A novel data-driven workflow combining literature and electronic health records to estimate comorbidities burden for a specific disease: a case study on autoimmune comorbidities in patients with celiac disease

We first selected the list of autoimmune diseases from MeSH® terminology. We then restricted this list to the most frequent autoimmune diseases associated with CD in the literature, based on the number of co-occurrences of MeSH® terms in MEDLINE®. We mapped selected autoimmune disorders to different terminologies to identify concepts and terms for data collection. Then we included CD patients from a local registry and completed by querying the hospital clinical data warehouse (CDW). Finally we identified status for selected autoimmune diseases for these patients by reviewing their EHR for mapped concepts.

To define autoimmune comorbidities burden in CD, we first extract a specific set of autoimmune disease. An initial list of autoimmune diseases was defined as: all the children of concept ‘Autoimmune Disease’ (D001327) in the MeSH® hierarchy, i.e., all the descriptors with a MeSH® Tree Number starting with C20.111. We identified amongst this list the autoimmune diseases most frequently associated with CD in MEDLINE® using the MeSH® co-occurrence file (MRCOC) provided by the U.S. National Library of Medicine as part of the UMLS Metathesaurus (2014AB release). The MRCOC file contains the number of times each pair of MeSH® descriptors occurred in MEDLINE® citations. We extracted all pairs of MeSH terms that contained both the MeSH descriptor for Celiac Disease (D002446) and any MeSH descriptor for autoimmune diseases. We arbitrarily restricted the list for this study to the 15 autoimmune diseases the most co-occurring with CD. This process allows to focus on a domain of comorbidities using the MeSH hierarchy and to a subset of comorbidities using the number of co-occurrences in MEDLINE®.

We used the Unified Medical Language System (UMLS) to map autoimmune disorders MeSH® concepts to other terminologies used in the CDW. To leverage diagnosis codes we used the International Classification of Diseases version 10 (ICD-10). We used the Anatomical Therapeutic Chemical Classification System (ATC) whenever a drug was specific to an autoimmune disease. These terminologies (MeSH, ICD-10 and ATC) also provided terms to constitute a catalog of words for helping textual data review.

The HEGP hospital has a specialized consultation for CD patients with about 50 new patients recruited per year in the last decade. The gastroenterology department has maintained a list of patients with CD for research purposes. The study has been approved and data access granted by the following commissions: the “Commission Nationale de l’Informatique et des Libertés” (declaration #174350) and the HEGP institutional review board (registration #00001072).

To complete the list of patients maintained by the gastroenterology department, we also extracted patients fulfilling the three following criteria: (i) at least one encounter in the 2000-2014 period of time (ii) presence of at least one ICD 10 code for CD (K90) in billing claims; (ii) one or more hospitalization stay or consultation in the gastroenterology department and (iii) at least one text document (discharge or letter) containing the term ‘celiac disease’ or one of its synonyms. We extracted these data from HEGP i2b2 CDW [25]. This CDW contains routine care data divided into several categories among them demographics (age, sex, and vital status), vital signs (e.g., temperature, blood pressure, weight…), diagnoses (ICD-10), procedures (French classification), clinical data (structured questionnaires from EHR), free text reports, pathology codes (French ADICAP classification), biological test results (LOINC), and Computerized Provider Order Entry (ATC) drug prescriptions [28].

The study population counted 741 patients and a corpus of 6340 clinical reports (patients’ inclusion flowchart available in Fig. 1).

We queried the CDW to identify for each patient the presence/absence of the selected autoimmune diseases. We used the selected ICD-10 diagnosis codes to query billing data and selected ATC codes to query Computerized Prescription Order Entry data.

We extracted every narrative report including discharge summaries, outpatient reports, multi-disciplinary expert meeting summaries and letters from all departments of the hospital and reviewed them manually to extract selected autoimmune disease status for each patient. To facilitate the manual review, we developed a browser-accessible software, FASTVISU [29], interfaced with the CDW to display and explore all the documents for a given patient. FASTVISU highlights terms with an entity recognition module based on a list of regular expressions (e.g., the pattern \bdiab\w + \b to match for diabetes) defined by the user and approximate matching techniques. For the 15 selected autoimmune comorbidities, a set of regular expressions was established, and used to query the CDW and obtain the CD corpus; then two trained physicians reviewed all the clinical narratives in the CD corpus using FASTVISU to validate the presence/absence of each of the selected autoimmune diseases for each patient.

Because autoimmune diseases are chronic diseases, a patient was considered having the autoimmune disorder if the condition was ascertained at least once in her/his EHR. The analysis was performed considering all sources together then each source separately.

Figure 2 illustrates our workflow.

Patients’ characteristics were summarized using median and interquartile ranges for quantitative variables and proportions for qualitative variables. We measured Sperrin’s I coefficient [30] to evaluate temporal irregularity of the data recorded in the EHR, denoting whether some encounters would provide more data and, consequently, whether other time periods would be less covered. Sperrin’s I coefficient was calculated for each patient with at least 2 encounters using formula (1):

Reviewers’ mutual-agreement in the text review was evaluated with Cohen’s alpha coefficient.

For each autoimmune disease, a case was defined by at least one ICD code, one drug prescription, or one mention in text. For each prevalence estimate, we computed the 95% confidence interval or Wilson’s interval if the proportion was close to 0%.

To compare the contributions of text, ICD codes and drugs to estimate autoimmune comorbidities, we computed the proportions of cases detected by text alone, structured information (ICD code or drug prescription on CPOE) alone, or both.

Finally, to assess the performance of literature-based comorbidities selection, we estimated the correlation between MeSH® and diseases prevalence, using Spearman’s correlation coefficient between number of publications indexed with MeSH® terms for both CD and an autoimmune disease, and the prevalence the corresponded disease obtained from our EHR extraction.

Statistical analysis was conducted with R (version 3.1.2).

The mining of EHR data allowed us to include 741 patients, one of the largest population of adult CD patients used to report autoimmune diseases prevalence in CD patients to the best of our knowledge. In this hospital based study, the prevalence of autoimmune comorbidity was 18.1% (95% CI 15.4 –21.0). The three most prevalent autoimmune comorbidities were thyroiditis with a prevalence of 12.6% (95% CI 10.1–14.9), type 1 diabetes mellitus with 2.28% (1.2 –3.4) and dermatitis herpetiformis with 2.0% (95% CI 1.0–3.0).

We compared our prevalence estimates with literature as an external validation. Our disease prevalence estimates are in the highest range compared to other published studies [14‐20, 23]. The first explanation is that we assessed prevalence from a hospital-based cohort from a CD specialized center, while most non-complicated CD disease, therefore with no additional autoimmune burden, are likely to be followed-up in ambulatory care only. Moreover, our study benefited from the coverage of the CDW, which includes a long follow-up and, therefore, increases the probability of mentioning an associated autoimmune disease. The quality of this longitudinal source of information was measured by Sperrin’s coefficient, which demonstrates a broad coverage of text documents during the follow-up period. In contrast, prevalence studies based on questionnaires [14, 15, 21] may underestimate prevalence, e.g., due to memory bias.

It would have been of interest to extract diagnosis date, but as we expected many missing data and approximations due to early childhood diagnoses, we did not extract this information.

In this study, most diagnoses were ascertained through text reports. This finding is consistent with the review by Shivade et al. of 97 studies using EHR for phenotype identification [2]. A typical example is thyroiditis, where about 98% of the cases were found in the text reports (92.5% only in text, and 5.3% in both text and structured data).

Few diagnoses cases were ascertained through ICD-10 codes. In France, as in many countries, ICD-10 coding is primarily used for billing purposes and limited to inpatients. Consequently, the coding does not aim to cover all the patient’s conditions [31]. Moreover narrative reports include extensive information such as a dedicated medical history section [10]. Additionally, autoimmune disease cases were identified in our study using documents produced during outpatient visits during which no ICD-10 codes were collected in line with French regulation (no ICD-10 codes are produced during outpatient visits in France). Similarly, Wei et al. analyzed the respective contributions of clinical notes, ICD codes and medications for detecting ten diseases in EHRs and showed that clinical narratives offered the best sensitivity (0.77) [12].

Our results showed the benefits of combining text mining and structured data extraction. Other examples are found in the literature in colon cancer [11], atrial fibrillation, Alzheimer’s disease, breast cancer, gout, human immunodeficiency virus infection, multiple sclerosis, Parkinson’s disease, rheumatoid arthritis, and types 1 and 2 diabetes mellitus [12].

To the best of your knowledge there was no clear synthesis of major CD autoimmune comorbidities. The novel approach mining literature presented in this study enabled to identify relevant comorbidities as attested by the fact that the attention from the literature was coherent with the prevalence found both in the literature and in our cohort: autoimmune dysthyroidism or type 1 diabetes mellitus appeared in the top co-occurring MeSH® terms and these comorbidities were described in the literature as being the most prevalent autoimmune comorbidities in CD patients [16, 20]. Furthermore, the three most prevalent autoimmune diseases in our adult CD population appeared as the top three in the co-occurrence ranking. Our method is flexible as domain restriction using MeSH® hierarchy and limiting the number of results with the number of co-occurrence are both optional, although we haven’t evaluated this method without these two types of restrictions. Moreover, this novel approach provided us with the most recent list of autoimmune diseases associated with CD in the literature. This is of interest because research subjects evolve over time and in this light Medline acts as a biomedical research collective memory and an up-to-date view of clinical expertise. For example, based on MEDLINE co-occurrences before year 2000, Autoimmune Hepatitis would not be in the top 15 selection, but Pemphigus and Pemphigoid, Bullous would be (see Additional file 2). Combined with EHR mining give us prevalence estimate of comorbidities that were not suspected as being associated to CD at the time patients’ diagnoses were recorded. Another advantage of this data-driven selection is to provide an automatable alternative to the usual elicitation step which classically determines relevant comorbidities by domain experts. This method allows to design more pragmatic studies, not relying on one or two experts’ opinion.

The workflow of this study, i.e. comorbidities selection from the MeSH® co-occurrences file, mapping from MeSH® to other terminologies, and case identification through text and coded data mining, can be automated to estimate comorbidities burden in other EHR-based studies.

Our workflow is in line with RECORD statement [32], in particular reporting a complete list of codes and algorithms for each comorbidities.

Moreover, we demonstrated that manual review could be performed easily using text visualization tools integrated with the CDW, even for non-English language based EHR [29].

Phenotyping quality is sometimes considered as a limit of EHR reuse. In the proposed workflow, we reinforce phenotyping quality by manual phenotype extraction by two readers in reasonable time thanks to assisted extraction using a visualization tool, FASTVISU. While FASTVISU is based on regular expressions which lack of precision, our workflow could be improved using a natural language based tool.