Introduction

As populations become more connected and diverse through travel and immigration, there is a growing interest in understanding the health needs of migrant populations to protect their health and that of the broader public [1]. Migrant travellers who travel to their birth countries to visit friends and relatives (VFR travellers) are an important and growing risk group for travel-related infectious diseases [2]. In ethnoculturally diverse regions in Canada and around the world, VFR travellers account for the majority of hepatitis A, malaria, and enteric fever cases reported to public health [3]. These diseases are preventable through safe and effective vaccines or chemoprophylaxis [4–6]; however, travel health services are not reimbursed in many countries with universal health insurance plans including Canada, although some countries (e.g. England and Australia) provide coverage for select services [7–9]. Given emerging trends in globalization and climate change that are increasing opportunities for exposure [10], it is prudent to measure the burden of travel and migration-related infectious diseases to evaluate current policies and develop effective public health interventions.

Reportable disease surveillance systems are the mainstay for burden estimation as many travel-related infectious diseases are notifiable by law [11]. Despite this, an important limitation is they usually lack information on country of birth, region of origin, or immigrant status, which are needed to appropriately target preventive measures. Accordingly, the European Centre for Disease Prevention and Control has called for better data to improve understanding of risk groups of imported malaria in Europe [12]. Another important challenge with public health surveillance data is poor sensitivity as cases may not seek health care or be reported to public health (which varies by disease [13, 14]), or they may seek health care or die from their infection while travelling [15, 16]. These limitations may explain the lack of studies estimating burden of travel or migration-related infectious diseases, despite the value of this information to decision-makers [17].

Routinely collected health administrative data (e.g. data generated during health care delivery such as records of hospitalization, outpatient physician visits, emergency department visits, etc.) offer an efficient and inexpensive alternative to public health surveillance data for burden estimation of medically-attended travel or migration-related diseases. They also provide an opportunity to overcome existing data limitations through population-based linkages with immigration data [18] and to address important gaps if diseases are removed from notifiable disease lists (as in Ontario recently for malaria). Because the validity of these data for identifying travel or migration-related infectious diseases is not well-understood, we sought to examine if health administrative data could be used to accurately identify cases of hepatitis A, malaria and enteric fever compared with reportable disease surveillance data as the reference standard.

Materials and methods

Study cohort

We constructed a cohort of medically-attended individuals with presumed hepatitis A, malaria or enteric fever in Peel region through deterministic linkage of the Canadian Institutes for Health Information’s hospital discharge abstract (DAD) and same-day surgeries databases (SDS), National Ambulatory Care Reporting System (NACRS), and Ontario Health Insurance Plan (OHIP) physician claims. These databases include all persons covered under Ontario’s universal health insurance plan (new immigrants with the exception of refugees have a three month wait period before they are eligible for universal health care [20]), and are held at ICES [21]. Diagnoses are coded using the International Classification of Diseases, 10th revision Canada (ICD-10-CA) for all databases aside from OHIP that uses a similar, but not identical, version of ICD-8 that identifies broad disease classifications.

We identified a set of sensitive diagnostic codes to select individuals (S1 File) that were felt to reflect a confirmed diagnosis, symptoms, or interpretation of a diagnostic test for the diseases of interest based on consultation with an infectious disease physician and the literature [22, 23]. For the primary analysis, we restricted the cohort to the most presumed specific codes (i.e. ICD-10 codes B15, B50-54, A01 or OHIP codes 070, 062, 002). There are OHIP diagnostic codes available for viral hepatitis (070) and typhoid and paratyphoid fevers (002) but none for malaria. As a result, we selected code 062 (mosquito-borne viral encephalitis) because of its reference to mosquito-based transmission. We also performed a post-hoc sensitivity analysis using code 136 (other infectious or parasitic diseases), which is referenced in the Resource Manual for Physicians for malaria [24].

Any records with a valid ICES key number (IKN), a unique identifier based on their health card, that met the diagnostic selection criteria between November 20, 2011 and February 11, 2015 were included. This time period provided a six-week buffer relative to the case report date, recognizing that individuals seek health care prior to being reported to public health, and may require follow-up visits.

Records meeting our inclusion criteria were linked deterministically across databases. These datasets were linked using unique encoded identifiers and analyzed at ICES. Records were further restricted to Peel region residents using the individual’s postal code in the Registered Persons Database (RPDB). We excluded individuals with a missing postal code (0.4%) (Fig 1).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Study flow chart.

https://doi.org/10.1371/journal.pone.0207030.g001

Individuals with health care encounters 3 to 24 months prior to cohort entry were flagged to explore potential misclassification (e.g. illness might have been chronic and non-travel related). Remaining records were linked to the Ontario portion of the Immigration, Refugee and Citizenship Canada’s permanent resident (IRCC-PR) database to ascertain immigrant status. IRCC-PR data capture landed immigrants to Ontario from 1985 to 2012 who obtained a valid Ontario health card number [18]. Immigrant status was of interest given that VFR travellers experience disproportionate travel-related morbidity [25–27] and may also seek care differently or have differential access to diagnosis versus long-term residents [28]. Any individual present in the IRCC-PR was classified as an immigrant, recognizing the important limitation that individuals who immigrated pre-1985 and post-2012 would not be classified. Additionally, new immigrants who currently reside in Ontario but who landed in another province were not classified. The dataset creation plan for the study cohort is in S2 File.

Results

Over the three year period, 55 confirmed cases of hepatitis A, 125 of malaria, and 141 of enteric fever (N = 321) were reported to PPH (Fig 1). In the study cohort, there were 788,912 unique individuals with presumed travel or migration-related infections with a corresponding 2,539,975 health care encounters (Fig 1). The majority of these encounters were outpatient/primary care physician visits (2,499,947 or 98.4%), followed by emergency department visits (33,314 or 1.3%), hospitalizations (5,279 or 0.2%) and then same-day surgeries (1,417 or 0.1%).

Probabilistic linkage of iPHIS cases to the RPDB was successful for 90.0% (289/321) of cases. Un-linked cases were more likely to be a recent immigrant or visitor to Ontario compared to linked cases (53.3% vs. 8.7%, P < 0.001), likely because they did not have an Ontario health card, but otherwise were similar in terms of sex, age, disease distribution and foreign-born status.

Incidence rates of hepatitis A, malaria and enteric fever were substantially higher in the health administrative data compared with the reference standard (Table 1).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Number of cases and disease incidence rates (per 100,000) as determined by the reference standard (iPHIS) compared to health administrative data, November 20, 2011 to February 11, 2015.

https://doi.org/10.1371/journal.pone.0207030.t001

Sensitivity varied by disease (Table 2) and was lowest for malaria given that there were zero true positives with 062-coded health encounters in OHIP. Over 80% of hepatitis A cases reported to public health were present in health administrative data with diagnostic codes 070 (OHIP) or B15 (ICD-10); however, only 0.3% of those with these codes were a laboratory-confirmed case reported to public health. PPV was similarly low for enteric fever but improved for malaria (PPV = 43.0%). In the post-hoc sensitivity analysis replacing OHIP code 062 with 136, the number of true and false positives rose to 89 and 7,556 respectively, with a decrease in false negatives to 24 [sensitivity = 78.8% (95% CI 71.2–86.3); PPV = 1.2% (95% CI 0.9–1.4)].

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Accuracy of diagnostic codes for hepatitis A, malaria and enteric fever used in health administrative data, overall and by data source, with Ontario’s reportable disease registry (iPHIS) as the reference standard.

https://doi.org/10.1371/journal.pone.0207030.t002

When stratified by health administrative database (Table 2), diagnostic codes in outpatient/primary care physician visit data (OHIP) were more sensitive than hospitalization (DAD) and emergency department visit (NACRS) data for both hepatitis A and enteric fever. For malaria, sensitivity was highest for emergency department visit data. Hospitalization data provided the highest PPV for all diseases.

Sensitivity estimates were similar in immigrants compared to long-standing residents across all diseases (87% vs. 81% hepatitis A; 69% vs. 68% malaria; 84% vs. 89% enteric fever). PPV was similar across groups for hepatitis A (0.3% vs. 0.3%) but higher amongst immigrants for malaria (57.1% vs. 26.5%) and enteric fever (7.8% vs. 4.9%).

We compared true and false positives (Table 3) and found that true positives were more likely to be younger (with the exception of malaria), to not have had similar health encounters in the lookback period, and to be an immigrant. True positives were also more likely to be hospitalized or visit the ED than false positives and to have a higher number of health encounters but spaced over a shorter duration. In total, 53 false positives had a specific diagnostic ICD-10 code assigned to a hospitalization, and likely represent true, unreported disease cases.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Characteristics of true (TP) and false positive (FP) individuals, by disease.

https://doi.org/10.1371/journal.pone.0207030.t003

A description of cases reported to public health that were not present in the health administrative data is provided in Table 4. Of the 61 false negatives, 18 had no health encounters matching any of the specified diagnostic codes. The remaining 43 cases were present in the health administrative data but did not have a specific diagnostic code. Amongst the 43 cases, 42 (97.7%) accessed primary care, while <6 visited the emergency department or were hospitalised. Overall, false negatives were predominantly younger and male. Almost one-quarter (21.3%) were not eligible for OHIP at the time of their illness. While all cases had PPH indicated as their responsible health unit, only 63.9% had a Peel region postal code assigned in the health administrative data based on their most recent contact with the health system.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Characteristics of false negatives.

https://doi.org/10.1371/journal.pone.0207030.t004

Broadening the diagnostic codes used to classify individuals as having hepatitis A, malaria, or enteric fever improved sensitivity by ~10–15%; however, PPV decreased near zero for every disease (Table 5). All exclusions made in post-hoc sensitivity analyses led to either no change or a slight decrease in sensitivity with only minor improvements in PPV (S5 File).

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Test characteristics of sensitive (low specificity) health administrative diagnostic codes of hepatitis A, malaria and enteric fever diseases with the reportable disease registry iPHIS as the reference standard.

https://doi.org/10.1371/journal.pone.0207030.t005

Discussion

We found that diagnostic codes in Ontario’s health care databases had good sensitivity (69–86%) to identify true cases of hepatitis A, malaria and enteric fever reported to public health. An additional 10–15% of cases were present in health administrative data but with less specific codes. Despite this, only a small proportion of individuals with diagnostic codes for these diseases were true cases identified in reportable disease surveillance data (PPV range: 0–41%).

While the utility and validity of health administrative data for chronic disease surveillance has been extensively studied worldwide, the use of these data for infectious diseases is still in its infancy [23, 31]. As the potential of these data expand, growing to encompass electronic health records and population-level integrated laboratory databases, characterization of the limitations and strengths of health administrative data for identifying infectious disease cases is needed. In our study, we identify several challenges in the context of travel and migration-related infectious diseases. The first is that recent immigrants may be underrepresented in burden studies reliant solely on health administrative data if universal health coverage policies exclude these new residents. Secondly, the predictive accuracy of these data is dependent on precise and up-to-date diagnostic coding systems that have substantial diagnostic specificity [32]. In Ontario, the lack of disease-specific codes in primary care largely drove the poor PPV we observed (e.g. codes did not distinguish between hepatitis types, nor was there a dedicated code for malaria). As primary care is often the first and only point of contact with the health care system for infected individuals, improving coding in this setting is particularly important. Updating the antiquated coding system of physician billings to the latest coding system (i.e. ICD-10) could vastly improve both the sensitivity and PPV of health administrative data for travel and migration-related diseases and bring alignment with other health care data. In the meantime, quality improvement initiatives may help standardize how current codes are being used by physicians.

Beyond coding systems, there are challenges inherent in assigning diagnoses for infections. For example, disease diagnoses are rarely confirmed by laboratory testing at the time the physician sees the individual. This uncertainty can lead to false positives if disease codes include those with suspected, not confirmed, infections [23], and also false negatives if generic codes are assigned to indicate no diagnosis (e.g. we found that half of the malaria false negatives in our study were assigned code 999 (no diagnosis) in OHIP). Linking primary care databases with electronic laboratory data could address this challenge. Lastly and perhaps the most difficult challenge to address are physician-related barriers to accurate coding. Through interviews with professional coding specialists, Tang et al. recently identified incomplete and nonspecific documentation, errors and discrepancies within patient charts, disconnects between physician and coder terminologies, and lack of communication between physicians and coders as key barriers to accurate coding [33]. Study authors acknowledge that these barriers have structural and systemic causes that often extend beyond the control of physicians and as such, are complex to address. Despite this, as the use of health administrative data for research, surveillance, and evaluation grows, and with it the need to evaluate health care to maximize efficiencies and health outcomes, improving the accuracy of these data should be prioritized so that they can reliably inform decision-making [33].

Despite these challenges, there are strengths of health administrative data for travel and migration-related infectious diseases that can be leveraged to further improve its accuracy and utility. First, health administrative data are linkable to reportable disease surveillance data and fill important gaps by providing objective measures of health care utilization and outcomes of travel or migration-related disease cases. Further linkage with immigration records can enable study on health care burden and cost of these infections broadly and also in migrants and be used for resource prioritization and setting. We also identified that hospitalization records, on their own, are likely useful for accurately identifying severe cases of hepatitis A, malaria and enteric fever given the relatively high PPV of diagnostic codes compared with ED and primary care data. Lastly, we identified a number of demographic and health encounter-related differences between true and false positives that may be exploited to improve the accuracy of health administrative data for infectious diseases. For example, Yu et al (2013) have shown that a classification algorithm that incorporates additional administrative information associated with a hospitalization can improve the accuracy of administrative data to detect individuals with community-acquired pneumonia [34]. Together, our findings suggest that health administrative data have the potential to add value to infectious disease surveillance, and merit further investigation of how their utility can be improved.

Although we took measures to be comprehensive in our selection of diagnostic codes, it is possible that alternative codes are being used by physicians, which may have underestimated the accuracy of health administrative data. In our study, 16% of laboratory-confirmed malaria cases reported to public health were not present in the health administrative data with any of the diagnostic codes we felt were relevant. In a subsequent study, we reviewed the complete clinical record of cases and found that a small proportion (<5%) of cases of malaria and enteric fever were additionally assigned OHIP codes 038 (septicemia) and 780 (convulsions, ataxia, vertigo, headache); however, as these codes were not detected in false negative cases, failure to include these would not have changed our sensitivity estimates (data not shown). We recommend any future studies of travel and migration-related infectious diseases using OHIP data conduct a chart review to validate case definitions first, where possible, to optimize accuracy.

A second limitation of our study is the use of postal codes to restrict our study cohort to Peel residents. Postal code is assigned based on the address reported on the individual’s health card but is supplemented with other ICES-held datasets that collect residence information at each health encounter [35]. Postal codes are estimated annually to select the most recent data as of July 1st as the reference date [35]. Therefore, movement in and out of Peel within one year may mean that individuals who lived in Peel at the time of their diagnosis but later moved in that same year may have been excluded (or vice-versa). During our study period, ~80,000 persons/year moved to Peel (<6% of total population) and ~65,000 moved out of Peel (<5%), and so it is not expected that misclassifications would appreciably change our findings.

Lastly, as iPHIS is an imperfect standard, some of the individuals we classified as false positives may have been true disease cases. To assess the impact of this potential misclassification, we used capture-recapture methods to estimate an additional 37 cases not captured in iPHIS or hospitalization records, for an iPHIS reporting completeness of 74% (95% CI 61–83%) for hepatitis A, 82% (95% CI 75–88%) for malaria, and 78% (95% CI 71–84%) for enteric fever (S2 File). These estimates are much higher than what was recently reported for pertussis in Ontario using a similar methodology [36], suggesting that the impact of this misclassification would likely be small. Indeed, if we moved these 37 cases from the false positive to true positive category, there would be <5% absolute improvement in sensitivity and <20% absolute improvement in PPV. Investigating false positives can identify if there are opportunities to improve hospital/physician reporting practices to public health.

Conclusion

Our findings highlight challenges and strengths of using health administrative data to identify cases of travel or migration-related infectious disease, and caution against the use of these data on their own (i.e. without linkage to records confirming infection) to identify cases. Linking these data with reportable disease surveillance data and immigration records can, however, bridge important gaps in estimating burden using either data source independently. Linkages with electronic laboratory data, and initiatives aimed at improving the coding of health administrative data, particularly in primary care settings, offer potential to further improve the utility of these data for understanding the health needs of migrants.  

Supporting information

Acknowledgments

The authors wish to thank Immigration, Refugees and Citizenship Canada for providing data to ICES where analyses were conducted. We also wish to thank Gregory Kujbida, Epidemiologist, Peel Public Health, for his assistance in preparing the reference data for linkage, as well as Alex Kopp and Laura Holder, ICES, for their assistance in preparing the health administrative cohort.