Ascertainment of Helicobacter pylori Infection and Eradication Treatment Using a Nationwide Electronic Health Record Database

Abstract

Background and Aims

There are limited contemporary population-based data on Helicobacter pylori epidemiology and outcomes in the United States. Our primary aim was to create a validated cohort of veterans with H pylori testing or treatment using Veterans Health Administration data.

Methods

Using Veterans Health Administration structured and unstructured data, we developed and validated 4 algorithms for H pylori infection (3 algorithms) and treatment status (1 algorithm). During the development phase, we iteratively modified each algorithm based on a manual review of random sets of electronic health records (reference standard). The a priori validation goal was to achieve a one-sided 95% confidence lower bound (LB) for positive predictive value (PPV) and/or negative predictive value (NPV) >90%. We applied the Bonferroni correction when both PPV and NPV were relevant.

Results

For H pylori infection, we achieved 99.0% PPV (LB = 94.6%) and 100% NPV (LB = 96.4%) for discriminating H pylori positive vs negative status using structured (ie, laboratory tests) and 95% PPV (LB = 90.3%) and 97.9% NPV (LB = 93.9%) using unstructured (ie, histopathology reports) data. Diagnostic codes achieved 98% PPV (LB = 93.0%) for H pylori diagnosis. The treatment algorithm was composed of multiple antimicrobial combinations and overall achieved ≥98% PPV (LB = 93.0%) for H pylori treatment, except for amoxicillin/levofloxacin (PPV<60%). Application of these algorithms yielded nearly 1.2 million veterans with H pylori testing and/or treatment between 1999 and 2018.

Conclusion

We assembled a validated national cohort of veterans who were tested or treated for H pylori infection. This cohort can be used for evaluating H pylori epidemiology and treatment patterns, as well as complications of chronic infection.

Introduction

Helicobacter pylori is a gram-negative bacterium that colonizes over half of the world’s population and is the most common chronic bacterial infection globally. Chronic H pylori infection leads to gastric inflammation. While the majority of people will be asymptomatic or have nonspecific gastrointestinal symptoms, a small percentage of individuals will experience complications such as peptic ulcer disease or, rarely, gastric cancer.¹ H pylori is the strongest known risk factor for gastric adenocarcinoma and is designated by the World Health Organization as a human carcinogen. Successful eradication of H pylori infection is associated with a reduced risk of these complications.² Thus, because of H pylori’s persistent nature, eradication treatment is typically recommended. In the United States, treatment entails a 10–14 day course of 1–3 antibiotics and a proton pump inhibitor in combination.^3,4

H pylori epidemiology, diagnostic and treatment practices, and outcomes such as eradication treatment failure are poorly described in the United States at a population level. The last large cohort of individuals evaluated for H pylori infection was the National Health and Nutrition Examination Survey cross-sectional study in 1988–1991 and 1999–2000.^5,6 This cohort evaluated H pylori infection via serology only and did not evaluate non-serological testing or H pylori treatment. Notably, the first clinical consensus and guideline statements recommending H pylori eradication treatment among individuals testing positive were not published until 1998.⁷ Ascertainment of H pylori episodes of care poses a challenge due to the several modalities of testing and treatment, as well as imperfect diagnostic coding. These knowledge gaps not only impact patient outcomes related to H pylori management but also resource allocation for at-risk groups.

Our primary aim was to assemble and validate a national cohort of veterans who had been tested for H pylori, diagnosed with H pylori based on diagnostic billing codes, or treated for H pylori as ascertained by structured and unstructured data contained within the Veterans Health Administration (VHA), one of the largest integrated health systems in the United States.

Methods

This study was approved by the committee on research ethics at Veterans Affairs (VA) San Diego and VA Tennessee Valley, in accordance with the Declaration of Helsinki.

Study Data Sources

We compiled data from the VHA, which provides longitudinal care to approximately 9 million US veterans. The VHA Corporate Data Warehouse aggregates longitudinal patient data from all US veterans receiving care through any VHA facility across the United States and its territories. Individual-level data include demographics, comorbidities, laboratory testing and results, procedures and results, progress notes, diagnoses, billing codes, vital status files, and all dates. We collated all demographics, laboratory tests and results, International Classification of Diseases (ICD) codes, and procedure information related to inpatient and outpatient encounters between 1999 and 2018. These data elements include a mix of both structured and unstructured data. We linked these individual-level data to the VHA pharmacy files to identify details of all filled prescriptions, including dates of therapy, quantity dispensed, and dosing instructions (ie, dose and frequency).

Algorithm Development and Iteration

Diagnosis

Three algorithms were developed to identify individuals tested for and/or diagnosed with H pylori (Figure).

Figure.

Schematic diagram of the validation approach for H pylori diagnosis or treatment status based on structured and unstructured Veterans Health Administration data.

Algorithm 1 development: ascertainment of H pylori status based on H pylori diagnostic laboratory tests

The following elements were ascertained from inpatient and outpatient structured data: H pylori date of testing, testing modality (serology vs non-serology), and testing result (positive, negative, equivocal). H pylori stool antigen testing, Campylobacter-like organism test, rapid urease test, urea breath test, and H pylori culture were classified as non-serological tests. Qualitative results of each test were recorded based on the data query; result categories included positive or negative. While the majority of serological lab tests were reported qualitatively, a minority were reported quantitatively (<10%). The lab value ranges and units of measurement for these quantitative results were explored. It was identified that the threshold values for categorizing these quantitative labs as positive vs negative varied depending on the test type, assay, and unit of measurement. Because there was no way to standardize the reference ranges and units across each quantitative lab measurement, we categorized values greater than the 75th percentile as ‘positive,’ less than the 25th percentile as ‘negative,’ and between the 25th and 75th percentile as ‘equivocal.’ Qualitative equivocal tests were not evaluated as they comprised <3% of all lab values.

Algorithm 2 development: ascertainment of H pylori status based on histopathology reports with gastric biopsies

This algorithm leveraged both keyword searching and natural language processing methods applied to unstructured data. We identified individuals who had undergone an inpatient or outpatient esophagogastroduodenoscopy (EGD) procedure as identified based on the Current Procedural Terminology code (Table A1) and had a pathology report available in the ± 30 days from the date of the EGD.

We then applied the keywords for ‘stomach’ and different gastric locations only to the portion of the histopathology report where biopsy results are reported. All reports demonstrated the same organization of unstructured data elements; for example, all biopsy results were reported in the section following the “gross description” section. In order to prevent interpreting historical or history of H pylori or other gastric diseases, we developed a code such that only the text after the phrase “gross description” in the pathology report was interpreted. Thus, if the histopathology report states “history of H pylori” in the indication section, but gastric biopsies were not obtained during the EGD, then this report would not be selected by the algorithm since H pylori status cannot be ascertained without gastric sampling.

Pathology report documents were divided into separate sentences for the purpose of identifying sentences with positive or negative diagnosis of H pylori. The end of a sentence was defined as having at least 2 characters and could include parentheses or letters in that character string. Gastric biopsies evaluate the presence vs absence of H pylori infection using a stain and often include positive and negative quality control results in the report. Thus, the initial algorithm was constructed such that word pairings or phrases related to the performance of the control stains were removed from consideration; these phrases included “positive and negative control stained appropriately,” “positive control,” “positive and negative,” “appropriately positive,” and “appropriately negative.” “With feature” and “with finding” were also excluded due to their high frequency and nonspecific context. Once this basic algorithm was developed, formal rounds of validation using randomly selected groups of algorithm positive vs negative individuals were completed, with algorithm modifications implemented as appropriate if threshold performance metrics were not achieved.

Algorithm 3 development: ascertainment of H pylori diagnosis based on ICD-9 and ICD-10 codes

We evaluated the performance of a single inpatient or outpatient ICD-9 (041.86) or ICD-10 (B96.81) code for identifying someone with a true diagnosis of H pylori infection, in the absence of other structured indicators of H pylori infection (ie, laboratory tests). A diagnosis was considered verified if a manual chart review confirmed that the individual had H pylori infection based on testing (either within or outside the VHA), treatment, or a clinical progress note documenting current or prior H pylori infection. ICD codes can only be used to identify H pylori-positive individuals since there are no corresponding ICD codes for H pylori-negative status (ie, someone who was tested for H pylori but tested negative). Thus, the negative predictive value (NPV) was not applicable to this algorithm.

H pylori eradication treatment

Algorithm 4 development: ascertainment of H pylori eradication treatment

We validated multiple guideline-recommended H pylori eradication regimens. H pylori eradication treatment consists of 1–3 antimicrobials ± bismuth along with twice per day proton pump inhibitor (PPI), in combination for 10–14 days. The US guideline-recommended H pylori eradication treatment regimens are provided in Table A2.³ Because PPIs are also obtained over the counter, these medications were not included in the initial algorithm development. For each H pylori eradication regimen, we evaluated the accuracy of individual treatment regimens for H pylori in the absence of a concomitant ICD code or H pylori laboratory test.

A minimum 7-day overlap was required for each prescription, which was based on an initial round of chart review and data exploration during algorithm development. The days of medication dispensed, that is “day supply,” was restricted to 5–30 days in order to exclude outliers; even though therapy is only prescribed for a maximum of 14 days, it is not uncommon for a 30-day days supply to be dispensed due to VHA benefits, which is why 30 days and not 14 days were selected as the upper limit. High-dose amoxicillin and PPI therapy (dual therapy) were not included in the query due to the nonspecific nature and due to the rarity of this regimen for H pylori treatment in the United States currently.^3,8 For each H pylori treatment regimen, we evaluated its positive predictive value (PPV) for H pylori indication in the absence of a concomitant ICD code or H pylori laboratory test. NPV was not applicable for this algorithm.

Chart Abstraction, Validation, and Analysis

Each algorithm underwent independent rounds of review of a randomly selected set of unique electronic health records (EHRs) that were categorized as algorithm “positive” or “negative.” Independent samples of cases and controls were drawn for each round of review. The EHR was considered the reference standard for categorizing true vs false positives and negatives. The EHR comprised all laboratory tests, medical encounter notes, procedure notes, medications, and billing documentation. Medical charts without any evidence or mention of H pylori testing, diagnostic code, or treatment were not considered in the validation process. The reviewers (M.L., H.Y., S.C.S.) were blinded to the algorithm output. Discordant findings were carefully reviewed and the algorithms were modified as needed based on the etiology of the discordance.

The a priori goal was to achieve a one-sided 95% confidence lower bound (LB) for PPV and/or NPV >90%. Iterative modifications to each algorithm were applied if the a priori thresholds for PPV or NPV were not achieved (Figure). The PPV was defined as the proportion of subjects who were correctly classified by the proposed algorithm as having the outcome (ie, proportion of true positives; eg, individuals who were classified by algorithm 1 or 2 as H pylori-positive and who were manually confirmed to have H pylori based on laboratory testing or histopathology, respectively). The NPV was defined as the proportion of subjects without the outcome among those who were classified by the proposed algorithm as not having the outcome (ie, proportion of true negatives; eg, individuals who were classified by algorithm 1 or 2 as H pylori-negative and who were manually confirmed to be H pylori-negative on laboratory testing or histopathology, respectively). The prevalence estimates for each of the PPV and/or NPV performance calculations were derived by applying the respective algorithms to the full cohort of individuals receiving longitudinal VHA care.

For those algorithms in which both PPV and NPV were relevant (algorithms 1 and 2), we applied the Bonferroni multiple comparison adjustment to ensure an overall confidence level of 95% for PPV and NPV estimates. Based on these target thresholds and the estimated PPV and NPV from the development phase, we projected that we would need to randomly sample at least 100 algorithm-positive cases and 100 algorithm-negative controls, respectively, for each iterative validation round.⁹ Based on the projected sample size, if the estimated PPV and NPV achieved 95% or greater in the validation round, the 95% one-sided confidence LBs would be at least 90%, and if the estimated PPV and NPV achieved 90% or greater, then the 95% one-sided confidence LBs would be at least 84%.⁹

H Pylori Cohort Creation

We applied each of the 4 validated algorithms to individuals >18 years old receiving longitudinal care within the VHA to assemble a nationwide cohort of H pylori tested and/or treated individuals between 1999 and 2018. Longitudinal care was defined as having at least one medical encounter, laboratory test, or prescription fill. We compiled basic demographic information for the cohort.

Results

The performance of each of the validation algorithms is summarized in Table.

Table.

Summary Performance of Validation Algorithms

Algorithm	Iterative roundsa	Prevalenceb	PPV (lower bound)c	NPV (lower bound)c
1 – Laboratory tests	1	0.26	0.99 (0.95)	1.00 (0.96)
2 – Histopathology	1	0.13	0.94 (0.87)	0.97 (0.915)
	2	0.12	0.95 (0.86)	0.97 (0.92)
	3	0.11	0.95 (0.89)	0.98 (0.94)
3 – ICD codes	1	0.10	0.98 (0.94)	n/a
4 – Eradication treatment	1	0.10	0.98 (0.94)	n/a
	2 (amoxicillin/levofloxacin combination only)	0.10	0.55 (0.46)	n/a

n/a, not applicable.

^a100 algorithm-positive and 100 algorithm-negative controls were manually reviewed for the validation phase in each round with the following exceptions: 110 algorithm-positive and 110 algorithm-negative EHRs were manually reviewed for round 2 of Algorithm 2, and 140 algorithm-positive and 140 algorithm-negative EHRs were manually reviewed for round 3 of Algorithm 2. Independent samples of cases and controls were drawn for each round of review.

^bPrevalence estimates are derived from applying the respective algorithms to the full data set of individuals ≥18 years old with longitudinal VHA care.

^c95% one-sided confidence lower bound for PPV and NPV was based on binomial exact test with Bonferroni correction for Algorithms 1 and 2.

Algorithm 1: H Pylori Laboratory Tests

Among individuals with any laboratory-based H pylori testing, the algorithm classified the testing result as positive vs negative with 99% PPV (LB 94.6%) and 100% NPV (LB 96.4%) (Table). This algorithm performance was achieved using the threshold for classifying quantitative H pylori results greater than the 75th percentile as positive and less than the 5th percentile as negative; as previously noted, quantitative results comprised <10% of the H pylori lab tests overall. The prevalence of H pylori based on the laboratory testing algorithm was 26%, which is consistent with recently published data reporting H pylori positivity rates among the VHA population with an indication for H pylori testing.¹⁰ The sensitivity of the algorithm was 100% and specificity was 99.6%.

Algorithm 2: Histopathology

In the first round of validation, the algorithm achieved 94% PPV (LB 87.4%) and 97% NPV (LB 91.5%) for H pylori positive and negative diagnosis, respectively (Table), which was above the a priori target performance threshold for PPV.

The algorithm was therefore modified based on the experience from the first round. We subsequently added variations of “neg helicobacter found stain” to the negative identifier for H pylori and removed the 4-part rule “no strength histopath abnormal,” in favor of using various 3-part combinations of that phrase instead (eg, “no strength histopath,” “no histopath abnormal”). The second independent random set of 100 algorithm-positive and 100 algorithm-negative EHRs were reviewed and the PPV improved to 94.5% (LB 88.5%), without significantly impacting the NPV (97.3%, LB 92.2%), but this was still above the target PPV.

In the third validation round, we again modified the algorithm based on the second-round experience. We adjusted the code to recognize complete sentences such that multiple punctuation characters would be accepted at the end of a sentence (eg, “..”). We also expanded the tenses for applicable positive phrases and added “none” and “non-” as synonyms for “no.” A third independent random sample of 100 algorithm-positive and 100 algorithm-negative EHRs were selected to validate and achieved 95.0% PPV (LB 90%) and 97.9% NPV (LB 93.9%). Thus, all a priori threshold performance values were met and this was considered the final algorithm. The prevalence of H pylori based on the final algorithm was 11%, which is consistent with estimates among the VHA population (unpublished data). The estimated sensitivity and specificity were 84.7% and 99.4%, respectively.

Algorithm 3: Diagnostic Codes

The presence of one inpatient or outpatient ICD-9 or ICD-10 code alone achieved a 98% PPV (LB 93.8%) for a true diagnosis of H pylori infection based on EHR documentation among individuals with at least one ICD-9 or ICD-10 code for H pylori (Table). Often this was confirmed based on H pylori testing that occurred outside of the VHA.

Algorithm 4: H Pylori Eradication Treatment

We validated the performance of the eradication treatment algorithm for correctly identifying that these regimens were indeed being used for H pylori treatment and not another indication, even in the absence of other structured data indicators of H pylori infection (ie, H pylori testing and/or ICD codes).

The first round of validation after the development phase yielded an observed PPV of 98% (LB 93%) for H pylori eradication treatment indication for each of the guideline-recommended therapies (see Table A2) with the exception of the combination of amoxicillin/levofloxacin (Table). EHR review of amoxicillin/levofloxacin cases revealed that this regimen was most often prescribed for skin and soft tissue or other non-H pylori infections. As such, in order to improve performance, we modified the algorithm for amoxicillin/levofloxacin to remove any patients with ICD codes for non-H pylori infections in the ±30 days surrounding the date of amoxicillin and levofloxacin prescription fills, as well as removing individuals who had prescription fills that were not consistent with H pylori treatment recommendations (ie, removing any day supply that was not 7 days, 10 days, or 14 days). We also narrowed the overlap period from 7 days to 5 days. Other elements of the prescription including the date prescribed and duration of therapy were 100% accurate; however, indication for the prescription combination could not achieve the a priori threshold for validity (observed PPV 55%, LB 44.7%). Thus, in the absence of other indicators such as H pylori laboratory testing or ICD code, the combination of amoxicillin/levofloxacin alone is not indicative of H pylori eradication treatment, and this combination was therefore removed from the algorithm. All other guideline-recommended regimens achieved very high PPV for an H pylori eradication treatment indication independent of any other structured H pylori data indicators in the VHA record (eg, laboratory tests, ICD codes) and thus comprised the final algorithm for H pylori eradication treatment.

H Pylori Cohort Details

Of 15,594,932 individuals who received care through the VHA between 1999 and 2018, we identified 1,199,032 individuals with H pylori testing (laboratory and upper endoscopy with gastric biopsies [algorithms 1 and 2]) (n = 913,328; 76.2%), H pylori ICD-9 or ICD-10 code only without evidence of laboratory testing [algorithm 3] (n = 60,472; 5.0%), or H pylori eradication treatment only without evidence of laboratory testing or ICD-coding [algorithm 4] (n = 225,232; 18.8%). The majority of eradication regimens with clarithromycin-based triple therapy (Table A2), while amoxicillin/levofloxacin comprised less than 7% of all eradication combinations. These 1,199,032 individuals comprised the starting VA-Helicobacter pylori (VA-HP) cohort. Basic demographics of the cohort included mean age 58.1 (standard deviation, 14.9) years, 90.4% men, 66.2% non-Hispanic White, 15.3% non-Hispanic Black, 6.1% non-Hispanic other, and 6.4% unknown, and 6.0% Hispanic.

Discussion

We achieved the primary aim of this study and validated a large nationwide cohort of individuals receiving care through the VHA who underwent testing and/or treatment for H pylori infection (VA-HP cohort). Indeed, of the nearly 1.2 million people identified, 913,328 (76.2%) had evidence of formal H pylori testing. By using both structured and unstructured data to develop the algorithms, we optimized both data capture and performance, as evidenced by most algorithms only requiring one round of manual chart review to achieve a priori performance thresholds. All algorithms for H pylori diagnosis based on testing or ICD codes achieved PPV 94%–99% (LB 90%–94.6%) and NPV 97.9%–100% (LB 93.9%–96.4%) and, with the exception of the combination of amoxicillin/levofloxacin, all guideline-recommended combinations of H pylori eradication treatments achieved PPV 98% (LB 93.8%) for H pylori treatment indication.

To our knowledge, this is the largest validated cohort of individuals with confirmed H pylori testing and results available. The VA-HP national cohort can be used to describe contemporary H pylori epidemiology, risk factors, diagnostic and treatment practices, and outcomes, including H pylori eradication treatment failure. In addition to validating the results of testing, we confirmed that we were able to ascertain related H pylori testing and procedural dates, testing modality (serology vs non-serology), and prescription details. Notably, the VHA data are readily linked to the VA Central Cancer Registry and National Death Index, which accurately identify cancer diagnoses and causes of death, respectively, as well as the Million Veteran Program, one of the largest actively enrolling genomic biobanks globally.¹¹ Accordingly, the VA-HP cohort can not only be leveraged for time-to-event analyses related to cancer and cancer mortality outcomes but also to characterize genetic and gene x environment underpinnings of H pylori susceptibility and its plethora of complications. Importantly, the VA-HP cohort includes all individuals tested for H pylori which reduces the potential for misclassification since H pylori infection is often asymptomatic and may therefore go undiagnosed unless someone is formally tested.

Our work extends limited prior literature reporting on the PPV and NPV of using administrative data to identify individuals with an H pylori diagnosis. Thirumurthi et al¹² evaluated the performance of H pylori ICD-9 code or prescription of triple or quadruple eradication therapy for a diagnosis of H pylori. They reported that the PPV of the H pylori ICD-9 code ranged between 97.4% and 100%, while their algorithm for triple and quadruple therapy had a PPV of 73.7% and 97.7%, respectively. Notably, this study did not validate laboratory tests or histopathology obtained from EGD with gastric biopsies. This study published in 2008 also predated the most recent 2017 US American College of Gastroenterology guidelines on H pylori treatment and predated the introduction of ICD-10 coding.³ It is also worth noting that, while ICD codes have excellent PPV for H pylori, these only capture a very small fraction of the pool of individuals with H pylori infection in the VHA and also miss all individuals who were tested but who tested negative.

Our study has several strengths. We leveraged comprehensive algorithm development using both structured and unstructured data and a rigorous validation process. We validated all guideline-recommended H pylori eradication treatments and reported on the accuracy of capturing the correct medications, dates of treatment, as well as indication (i.e., treatment of H pylori infection). Lastly, we were able to apply our algorithm across the VHA nationwide database, which includes over 160 distinct VHA stations, to create the VA-HP cohort spanning 1999–2018. Our study also has limitations. We are not able to reliably capture care that occurs outside of the VHA, and thus there is still the small possibility of misclassification. Our algorithms were validated against the VHA database and may perform differently in other electronic health record systems; future work can assess the performance of these algorithms on non-VA data.

In conclusion, we created a large, nationwide validated cohort of individuals with established VHA care tested or treated for H pylori infection. H pylori infection is common, particularly in high-risk groups such as non-White racial and ethnic groups, early-generation immigrants from countries where H pylori infection is endemic, and older individuals. Chronic H pylori infection is associated with potentially serious complications, including gastric cancer. Accordingly, this powerful cohort comprising nearly 1.2 million individuals can be used to evaluate H pylori epidemiology, treatment patterns, and outcomes, as well as risk factors for disease complications related to chronic H pylori infection.