The burden of infectious intestinal disease (IID) is considerable. The World Health Organization (WHO) estimated that foodborne disease from 22 pathogens accounted for 22 diseases resulted in 2 billion cases, over 1 million deaths, and 78.7 million disability-adjusted life years in 2010 [1]. The IID in the United Kingdom (IID2 study) [2] reported 274 cases per 1000 person-years, with 17.7 (95% CI 14.4-21.8) presenting to primary care. However, this may be an underestimate. Less restrictive – more representative (of coding practice) diagnostic criteria would greatly increase, for example, their estimate of norovirus by 26% to 59/1000 (95% CI 52.32-64.98) person-years equating to 3.7 (3.3-4.1) million infections annually [3].

Reported incidence rates for IID vary between 0.5% and 20% annually in the developed world [4-9]. Variation can be greatly attributed to underreporting and data types used to calculate rates [10]. Data used to report IID rates include: primary care records, hospital and other secondary care settings, prospective and retrospective surveys or questionnaires, notifications of disease to authorities, and reports of laboratory detection of pathogens [7,11,12]. Studies have concluded that approximately 1 in 20 IID patients in the community consult a general practitioner (GP) [7,13,14], hence incidence rates calculated based on primary care data are 0.5-3.3%—much lower than rates calculated with other methods [4,13,14].

Published variations may also be caused by imprecise or interchangeable use of the terms such as “diarrheal disease,” “acute infectious gastroenteritis” and “IID” and differing methods for describing cases, underscoring the importance of transparency when defining the disease [6,15]. The more general term “diarrheal disease” is used by the WHO and others in international public health as a symptom-based definition: infectious diarrhea and/or vomiting [6,11,16,17]. The terms “IID” and “acute gastroenteritis” tend to be more limited terms used to define patients with loose stools and/or vomiting for specific time periods and excluding chronic infections. Generally IID is defined as lasting less than 2 weeks, in the absence of known noninfectious causes, preceded by 2-3 symptom-free weeks [14,15]. Many studies list pathogens in their definition of IID or acute gastroenteritis; chronic or systemic conditions such as typhoid/paratyphoid and Helicobacter infections are often excluded [2,13].

Ontologies provide a method of systematically and transparently defining concepts and their relationships. They are used to clarify case finding and more accurately calculate disease incidence based on disease definitions that balance sensitivity and specificity [18,19]. In this study, we used a three-layer approach developed previously by the University of Surrey to develop an IID ontology [18]; we then used the ontological definition to calculate the incidence rate. The three-layer approach, an iterative process, includes development of disease concepts into an ontology, code collection, and logical data extraction [20].

UK general practice is highly computerized. Electronic registration–based systems ensure accurate denominators, and data from general practice provide opportunities for health research [21,22]. Most consultations are recorded on computers with key data—diagnosis, symptoms, investigative tests, and treatments—using a system called the Read codes [23]. The majority of UK practices are electronically connected to pathology laboratories, with generalized pathology results coded back into clinical records. Any laboratory results indicating pathogen detection should be coded directly by the clinician.

We aimed to test new technologies that provide general practitioners near real-time test results for a wide range of pathogens associated with IID [24]. We carried out this analysis to determine IID incidence from routine data using an ontological approach to make case finding more transparent and allow comparisons to other studies and data. We compared rates calculated using standard Royal College of General Practitioners (RCGP) Research and Surveillance Centre (RSC) and ICD-10 definitions with an ontological approach and reported impact on incidence rate.

We reviewed common IID case definitions published in the literature and standard coding systems used to record IID diagnoses in primary care settings and chose three IID definitions (Textbox 1).

In the United Kingdom, the RCGP RSC case definition used for calculating weekly incidence of IID for the RSC’s weekly communicable and respiratory disease report is the established “gold standard” for surveillance [14,25]. IID incidence rates for the RSC weekly report are generated using codes from the IID chapter of Read codes version 2 (5-byte set), the GP coding system most commonly used in primary care since 1985 to enter data into electronic health records. The RSC definition includes Read codes for conditions in WHO’s International Classification of Diseases ICD-8/9 infectious intestinal diseases chapter (A00-A09 codes) [25]. To maintain consistency while monitoring long-term year-over-year trends in infections and outbreaks, RSC has conducted IID surveillance following the ICD-9 infectious intestinal diseases chapter, and as a result, many conditions not currently included in the newer WHO ICD-10 definition of IID continued to be included in the weekly returns report after ICD-10 was released [25]. To examine coding differences and relationships, we mapped IID codes between three ICD classifications and back to RSC weekly report codes.

For the ontological case definition of IID, we selected the more restrictive, well-documented case definition used during the Second Study of Infectious Intestinal Disease in the Community (IID2 Study), an extensively published, longitudinal study of IID incidence carried out in UK primary care [2,14,26]. The study defines IID as an infectious intestinal condition always causing diarrhea and sometimes other symptoms such as vomiting or nausea lasting 2 weeks or less [26].

We used a three-level approach previously developed by the University of Surrey to establish an ontology based on IID case definition [20]. We formalized this ontology using Protégé, which is supported by grant GM10331601 from the National Institute of General Medical Sciences of the US National Institutes of Health [27].

The design of the ontology followed the structure used in problem-orientated records (POMR) and their associated coding system. This has its roots in the work of Lawrence Weed who created the idea of separating subjective (history) from objective (findings) and analysis (often diagnosis or problem) from plan (prescription or treatment). This is known internationally as Weed’s SOAP [28-30]. The classes in our ontology (Multimedia Appendix 1) broadly followed the components of SOAP: subjective (S), clinical features; objective (O), findings from laboratory tests, but could include objective clinical features such as fever if measured; analysis (A), the problem title or diagnoses; and plan (P), which includes the process of care code (which are often nonspecific) and an prescription or referral for further care. The computerized medical record (CMR) systems in the United Kingdom were historically strictly problem orientated, though those that are now in ascendency are more episode orientated [31]. The coding systems used within these systems have historically been hierarchical and used “chapters” that fit with the POMR structure [23].

We applied the ontology to the Read Code list by searching for codes indicative of IID diagnosis and mapped each into one of the following three classes [18,32]. Complete ontology and code lists are presented in supplementary tables (Multimedia Appendices 1 and 2).

Codes that refer to chronic conditions or non-intestinal conditions were defined as not mapping to IID and were excluded (eg, Helicobacter, Salmonella arthritis). We found that case finding was barely affected by the inclusion of codes in the least restrictive “no clear mapping” class and therefore did not use these codes in any analyses.

This study used primary care data recorded during a 52-week period spanning July 2014-July 2015 from the RCGP RSC, a sentinel network representative of the English population [33]. The cohort included patients with a recorded event, registered for the entire period. These data were used to determine the denominator. Data were extracted using SQL (Structured Query Language) software [34].

We calculated case numbers and incidence rates for the three IID definitions. When clinicians record a diagnosis, they assign episode type, which differentiates incident (first, new) cases from prevalent (follow-up, ongoing) cases. Records with “first” or “new” episode types were counted when counting cases and calculating incidence rates using diagnostic codes. When cases were found using directly mapping investigation and process care codes, all episode types were included because it is not standard clinical practice to code these events as “first” or “new.” Patients with excessive IID diagnostic records (>4 per year) were excluded from case counts as they likely had chronic gastrointestinal conditions, although this represented fewer than 10 people over the one-year study period.

We further investigated differences between case definitions and the validity of using an ontological case definition by searching patients who had been already counted as a case for codes relating to four supporting concepts: (1) symptoms (diarrhea, vomiting, and fever), (2) pathology investigations (stool sample sent to laboratory, to test for specific pathogens), (3) process of care (notification of dysentery or food poisoning), and (4) therapies (loperamide or oral rehydration therapy). We used a 2-week sliding window due to IID’s acute nature: all events for supporting concepts recorded with any episode type had to occur 2 weeks before or after the patient’s diagnosis event to be included. In addition, multiple events coded for any one factor within the 2-week window (eg, three investigation codes in one week) were counted as one event. Complete code lists for supporting factors are presented in Multimedia Appendix 2. We counted occurrences of each of the four supporting concepts and created Venn diagrams using R software [34].

The “Integrate” study received a favorable ethical opinion from the NHS NRES Committee North West-Greater Manchester East (Ref: 15/NW/0233). Patient-level data were automatically extracted and pseudonymized at the point of extraction. Data were stored at the University of Surrey Clinical Informatics and Health Outcomes Research Group data and analysis hub such that patients could not be identified from records used during the study.

An ontological approach to IID case finding changed IID incidence rate, increasing case detection. The ontological approach is also more transparent and independent of coding systems.

The ontological approach may address elements of IID underestimation due to low rates of case finding using electronic data alone [36], depending upon the case definition used [8,15,37]. However, the major limitation to accurate case finding remains that many community cases of IID do not seek health care [38].

GPs appear more likely to enter symptom codes, which from the ontological perspective are less helpful as they overlap with other conditions rather than being specific to IID, unless the symptoms are supported by another code indicating pathogen detection [39]. Results of the ontological approach have highlighted how use of symptom codes contributes to underreporting IID patients who do not have appropriate diagnostic or surveillance codes entered into the patient record.

An ontological approach provides insights into what types of data are available for case ascertainment. Although this approach offers benefits, and has limitations, our recommendation is to start by making the laboratory results recorded much more specific.

The mechanism for transferring results from stool sampling to GPs needs to be updated. Currently UK laboratories electronically report stool sample results to GPs using the Pathology Bounded Code List (PBCL), a subset of Read codes. However, there is no standardized algorithm for reporting results, and the PBCL code list for Microscopy, Culture & Sensitivity (MC&S) results has not kept pace with developments in pathology services. For example, typical laboratory protocol is to report one generic stool sampling code per test request, regardless of the range of pathogens being screened or detected, or of the sensitivity or specificity of the testing method. When a GP receives electronic results of a stool sample, the electronic report only contains generic MC&S Read codes, indicating that a stool sample was analyzed. This is followed by a “free-text” message (ie, not coded) indicating any detected pathogens. If pathogens are detected, the clinician must then code this information manually into the computerized medical record (CMR) system. This means that, inevitably, laboratory findings are under-coded. Furthermore, for some pathogens there is only one PBCL code specifically for test requests, not for recording results. Many IID pathogens have no designated PBCL code at all, and where appropriate pathogen codes are available, they are often not used. Given likely advances in stool microbiology testing in the future, with a move away from MC&S to nonculture, PCR-based methods, the way microbiology results are reported and coded via PBCL needs to be reviewed and modernized. There might be scope to draw lessons from biochemistry and hematology where, with the exception of glucose provenance and use of nonnumeric keys [39] and the use of nonnumeric characters, results with coded data are generally readily filed into the CMR system.

The principal limitation of this study is the lack of a gold standard; we do not know the “true” incidence of IID. There has been no back-to-case records review to validate this approach, though the authors have gone back to records to demonstrate the reliability of case finding from clinical records in other domains, for example, chronic kidney disease [40] and diabetes [41,42]. We have also reported where we consider conclusions to be unsafe because the wrong codes were selected [43].

In addition, ontologies are developed as an iterative process; therefore, we recommend testing by running data extracts to improve sensitivity and specificity. Our ontology is online and may be superseded by better laboratory coding, advances in near-patient testing, or other unforeseen advances. For example, there was no attempt to include social media data in this exercise. Techniques are emerging to do this and should be considered as part of future investigations and for inclusion in the subjective elements of the ontology [44,45].

Bias of many types can affect the quality of data recording in CMR systems. This can be around financial incentives to adopt CMR systems which then may not get used [46]; and around pressures within systems to either investigate, refer, or prescribe more (or less) depending on the constraints within the individual health care system at the time. These effects are probably best reported for drug safety studies where the availability of a large number of CMR records or administrative datasets had not obviated the need for other mechanisms of drug safety recording [47,48].

Finally, use of a new ontological approach to measuring disease incidence might result in further discrepancies between different surveillance systems that monitor the IID incidence. Harmonization of coding systems across different systems and countries is important from an epidemiological perspective to ensure that estimates of disease burden are comparable.

Our study indicates that use of the standard definition of IID to identify cases in primary care results in the underestimation of disease incidence. To capture a larger proportion of new IID cases in primary care, an ontological approach should be adopted to expand the case definition to include those patients with codes falling outside more restrictive standard definitions, as well as improving the PBCL coding list used by laboratories returning pathology results. Given the high burden of IID in the community, identifying what specific organisms are circulating within a community would help GPs and public health services. For GPs this would reinforce the importance of stressing simple and important control measures, such as hand washing, and trigger the implementation-specific interventions for specific infections. Local and regional public health services would more accurately know the disease burden and be able to intervene; nationally and internationally more accurate data would enable better policy evaluation and development around hygiene and food chain management.

Using these approaches will provide a better picture for clinicians, epidemiologists, and public health officials of the burden of IID in the community and the impact of seasonal infectious disease outbreaks.