Medical claims data, electronic medical records (EMRs), and electronic health records (EHRs) are familiar sources of real-world data (RWD). They are often used secondarily to complement limitations in clinical trials. For example, they can characterize patient subgroups that are excluded from clinical trials by following eligibility criteria such as comorbidities or age. Findings obtained through long-term, naturalistic observations of a large and diverse patient population can be easily generalized to other populations. Other advantages are that these data have high external validity, a single data source can be used for different study purposes, and prospective data collection is not required.

Claims data are electronic records of transactions between patients and health care providers. They include information on bills (claims) submitted by providers (hospitals, clinics, and pharmacies) to third-party payers (health insurance associations). There are already some large-scale commercial and nonprofit claims databases available in Japan [1-6] that aggregate information from multiple health care providers for secondary use. Recently, the EMR and EHR data have become widely available. The EMR data are the details of the encounters with patients recorded by physicians through EMR systems. They contain rich clinical information such as laboratory test results, diagnostic images, pathology findings, and patient symptoms. As different facilities may use different EMR systems, domestic EMR data are currently available from ≥1 medical institution. The EHR data are electronic records of all health-related information of individual patients created and managed by clinical professionals, which can be shared and used among various medical facilities. Current EHR databases in Japan include both patient claims data and medical records.

In recent years, claims data, EMRs, and EHRs have been increasingly used in clinical epidemiology studies. Such studies include cost-effectiveness analysis of drugs (including disease burden and assessment of medical technology), risk factor analysis, investigation of the actual status of drugs (including preclinical feasibility valuation, marketability study, and detection of prescription patterns), and evaluation of drug efficacy in actual clinical practice. Because these data are not designed for research purposes, the secondary use requires an understanding of their limitations and the ability to generate clinical questions, epidemiological skills to construct a study design, and statistical skills to analyze retrospective observational data. Previous approaches have addressed the limitations and challenges of using these data [7-12]. Understanding their application status based on these advanced guidelines is essential. However, investigations of existing epidemiological studies based on these data are lacking.

We conducted a comprehensive narrative literature review to clarify the secondary use of claims, EMRs, and EHRs in clinical epidemiology in Japan. We focused on 7 categories of key information, including organization, study design, RWD type, database, disease, outcome, and statistical method. We expect that this review would help in the design of high-quality epidemiological studies.

This is a comprehensive narrative literature review that investigated the secondary use of claims data, EMRs, and EHRs in epidemiology in Japan. Referring to PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [13] and procedures used in previous review studies [14-18], we conducted this review by searching for biomedical articles in PubMed.

We searched peer-reviewed publications that satisfied the eligibility criteria for this narrative literature review in PubMed from January 1, 2006, to June 30, 2021 (the date of search).

Keywords used to search PubMed consisted of “real world,” “database,” “claim,” “receipt,” “administrative,” “emr,” “ehr,” “japan,” “electronic medical record,” “electronic health record,” and Medical Subject Headings (MeSH) terms including, “Electronic Health Record,” “Administrative Claims, Healthcare,” “Insurance Claim Review/statistics and numerical data,” and “Japan/epidemiology.” We initially identified related articles by using various combinations of these keywords. The details of the search string are available in Multimedia Appendix 1.

On the basis of the search strategy, we identified articles whose titles and abstracts satisfied the following three inclusion criteria: (1) involvement of claims, EMRs, EHRs, or medical receipt data; (2) mention of Japan; and (3) published from January 1, 2006, to June 30, 2021. Eligible articles were then filtered out by satisfying any of the following six exclusion criteria: (1) review articles; (2) non–disease-related articles; (3) articles in which the Japanese population is not the sample; (4) articles without claims, EMRs, or EHRs; (5) unavailability of full-text articles; and (6) articles without statistical analysis.

The second author (TT) conducted the article search based on the search strategy. Both authors jointly reviewed all searched publications and performed 2 rounds of screening to identify target eligible articles. In the first round, we removed duplicates and articles that met any of the 6 exclusion criteria by screening the titles and abstracts. Review articles were automatically identified by a section classification model [19] trained on the PubMed 200k data set [20], which classified sentences in the abstracts into 5 sections (introduction, objective, method, result, and conclusion). On the basis of the hypothesis that review articles do not have sentences describing the results, we considered those without result sentences as review articles and removed them from the target articles. Artificially, we filtered out articles that met the exclusion criteria (2)-(5). In the second round of screening, the first author (YZ) reviewed the full text of the remaining articles and removed those that did not include statistical analysis. The 2 authors double checked the results to ensure accuracy and finalized the eligible articles.

We performed a descriptive statistical analysis of the collected data by describing their counts and percentages. In addition, we calculated the percentages of outcomes and databases for each disease. The percentages of statistical methods used to assess different outcomes were also analyzed. All codes used for data collection and descriptive analyses were performed using Python (version 3.8.8, 2021).

A total of 620 eligible articles were identified for this narrative literature review. Figure 2 [13-18] illustrates the selection process and the results of each screening step. We also illustrate the publication years of these articles in Multimedia Appendix 3. The distribution indicated that 68.7% (426/620) of the articles were published after 2018, suggesting that the secondary use of the 3 RWD types in epidemiological research in Japan was prevalent in approximately the last 5 years.

A comprehensive narrative literature review was conducted to understand the secondary use of nationwide claims data, EMRs data, and EHRs data in clinical epidemiology in Japan. On the basis of the search strategy and eligibility criteria, a total of 620 eligible articles were identified from PubMed between January 1, 2006, and June 30, 2021 (the date of search).

We quantified 7 categories of key information from these 620 eligible articles. The main findings were that (1) most of the research has been done by academic institutions, whereas nonacademic institutions tend to collaborate with academic institutions; (2) the cohort study was the major design that longitudinally measured outcomes of proper patients; (3) most studies used claims data; (4) the JMDC, DPC database (MHLW), MDV, and NDB were mostly used, whereas only a few studies used EMRs or EHRs from a single hospital or multiple hospitals, which do not have a large patient population; (5) the top rank of diseases studied in the current research were infections, cardiovascular diseases, neoplasms, and nutritional and metabolic diseases; (6) treatment patterns, physiological or clinical outcomes, and mortality were the most assessed in these articles; and (7) multivariate models were commonly used, during which logistic regression and linear regression were shown to be the first choice for analyzing categorical variables and continuous variables, respectively.

The findings on the percentage of outcomes for different diseases hint at the tendency of existing studies to examine different diseases. For some common, chronic, and psychiatric diseases, current studies tended to assess their treatment patterns, whereas for some sudden onset severe diseases, patient mortality and hospitalization or hospital stay were assessed more often. Existing studies have focused more on assessing treatment modalities, physiological or clinical outcomes, and mortality when targeting diseases such as infections, cardiovascular diseases, and neoplasms. Furthermore, although strong trends were detected between eye diseases and treatment patterns, hemic and lymphatic diseases versus mortality, and mental disorders versus mortality (Figure 3), it was difficult to draw any conclusions that reflect clinical importance because of the small sample size. However, these results indicated different distributions of outcomes measured in different diseases, from which we can learn the focus and shortcomings of the existing studies. In addition, the total number of studies measuring guideline adherence was relatively small (n=32). During this period, 63% (20/32) of the studies were conducted on “cardiovascular diseases” and “nutritional and metabolic diseases.” These results also revealed a relative lack of studies measuring guideline adherence in infections. We expect that RWD research on guideline adherence would receive more attention in future.

The percentage of databases used for different diseases implied the selection of databases for observing different diseases. The JMDC databases and DPC database (MHLW) showed opposite use trends in diseases, especially nutritional and metabolic diseases, musculoskeletal diseases, mental disorders, hemic and lymphatic diseases, eye diseases, and skin diseases.

According to the investigation of statistical methods used to assess different outcomes, multivariate models were the most commonly used in assessing mortality. Regardless of the outcome, multivariate modeling was accompanied by adjustments for various confounders (Figure 4B). Mortality, hospitalization or hospital stay, and resource use or costs have been analyzed using multilevel models or marginal models (eg, GEE) more than others. This implies that hospital-related outcomes tended to be assessed by models that took clustering into account. Logistic regression was the first choice for measuring many of the outcomes, with the exception of hospitalization or hospital stay and resource use or costs, for which linear regression was commonly used. Cox proportional hazards regression was suggested as the second choice when assessing mortality, physiological or clinical outcomes, and treatment patterns. Although the PS technique has been proven effective in balancing confounders between groups, it has not been widely used in existing studies. There is a relative preference for this technique in studies assessing mortality.

In this subsection, we compare this review with 2 similar studies [27,28]. Hirose et al [27] conducted a narrative review of 68 studies on the secondary use of claims data in a specific database, NDB, from October 2016 to June 2019. They summarized 5 key pieces of information, including study design, research area, setting or sample, outcomes, and strengths and limitations. Subsequently, Fujinaga and Fukuoka [28] conducted a similar narrative review of 643 studies on the secondary use of claims data in 4 large-scale domestic databases: NDB, DPC database (MHLW), JMDC, and MDV, from January 2015 to October 2020, from which 3 categories of research type, design, and area were analyzed descriptively. Both studies used a classification of the journals in which the target articles were published to extract information about the research area [29]. These classifications mixed disciplinary categories, such as clinical medicine, pharmacology and pharmacy, pharmacology and toxicology, and immunology; disease categories, such as infectious diseases; and general categories, such as social sciences and public environmental health. In addition, only the primary outcomes were analyzed in these 2 studies. As a result, the distribution of articles in each category was summarized in these studies.

Because of the partial overlap in search periods, as well as the fact that PubMed was used for the search, there were some articles that were reviewed in both this study and these 2 prior studies. In contrast to these 2 studies, which used 1 or more specific claim databases without specifying a research area, our review investigated domestic epidemiological studies based on the secondary use of 3 types of RWD: claims, EMRs, and EHRs. A further difference is that we defined 7 categories for data collection to assess the status and trends of the existing studies. One of the novelties is that we classified the outcomes with reference to the paper by Abaho et al [24] paper and collected information on all the outcomes measured in the target articles. The advantage of this classification is that these outcomes are also applicable to clinical trial studies and can be automatically identified from biomedical articles [24]. Another innovative point is that we proposed a hierarchical approach to classify the statistical methods that appear in the target articles. For the results of the data collection, we summarized the distribution of the target articles in each category. Additional comparative analyses were performed for diseases versus outcomes (Figure 3), outcomes versus statistical methods (Figure 4), and diseases versus databases (Multimedia Appendix 5), which revealed trends in the assessment of outcomes across different diseases, trends of statistical methods used for different outcomes, and trends in database selection when analyzing different diseases. Moreover, our findings shed light on the focus and shortcomings of previous studies.

In addition, we identified several other review studies on the secondary use of RWD data [30-32]. The paper by Ferver et al [30] provided a narrative review of 1956 claims-based studies in 5 health care journals from 2000 to 2005 by summarizing the research types and areas. The paper by Hutchings et al [31] provided a systematic literature review of 18 studies to investigate the attitudes of relevant practitioners toward the secondary use and sharing of health administrative and clinical trial data. Schlegel et al [32] conducted a literature review of 941 studies on the secondary use of health care data in 2016 to select the best performing articles. We summarized these additional studies to understand other investigations on the secondary use of RWD data. Comparisons were not made because of the survey years or different research purposes.

The first limitation of this review is that we only searched the literature in PubMed, which may have led to significant publication bias. Second, we only investigated studies conducted in Japan. In the future, a comparison of studies from other countries, such as the United States, will be necessary to understand the Japan-specific trends of such studies. In addition, searches of multiple electronic databases should be considered to reduce potential publication bias.

In this subsection, we discuss the future perspectives for the use of claims, EMRs, and EHRs in epidemiology in the Japanese context, in terms of the findings of this large narrative literature review.

This literature review provides a good understanding of the current status and trends in the use of claims, EMRs, and EHRs in clinical epidemiology in Japan. The results demonstrated appropriate statistical methods regarding different outcomes, Japan-specific trend of disease areas, and lack of use of artificial intelligence techniques in existing studies. We hope that the results of this narrative review will provide useful information for researchers to design relevant studies. In the future, a more precise comparison of relevant domestic research with worldwide research will be conducted to clarify the Japan-specific status and challenges.