables, the diagnosis coded according to ICD-10 is used to categorize the syndromes. Several syndromes are monitored including the syndromic grouping for conjunctivitis and unexplained fever.For early outbreak detection, a control chart is used based on an adaptation of the Cusum methods developed by the CDC within the framework of the EARS program (3).

Results

Each week, about 700 patients attend the ED of the hospital. The syndromic surveillance system allowed to detect an outbreak of conjunctivitis from week 10 (Figure 2). During the epidemic peak on week 12, conjunctivitis consultations represented 5% of all consultations. The data of the sentinel practitioner network confirmed this epidemic and the laboratory isolated Enterovirus (4). At the same time, an unusual increase of unexplained fever was detected.Open in a separate windowFigure 2Weekly number of conjonctivitis and unexplained fever consultations and statistical alarms detected

Conclusions

Due to its geographical and socio-demographical situation, the population of Mayotte is widely exposed to infectious diseases. Even on a small island, syndromic surveillance can be useful to detect outbreak early leading to alerts and to mobilize a rapid response in addition to others systems.     

Syndromic Surveillance Based on Emergency Visits: A Reactive Tool for Unusual Events Detection

Objective

To show with examples that syndromic surveillance system can be a reactive tool for public health surveillance.

Introduction

The late health events such as the heat wave of 2003 showed the need to make public health surveillance evolve in France. Thus, the French Institute for Public Health Surveillance has developed syndromic surveillance systems based on several information sources such as emergency departments (1). In Reunion Island, the chikungunya outbreak of 2005–2006, then the influenza pandemic of 2009 contributed to the implementation and the development of this surveillance system (2–3). In the past years, this tool allowed to follow and measure the impact of seasonal epidemics. Nevertheless, its usefulness for the detection of minor unusual events had yet to be demonstrated.

Methods

In Reunion Island, the syndromic surveillance system is based on the activity of six emergency departments. Two types of indicators are constructed from collected data:

• - Qualitative indicators for the alert (every visit whose diagnostic relates to a notifiable disease or potential epidemic disease);
• - Quantitative indicators for the epidemic/cluster detection (number of visits based on syndromic grouping).

Daily and weekly analyses are carried out. A decision algorithm allows to validate the signal and to organize an epidemiological investigation if necessary.

Results

Each year, about 150 000 visits are registered in the six emergency departments that is 415 consultations per day on average. Several unusual health events on small-scale were detected early.In August 2011, the surveillance system allowed to detect the first autochthonous cases of measles, a few days before this notifiable disease was reported to health authorities (Figure 1). In January 2012, the data of emergency departments allowed to validate the signal of viral meningitis as well as to detect a cluster in the West of the island and to follow its trend. In June 2012, a family foodborne illness was detected from a spatio-temporal cluster for abdominal pain by the surveillance system and was confirmed by epidemiological investigation (Figure 2).Open in a separate windowFigure 1Epidemic curve of measles casesOpen in a separate windowFigure 2Line-list of patient characteristics in an abdominal pain cluster.

Conclusions

Despite the improvement of exchanges with health practitioners and the development of specific surveillance systems, health surveillance remains fragile for the detection of clusters or unusual health events on small scale. The syndromic surveillance system based on emergency visits has proved to be relevant for the identification of signals leading to health alerts and requiring immediate control measures. In the future, it will be necessary to develop these systems (private practitioners, sentinel schools) in order to have several indicators depending on the degree of severity.     

Evaluating a Social Network Analytic Tool to Support Outbreak Management and Contact Tracing in an Outbreak of Pertussis

Objective

To determine the feasibility and value of a social network analysis tool to support pertussis outbreak management and contact tracing in the state of Utah.

Introduction

Pertussis (i.e., whooping cough) is on the rise in the US. To implement effective prevention and treatment strategies, it is critical to conduct timely contact tracing and evaluate people who may have come into contact with an infected person. We describe a collaborative effort between epidemiologists and public health informaticists at the Utah Department of Health (UDOH) to determine the feasibility and value of a network-analytic approach to pertussis outbreak management and contact tracing.

Methods

The partnership: In early 2012, epidemiologists from UDOH’s Vaccine Preventable Disease Program and UDOH’s public health informaticists formed a partnership to determine the feasibility and value of the Organizational Risk Analyzer (ORA) in pertussis outbreak management and contact tracing (1). Both entities have a longstanding partnership. A characteristic that has made the collaboration particularly strong and mutually beneficial is that both partners have expertise in disease surveillance and outbreak management. In addition, the informaticists have expertise in devising systems that help frontline healthcare providers.The Organizational Risk Analyzer (ORA): ORA is a computational tool that extends network analysis by using a meta-matrix model. A meta-matrix is defined as a network of connecting entities. The tool uses one or more matrices in an organization’s meta-matrix as input. From this input the tool calculates measures that describe the relationships and ties among the entities. ORA contains over 50 network and node level measures which are categorized by the type of risk they detect (1).Procedures: Following approval from UDOH’s Institutional Review Board, we analyzed records from 629 deidentified pertussis patients from the UT-NEDSS database from January 2011 to December 2011. The test data included demographics and epidemiological information. We used Excel to create .csv data files, uploaded the data into ORA, and displayed the data in meta-matrices consisting of nodes (cases/contacts) and edges (relationships). We used ORA’s visualizer to check for data-entry errors before performing the network analysis.Data Analysis: ORA’s centrality measures (degree, closeness, betweenness, hub, and eigenvector) were used to identify geographic locations with high infection rates and the patients who were central to sustaining the outbreak. Next, we applied a concor algorithm to find groups in the meta-network that might be hard to spot visually. Visualizations were used to supplement the metrics.

Results

The ORA analysis identified 5 individuals who were central to perpetuating the outbreak in that their centrality measures were higher than other patients in the network. The index patient (Fig 1) was traced back to Utah County and was linked to 6 direct contacts in the same county and several indirect ties in adjacent counties. The individual was highly connected to others within the network (hub centrality = 1.41 and eigenvector centrality = 1.00). Salt Lake County had the highest number of cases, followed by Utah County and Weber County. The concor analysis revealed hidden networks, including a cluster of patients grouped by age group and case status (Fig 2).Open in a separate windowFig. 1The index case.Open in a separate windowFig. 2Concor cluster of patients, by age group and case status.

Conclusions

The ORA was found to be a valuable tool for supporting pertussis outbreak management and contact tracing. Although network analysis is relatively new to public health, it can increase public health’s understanding of how patterns of social relationships can aid or inhibit the spread of communicable diseases and provide the information needed to target intervention efforts effectively.     

Tracking Drug Overdose Trends in Ohio using ED Chief Complaints

Objective

Preliminary analysis was completed to define, identify, and track the trends of drug overdoses (OD), both intentional and unintentional, from emergency department (ED) and urgent care (UC) chief complaint data.

Introduction

The State of Ohio, as well as the country, has experienced an increasing incidence of drug ODs over the last three decades [3]. Of the increased number of unintended drug OD deaths in 2008, 9 out of 10 were caused by medications or illicit drugs [1]. In Ohio, drug ODs surpassed MVCs as the leading cause of injury death in 2007. This trend has continued through the most current available data [3]. Using chief complaint data to quickly track changes in the geographical distribution, demographics, and volume of drug ODs may aid public health efforts to decrease the number of associated deaths.

Methods

Chief complaint data from ED/UC visits were collected and analyzed from Ohio’s syndromic surveillance application for 2010–2012. Ninety-six percent of all Ohio ED visits were captured during this timeframe. Due to the nonspecific nature of chief complaints as well as the lack of detail given upon registration at the ED/UC, attempting to separate visits into intentional vs. unintentional was not feasible. Therefore, a fairly specific classifier was created to define all potential ED/UC visits related to drug ODs. The data were analyzed, using SAS v 9.3, via time series analyses, and stratified by age, gender, and geographic region. Although these data types are pre-diagnostic in nature, they are more readily accessible than discharge data.

Results

On average, Ohio observed approx 66 ED/UC visits per day related to drug ODs from 2010–2012. The data show an increasing trend from 2010 through 2012 as well as a slight seasonal trend with higher visits observed in the spring/summer months as opposed to the autumn/winter months (Figure 1). The data showed that females attributed to a higher frequency of the drug ODs than males by approximately 4 ED/UC visits per day. Other data sources show a higher incidence in males than females related to unintentional drug ODs [3]. The highest age category attributing to the increase was the 18–39 years of age for both males and females, as shown in Figure 2. Population rates were calculated to identify those counties most affected by drug ODs. The data showed the highest rates of ED/UC visits related to drug ODs to be found in mostly rural areas of Ohio.Open in a separate windowFigure 1ED Visits Related to Drug Overdoses by Day, Ohio, 2010–12Open in a separate windowFigure 2ED Visits Related to Drug Overdoses by Age Group, Ohio, 2010–12

Conclusions

The annual death rate from unintentional drug poisonings by Ohio residents has increased from 3.6 in 2000 to 13.4 per 100,000 population in 2010[3]. As a result, the Ohio Governor created a Drug Abuse Task Force in 2009[4]. Ohio legislation (HB 93) implemented a prohibition on the operation of pain management clinics without a license on June 19, 2011[3]. According to this preliminary analysis, ED/UC visits related to drug ODs 1 year post-implementation of HB 93 continue to increase. It is unclear if HB 93 has slowed the rate of increase. Additionally, pre-diagnostic data has significant limitations including the significant possibility of misclassifying non-OD patient encounters as ODs. Further study of post-diagnostic data to confirm these trends is warranted.     

Increasing Mild Enterovirus Cases Provides An Important Signal of Up-coming Trends in Elevating Severe Enterovirus Cases

Objective

This study was to elucidate the spatio-temporal correlations between the mild and severe enterovirus cases through integrating enterovirus-related three surveillance systems in Taiwan. With these fully understanding epidemiological characteristics, hopefully, we can develop better measures and indicators from mild cases to provide early warning signals and thus minimizing subsequent numbers of severe cases.

Introduction

In July 2012, the 54 children infected with enterovirus-71(EV-71) were died in Cambodia [1]. The media called it as mystery illness and made Asian parents worried. In fact, the severe epidemics of enterovirus occurred frequently in Asia, including Malaysia, Singapore, Taiwan and China [2]. The clinical severity varied from asymptomatic to mild (hand-foot-mouth disease and herpangina) and severe pulmonary edema/hemorrhage and encephalitis [3]. Up to now, the development of vaccine for EV-71 and the more effective antiviral drug was still ongoing [4]. Therefore, surveillance for monitoring the enterovirus activity and understanding the epidemiological characteristics between mild and severe enterovirus cases was crucial.

Methods

Three main databases including national notifiable diseases surveillance, sentinel physician surveillance and laboratory surveillance from July 1, 1999 to December 31, 2008 were analyzed. The Pearson’s correlation coefficient was applied for measuring the consistency of the trend. The Poisson space-time scan statistic [5] was used for identifying the most likely clusters. We used GIS (ArcMap, version9.0; ESRI Inc.,Redlands, CA, USA) for visualization of detected clusters.

Results

Temporal analysis found that the Pearson’s correlation between mild EV cases and severe EV cases occurring in the same week was 0.553 (p<0.01) in Figure 1. Such a correlation became moderate (data) when mild EV cases happened in 1∼4 weeks before the current severe EV cases. Among the 1,517 severe EV cases notified to Taiwan CDC during the study period, the mean age was 27 months, 61.4% was male and 12% were fatal. These severe EV cases were significantly associated with the positive isolation rate of EV-71, with much higher correlation than the mild cases [ 0.498 p<0.01 vs. 0.278, p<0.01]. Using the space-time cluster method, we identified three possible clusters in June 2008 in six cities/counties (Figure 2).Open in a separate windowFigure 1The temporal trend between mild and severe EV cases.Open in a separate windowFigure 2The spatio-temporal clusters of EV severe cases.

Conclusions

Taiwan’s surveillance data indicate that local public health professionals can monitor the trends in the numbers of mild EV cases in community to provide early warning signals for local residents to prevent the severity of future waves.     

Using Syndromic Emergency Department Data to Augment Oral Health Surveillance

Objective

To utilize an established syndromic reporting system for surveillance of potentially preventable emergency department (ED) oral health visits (OHV) in New York City (NYC).

Introduction

NYC Department of Health and Mental Hygiene recently reoriented its oral health care strategy to focus on health promotion and expanded surveillance. One surveillance challenge is the lack of timely OHV data; few dental providers are in our electronic health record project, and statewide utilization data are subject to delays. Prior research has examined OHV using ICD-9-CM from ED records, and has suggested that diagnostic specificity may be limited by ED providers’ lack of training in dental diagnoses (1–3). We considered our existing ED syndromic system as a complement to periodic population-based surveys. This system captures approximately 95% of all ED visits citywide; 98% of records have a completed chief complaint text field whereas only 52% contain an ICD-9-CM diagnosis.

Methods

We used chief complaint text to define OHV in two ways: (1) a basic definition comprised of ‘TOOTH’ or ‘GUM’ in combination with a pain term (e.g., ‘ACHE’); (2) a more inclusive definition of either specific oral health diagnoses (e.g., ‘PULPITIS’) or definition (1). For both definitions, we excluded visits likely to have stemmed from trauma (e.g., ‘ACCIDENT’). Data from 2009–2011 were analyzed by facility, patient age and residential zip code, and day/time using SAS v9.2 (SAS Institute; Cary, NC).

Results

OHV in 2009–2011 totaled 72,410 (def. 1) and 103,594 (def. 2), or 0.6% and 0.9% of all ED visits, respectively. OHV (def. 2) spiked at age 18 and were highest among 18 to 29 year olds (Fig. 1). Neighborhood OHV rates (def. 2) ranged from 74 to 965 per 100,000 persons. 59% of OHV occurred between 8am and 6pm (Fig. 2). Highly specific dental conditions were rare; terms such as “tooth ache” were most common.Open in a separate windowFig 1OHV (def.2) by age, 2009–2011Open in a separate windowFig 2OHV (def.2) by day/time, 2009–2011

Conclusions

Findings suggest that OHV are a particular problem among ages 18 to 29. This pattern may reflect lower insurance coverage among young adults. The proportion of daytime visits suggests that EDs are substituting for regular dental treatment and there may be opportunities to promote daytime linkages to office-based dental providers.A well-established syndromic reporting system holds promise as a method of OHV surveillance. Strengths include near complete chief complaint reporting, rapid availability, and the potential to identify populations and facilities that could benefit from expanded access and preventive education. Limitations include the need to gather site-specific facility information (e.g., presence of dental residents, coding practices) to better understand patterns. Also, the absence of some important fields in the syndromic system (e.g., insurance coverage, income) limit assessment of the degree to which cost barriers may be driving OHV.     

Open Source Health Intelligence (OSHINT) for Foodborne Illness Event Characterization

Objective

We propose a cloud-based Open Source Health Intelligence (OS-HINT) system that uses open source media outlets, such as Twitter and RSS feeds, to automatically characterize foodborne illness events in real-time. OSHINT also forecasts response requirements, through predictive models, to allow more efficient use of resources, personnel, and countermeasures in biological event response.

Introduction

An increasing amount of global discourse reporting has migrated to the online space, in the form of publicly accessible social media outlets, blogs, wikis, and news feeds. Social media also presents publicly available and highly accessible information about individual, real-time activity that can be leveraged to detect, monitor, and more efficiently respond to biological events.

Methods

Salmonella and Escherichia Coli (E. coli) events were selected based on the magnitude and number of reported outbreaks to the Centers for Disease Control (CDC) in the last ten years (1). These events affect multiple states and were large enough to ensure appropriate confidence levels when developing response metrics obtained from our prediction models. We collected social media data between 2006 – 2012 due to the emergence of Twitter, Facebook, and other social media utilization during this time period.Characterization is defined as the process of identifying specific event features that inform overall situational awareness. The number hospitalized, dead, or injured, in addition to patient demographics and symptoms were determined to be useful for our characterization and forecast event metrics. Analytical methods, such as term-frequency-inverse document frequency (TF-IDF), natural language processing (NLP), and information extraction, were used to characterize events according to our metrics. Lexicon development, during NLP implementation, was generated from online news articles used to describe the events. Lastly, forecasting algorithms were developed to predict the potential response based on similar historical events that were initially characterized by our information extraction algorithms.

Results

The OSHINT system was developed in Amazon Web Services and includes real-time social media collection for event characterization (see Figure 1). OSHINT currently characterizes number of victims ill, hospitalized, and dead due to foodborne illness events.Open in a separate windowFigure 1OSHINT System in Amazon Web Services.OSHINT was used to characterize the recent national 2012 Salmonella event related to cantaloupes, during which OSHINT characterized social media posts related to the event, as news articles and Twitter tweets streamed into the system (Figure 2). On August 17, 2012 the OSHINT system identified a large increase in Twitter tweets mentioning salmonella. Social media data found absent (victims missing work or school day), death, hospital, and sick events to involve 2, 4, 17, 283 media mentions, respectively. Our TF-IDF algorithm characterized the salmonella event impact as two dead and 150 sickened by salmonella-tainted cantaloupe. Retrospective analysis of CDC reported data on August 30, 2012 indicated the salmonella event involved two deaths in 204 cases (2).Open in a separate windowFigure 2:2012 Salmonella Outbreak in Cantaloupe

Conclusions

The OSHINT team is continually developing and refining characterization and forecasting algorithms used in the system. Upon completion, OSHINT will characterize symptoms, geography, and demographics for E. coli and Salmonella events. The system will also forecast number sick, dead, and hospitalized for an effective and quick response. We will refine our algorithms and evaluate the system against past and future events to provide confidence in our results.     

Statistical Models for Biosurveillance of Multiple Organisms

Objective

To look at the diversity of the patterns displayed by a range of organisms, and to seek a simple family of models that adequately describes all organisms, rather than a well-fitting model for any particular organism.

Introduction

There has been much research on statistical methods of prospective outbreak detection that are aimed at identifying unusual clusters of one syndrome or disease, and some work on multivariate surveillance methods (1). In England and Wales, automated laboratory surveillance of infectious diseases has been undertaken since the early 1990’s. The statistical methodology of this automated system is described in (2). However, there has been little research on outbreak detection methods that are suited to large, multiple surveillance systems involving thousands of different organisms.

Methods

We obtained twenty years’ data on weekly counts of all infectious disease organisms reported to the UK’s Health Protection Agency. We summarized the mean frequencies, trends and seasonality of each organism using log-linear models. To identify a simple family of models which adequately represents all organisms, the Poisson model, the quasi-Poisson model and the negative binomial model were investigated (3,4). Formal goodness-of-fit tests were not used as they can be unreliable with sparse data. Adequacy of the models was empirically studied using the relationships between the mean, variance and skewness. For this purpose, each data series was first subdivided into 41 half-years and de-seasonalized.

Results

Trends and seasonality were summarized by plotting the distribution of estimated linear trend parameters for 2250 organisms, and modal seasonal period for 2254 organisms, including those organisms for which the seasonal effect is statistically significant.Relationships between mean and variance were summarized as given in Figure 1.Open in a separate windowFigure 1Relationships between mean and variance. (top) Histogram of the slopes of the best fit lines for 1001 organisms; the value 1 corresponds to the quasi-Poisson model; (bottom) log of variance plotted against log of mean for one organism. The full line is the best fit to the points; the dashed line corresponds to the quasi-Poisson model; the dotted line corresponds to the Poisson model.Similar plots were used to summarize the relationships between mean and skewness.

Conclusions

Statistical outbreak detection models must be able to cope with seasonality and trends. The data analyses suggest that the great majority of organisms can adequately – though far from perfectly – be represented by a statistical model in which the variance is proportional to the mean, such as the quasi-Poisson or negative binomial models.     

Searching for Complex Patterns Using Disjunctive Anomaly Detection

Objective

Disjunctive anomaly detection (DAD) algorithm [1] can efficiently search across multidimensional biosurveillance data to find multiple simultaneously occurring (in time) and overlapping (across different data dimensions) anomalous clusters. We introduce extensions of DAD to handle rich cluster interactions and diverse data distributions.

Introduction

Modern biosurveillance data contains thousands of unique time series defined across various categorical dimensions (zipcode, age groups, hospitals). Many algorithms are overly specific (tracking each time series independently would often miss early signs of outbreaks), or too general (detections at state level may lack specificity reflective of the actual process at hand). Disease outbreaks often impact multiple values (disjunctive sets of zipcodes, hospitals, multiple age groups) along subsets of multiple dimensions of data. It is not uncommon to see outbreaks of different diseases occurring simultaneously (e.g. food poisoning and flu) making it hard to detect and characterize the individual events.We proposed Disjunctive Anomaly Detection (DAD) algorithm [1] to efficiently search across millions of potential clusters defined as conjunctions over dimensions and disjunctions over values along each dimension. An example anomalous cluster detectable by DAD may identify zipcode = {z1 or z2 or z3 or z5} and age_group = {child or senior} to show unusual activity in the aggregate. Such conjunctive-disjunctive language of cluster definitions enables finding real-world outbreaks that are often missed by other state-of-art algorithms like What’s Strange About Recent Events (WSARE) [3] or Large Average Submatrix (LAS) [2]. DAD is able to identify multiple interesting clusters simultaneously and better explain complex anomalies in data than those alternatives.

Methods

We define the observed counts of patients reporting on a given day as a random variable for each unique combination of values along all dimensions. DAD iteratively identifies K subsets of these variables along with corresponding ranges of their values and time intervals that show increased activity that cannot be explained by random fluctuations (K is generally unknown and could be 0). The resulting set of clusters maximizes data likelihood while controlling for overall complexity. We have successfully derived a versatile set of scoring functions that allow Normal, Poisson, Exponential or Non-parametric assumptions about the underlying data distributions, and accommodate additive-scaled, additive-unscaled or multiplicative-scaled models for the clusters.

Results

We present results of testing DAD on two real-world datasets. One of them contains daily outpatient visit counts from 26 regions in Sri Lanka involving 9 common diseases. The other data contains semi-synthetically generated terrorist activities throughout regions of Afghanistan (Sigacts). Both span multiple years and are representative of data seen in biosurveillance applications.Figure 1 shows DAD systematically outperforming WSARE and LAS. Each algorithm’s parameters were tuned to generate one false positive per month in baseline data. The graphs represent average days-to-detect performance of 100 sets with synthetically injected clusters using additive-scaled (AS), additive-unscaled (AU), and multiplicative-scaled (MS) models of cluster interactions.Open in a separate windowFigure 1:Alg. performance (a) Srilanka, (b) Sigacts

Conclusions

We extend applicability of DAD algorithm to handle wide variety of input data distributions and various outbreak models. DAD efficiently scans over millions of potential outbreak patterns and accurately and timely reports complex outbreak interactions with speed that meets requirements of practical applications.     

Refactoring INDICATOR into an Advanced Information System for One Health Monitoring

Objective

To redesign INDICATOR for One Health, establish a common data format, and provide for long term scalability.

Introduction

INDICATOR is a multi-stream open source platform for biosurveillance and outbreak detection, currently focused on Champaign County in Illinois[1]. It has been in production since 2008 and is currently receiving data from emergency departments, patient advisory nurse call center, outpatient convenient care clinic, school absenteeism, animal control, and weather sources. Long term scalability was however compromised during the 2009 H1N1 influenza pandemic as immediate public health needs took priority over our systematic development plan. With the impending addition of veterinary clinic data and recognizing that the health of a community also depends on animal and environmental factors, we decided to revisit the INDICATOR architecture and redesign it to be a more holistic and scalable system. We also decided to revisit the data submission format, keeping in line with the philosophy of making opportunistic secondary use of as much data about the health of a community that we can obtain.

Methods

Following a formal evaluation of the existing production version of INDICATOR we established the systems architecture shown in Figure 1 to leverage work in other cyberinfrastructure projects at NCSA.Open in a separate windowFigure 1INDICATOR system architecture

Results

We have now implemented the back end changes, including unifying the multiple physical MySQL database systems and multiple Apache Tomcat application engines into a single system. A web application, using service oriented principles and the GWT library, has been developed that can query and display the newly unified data and provide new options for input of data to the system.In order to streamline and simplify the data format we decided to define a single format that can be used by different kinds of healthcare providers, both human and veterinary. Although we recognize the limitations in this approach we define a reported event to be a simple what, when, and where containing the following seven fields or the relevant subset based roughly on the ISDS meaningful use recommendations [2]

1. Date of incident
2. ICD-9 code for the primary diagnosis
3. Free text of the diagnosis (not the text definition of the ICD-9 code)
4. Text chief complaint at triage
5. Location
6. Count
7. Species

In this way we can handle, in a single format, data from emergency departments, convenient care clinics, patient advisory nurse call centers, veterinary clinics, veterinary labs, and veterinary poison control centers.

Conclusions

INDICATOR has been significantly redesigned and is now more integrated, scalable, and secure. It is now placed to become a One Health integrated monitoring system.     

Use of Control Bar Matrix for Outbreak Detection in Syndromic Surveillance System

Objective

To develop and test the method of incorporating different control bars for outbreak detection in syndromic surveillance system.

Introduction

Aberration detection methods are essential for analyzing and interpreting large quantity of nonspecific real-time data collected in syndromic surveillance system. However, the challenge lies in distinguishing true outbreak signals from a large amount of false alarm (1). The joint use of surveillance algorithms might be helpful to guide the decision making towards uncertain warning signals.

Methods

A syndromic surveillance project (ISSC) has been implemented in rural Jiangxi Province of China since August 2011. Doctors in the healthcare surveillance units of ISSC used an internet-based electronic system to collect information of daily outpatients, which included 10 infectious related symptoms. From ISSC database, we extracted data of fever patients reported from one township hospital in GZ town between August 1st and December 31st, 2011 to conduct an exploratory study. Six different control bar algorithms, which included Shewart, Moving Average (MA), Exponentially Weighted Moving Average (EWMA) and EARS’ C1, C2, C3, were prospectively run among historical time series of daily fever count to simulate a real-time outbreak detection. Each control bar used 7 days’ moving baseline with a lag of 2 days [the baseline for predicting Day(t) starts from Day(t-9) to Day(t-3), C1 method used a lag of zero day]. We set the threshold of μ+2σ for Shewart and MA, and 2.1 for EWMA C1, C2 and C3. An alarm was triggered when the observed data exceeded threshold, and the detailed information of each patient were checked for signal verification. Microsoft Excel 2007 was used to calculate the simulation results.

Results

During the 5 months, GZ township hospital reported 514 outpatients with fever symptom, with an average of 3.4 per day. All control bars were simultaneously operated among daily counts of fever cases. Of the 153 days on surveillance, 29 triggered alarms by at least one of the control bars. Nine days triggered alarms from >= 3 control bars while on one day (12/30) all 6 algorithms raised alarms. Figure 1 shows the date, fever count, algorithm and warning level (color) of each alarm, which we called a control bar matrix. It can be seen that C3 and EWMA present a higher sensitiveness towards tiny data change whereas C1, C2 and MA focus on large increase of data. C3 also had a memory effect on recent alarms. No infectious disease epidemic or outbreak event was confirmed within the signals. Most fever patients on the nine high-warning days (red and purple) were diagnosed as upper level respiratory infection. However, we discovered that the sharp increase of fever cases on 12/30 was attributed to 5 duplicate records mistakenly input by the staff in GZ hospital.Open in a separate windowFigure 1:Detailed information of alarm signals generated by control bar matrix (No-alarm days were omitted).

Conclusions

By combining control bars with different characteristics, the matrix has potential ability to improve the specificity of detection while maintaining a certain degree of sensitivity. With alarms categorized into hierarchical warning levels, public health staffs can decide which alarm to investigate according to the required sensitivity of surveillance system and their own capacity of signal verification. Though we did not find any outbreak event in the study, the possibility of localized influenza epidemic on high-warning days cannot be wiped out, and the matrix’s ability to detect abnormal data change was apparent. The proper combination, baseline and threshold of control bars will be further explored in the real-time surveillance situation of ISSC.     

Multiagent Simulation of the Hepatitis B Epidemic Process

Objective

To develop multiagent model of hepatitis B (HBV) infection spreading.

Introduction

The standard approaches to simulation include solving of differential equation systems. Such approach is good for obtaining general picture of epidemics (1, 2). When the detailed analysis of epidemics reasons is needed such model becomes insufficient. To overcome the limitations of standard approaches a new one has been offered. The multiagent approach has been offered to be used for representation of the society. Methods of event-driven programming give essential benefits of the processing time of the events (3).

Methods

For model development C# computing language has been used. We have used demographical data, the incidence rate of HBV infection of all population and different population groups (age, professional and other groups), coverage of hepatitis B vaccination, the proportion of HBV carriers in population, the prevalence rate of chronic HBV infection, percent of dominated transmission routes and factors and other rates in Kharkiv region. All parameters, expressed in the model were estimated using sero-surveys data and data of epidemiological surveillance of Kharkiv region sanitary-epidemiological station. Also the theoretical knowledge about HBV infection has been used. 26 conditions have been derived from the problem domain. The transition from one condition to another depends on stochastic value and time of the event change. All events are organized in priority queue which results in high rate of computation performance. The dependence on time and random value determines automata theory conceptions.

Results

The prototype of software system, which includes a subsystem of the multiagent simulation and specialized statistical and mathematical sub-system which can process the simulation results and perform a conditional optimization of the selected objective functions (morbidity, the effectiveness of specific preventive and control activities and their price, measure of reducing the socio-economic impact of HBV infection, etc.) have been developed. Screen form is presented in Figure 1.Open in a separate windowFig. 1.The main panel of simulation management and graphic visualization.

Conclusions

The multiagent simulation model of the HBV infection epidemic situation development, based on data obtained in Kharkiv (Ukraine) has been created.The simulation results allow us to:

1. predict the dynamics of the epidemic process in time in a particular area, taking into account specific epidemic situation;test the effectiveness of various preventive measures (sterilization of instruments, coverage of hepatitis B vaccination of certain groups of people, etc.).

Using the present model in public health system suggests improvment of the epidemiological diagnostics of HBV infection and of the quality of management decisions about epidemiological surveillance. The evolution of multiagent simulation in epidemiology will broaden the possibilities of epidemiological surveillance and control.     

Using Google Dengue Trends to Estimate Climate Effects in Mexico

Objective

To evaluate the association between Dengue Fever (DF) and climate in Mexico with real-time data from Google Dengue Trends (GDT) and climate data from NASA Earth observing systems.

Introduction

The incidence of dengue fever (DF) has increased 30 fold between 1960 and 2010 [1]. The literature suggests that temperature plays a major role in the life cycle of the mosquito vector and in turn, the timing of DF outbreaks [2]. We use real-time data from GDT and real-time temperature estimates from NASA Earth observing systems to examine the relationship between dengue and climate in 17 Mexican states from 2003–2011. For the majority of states, we predict that a warming climate will increase the number of days the minimum temperature is within the risk range for dengue.

Methods

The GDT estimates are derived from internet search queries and use similar methods as those developed for Google Flu Trends [3]. To validate GDT data, we ran a correlation between GDT and dengue data from the Mexican Secretariat of Health (2003–2010). To analyze the relationship between GDT and varying lags of temperature, we constructed a time series meta-analysis. The mean, max and min of temperature were tested at lags 0 –12 weeks using data from the Modern Era Retrospective-Analysis for Research and Applications. Finally, we built a binomial model to identify the minimum 5° C temperature range associated with a 50% or higher Dengue activity threshold as predicted by GDT.

Results

The time series plot of GDT data and the Mexican Secretariat of Health data (2003– 2010) (Figure 1) produced a correlation coefficient of 0.87. The time series meta-analysis results for 17 states showed an increase in minimum temperature at lag week 8 had the greatest odds of dengue incidence, 1.12 Odds Ratio (1.09–1.16, 95% Confidence Interval). The comparison of dengue activity above 50% in each state to the minimum temperature at lag week 8 showed 14/17 states had an association with warmest 5 degrees of the minimum temperature range. The state of Sonora was the only state to show an association between dengue and the coldest 5 degrees of the minimum temperature range.Open in a separate windowFigure 1Time Series Correlation: Google Dengue Trends vs. Secretariat of Health, Mexico 2003–2010

Conclusions

Overall, the incidence data from the Mexican Secretariat of Health showed a close correlation with the GDT data. The meta-analysis indicates that an increase in the minimum temperature at lag week 8 is associated with an increased dengue risk. This is consistent with the Colon-Gonzales et al. Mexico study which also found a strong association with the 8 week lag of increasing minimum temperature [4]. The results from this binomial regression show, for the majority of states, the warmest 5 degree range for the minimum temperature had the greatest association with dengue activity 8 weeks later. Inevitably, several other factors contribute to dengue risk which we are unable to include in this model [5]. IPCC climate change predictions suggest a 4° C increase in Mexico. Under such scenario, we predict an increase in the number of days the minimum temperature falls within the range associated with DF risk.     

An Integrated Syndromic Surveillance System for Monitoring Scarlet Fever in Taiwan

Objective

To develop an integrated syndromic surveillance system for timely monitoring and early detection of unusual situations of scarlet fever in Taiwan, since Hong Kong, being so close geographically to Taiwan, had an outbreak of scarlet fever in June 2011.

Introduction

Scarlet fever is a bacterial infection caused by group A streptococcus (GAS). The clinical symptoms are usually mild. Before October, 2007, case-based surveillance of scarlet fever was conducted through notifiable infectious diseases in Taiwan, but was removed later from the list of notifiable disease because of improved medical care capacities. In 2011, Hong Kong had encountered an outbreak of scarlet fever (1,2). In response, Taiwan developed an integrated syndromic surveillance system using multiple data sources since July 2011.

Methods

More than 99% of the Taiwan population is covered by National Health Insurance. We first retrospectively evaluated claims data from the Bureau of National Health Insurance (BNHI) by comparing with notifiable diseases reporting data from Taiwan Centers for Disease Control (TCDC). The claims data included information on scarlet fever diagnosis (ICD-9-CM code 034.1), date of visits, location of hospitals and age of patients from outpatient (OPD), emergency room (ER) and hospital admissions. Daily aggregate data of scarlet fever visits or hospitalizations were prospectively collected from BNHI since July 2011. Over 70% of the deaths in Taiwan are reported to the Office of Statistics of Department of Health electronically. We obtained daily data on electronic death certification data and used SAS Enterprise Guide 4.3 (SAS Institute Inc., Cary, NC, USA) for data management and analysis. Deaths associated with scarlet fever or other GAS infections were identified by text mining from causes of death with keywords of traditional Chinese ‘scarlet fever’, ‘group A streptococcus’ or ‘toxic shock syndrome’ (3).

Results

From January 2006 to September 2007, the monthly OPD data with ICD-9-CM code 034.1 from BNHI showed strong correlation with TCDC’s notifiable disease data (r=0.89, p<0.0001). From July 6, 2008 (week 28) through July 28, 2012 (week 30), the average weekly numbers of scarlet fever visits to the OPD, ER and hospital admissions were 37 (range 11–70), 7 (range 0–20) and 3 (range 0–9). Eighty-five percent of the scarlet fever patients were less than 10 years old. In Taiwan, scarlet fever occurred year-round with seasonal peaks between May and July (Fig. 1). From January 2008 to July 2012, we identified 12 potential patients (9 males, age range 0–82 years) who died of GAS infections. No report had listed ‘scarlet fever’ as cause of death during the study period.Open in a separate windowFig. 1.Weekly numbers of nationwide scarlet fever OPD and ER visits, and hospital admissions, with baseline OPD visits and 95% confidence interval calculated by a Serfling’s model, week 28 of 2008 to week 30 of 2012.

Conclusions

Taiwan has established an integrated syndromic surveillance system to timely monitor scarlet fever and GAS infection associated mortalities since July 2011. Syndromic surveillance of scarlet fever through BNHI correlated with number of scarlet fever cases through notifiable disease reporting system. Text mining from cause of death with the used keywords may have low sensitivities to identify patients who died of GAS infection. In Taiwan, syndromic surveillance has also been applied to other diseases such as enterovirus, influenza-like illness, and acute diarrhea. Interagency collaborations add values to existing health data in the government and have strengthened TCDC’s capacity of disease surveillance.     

Disease Mapping with Spatially Uncertain Data

Objective

Uncertainty regarding the location of disease acquisition, as well as selective identification of cases, may bias maps of risk. We propose an extension to a distance-based mapping method (DBM) that incorporates weighted locations to adjust for these biases. We demonstrate this method by mapping potential drug-resistant tuberculosis (DRTB) transmission hotspots using programmatic data collected in Lima, Peru.

Introduction

Uncertainty introduced by the selective identification of cases must be recognized and corrected for in order to accurately map the distribution of risk. Consider the problem of identifying geographic areas with increased risk of DRTB. Most countries with a high TB burden only offer drug sensitivity testing (DST) to those cases at highest risk for drug-resistance. As a result, the spatial distribution of confirmed DRTB cases under-represents the actual number of drug-resistant cases[1]. Also, using the locations of confirmed DRTB cases to identify regions of increased risk of drug-resistance may bias results towards areas of increased testing. Since testing is neither done on all incident cases nor on a representative sample of cases, current mapping methods do not allow standard inference from programmatic data about potential locations of DRTB transmission.

Methods

We extend a DBM method [2] to adjust for this uncertainty. To map the spatial variation of the risk of a disease, such as DRTB, in a setting where the available data consist of a non-random sample of cases and controls, we weight each address in our study by the probability that the individual at that address is a case (or would test positive for DRTB in this setting). Once all locations are assigned weights, a prespecified number of these locations (from previously published country-wide surveillance estimates) will be sampled, based on these weights, defining our cases. We assign these sampled cases to DRTB status, calculate our DBM, repeat this random selection and create a consensus map[3].

Results

Following [2], we select reassignment weights by the inverse probability of each untested case receiving DST at their given location. These weights preferentially reassign untested cases located in regions of reduced testing, reflecting an assumption that in areas where testing is common, individuals most at risk are tested. Fig. 1 shows two risk maps created by this weighted DBM, one on the unadjusted data (Fig.1, L) and one using the informative weights (Fig. 1, R). This figure shows the difference, and potentially the improvement, made when information related to the missingness mechanism, which introduces spatial uncertainty, is incorporated into the analysis.

Conclusions

The weighted DBM has the potential to analyze spatial data more accurately, when there is uncertainty regarding the locations of cases. Using a weighted DBM in combination with programmatic data from a high TB incidence community, we are able to make use of routine data in which a non-random sample of drug resistant cases are detected to estimate the true underlying burden of disease.Open in a separate window(L) Unweighted DBM of risk of a new TB case that received DST being positive for DRTB, compared to all new TB cases that received DST. (R) Weighted DBM of the risk of a new TB case that received DST being positive for DRTB, based on lab-confirmed DRTB cases and IPW selected non-DST TB cases, compared to all new TB cases.     

MoH+: A Global,Integrated, and Automated View of Official Outbreak Reporting

Objective

To introduce MoH+, HealthMap’s (HM) real-time feed of official government sources, and demonstrate its utility in comparing the timeliness of outbreak reporting between official and unofficial sources.

Introduction

Previous studies have documented significant lags in official reporting of outbreaks compared to unofficial reporting (1,2). MoH+ provides an additional tool to analyze this issue, with the unique advantage of actively gathering a wide range of streamlined official communication, including formal publications, online press releases, and social media updates.

Methods

Outbreaks reported by official sources were identified through MoH+ (healthmap.org/mohplus), which collects surveillance data published globally by ministries of health (MoH), other related ministries, government portals, government-affiliated organizations, and international governing bodies (Fig. 1). Reporting of these outbreaks was also identified in unofficial sources using various HM feeds including Google News, ProMED, and participatory surveillance feeds.Open in a separate windowFig. 1Interactive visualization of HealthMap MoH+, at healthmap.org/mohplusOf the 109 outbreaks identified since May 2012, 65 were excluded as they started before data collection, 7 were excluded as they were not reported by unofficial sources, and 1 was excluded as it was a non-natural outbreak. For the remaining 36 outbreaks, the median difference in first date of report between official and unofficial sources was analyzed using a Wilcoxon sign rank test.

Results

Outbreak reporting in official sources lagged by a statistically significant median of 2 days (p=0.003). Among unofficial sources, online news most often (75%) was the fastest to report an outbreak, followed by ProMED (22%) and participatory surveillance (3%). Among official sources, national government affiliated institutes were most often (41%) the fastest, and repeatedly providing prompt outbreak reports were the US Centers for Disease Control and Prevention (CDC), Public Health Agency of Canada, Finnish Food Safety Authority, Health Protection Scotland, UK Health Protection Agency, and French Institute of Public Health Surveillance (FIPHS). Following such institutes were the European CDC (ECDC) with 22% of first reports of outbreaks; MoH’s (17%); and WHO (10%). There were 4 instances in which official sources reported before unofficial sources—3 by the ECDC and 1 by FIPHS.

Conclusions

Compared to the Chan study reporting a 16 day lag between first public communication and WHO Outbreak News (1) and the Mondor study reporting a 10 day lag between non-government and government sources (2), the present study shows a much condensed lag of 2 days between unofficial and official sources. Because the two earlier studies cover a much broader historical time frame, one explanation for the reduced lag time is increased adoption of online communication by official government agencies. However, despite such improvements in communication, the lag persists, pointing to the importance of using informal sources for outbreak surveillance.The present study was limited by small sample size, as the study is in its early stages. We will continue to gather data and all numbers will be updated in time for the presentation to reflect the larger database. Future directions of this study include characterization of official and unofficial reporting by region, language, disease, and source.