Introduction

Lyme disease (LD) (also known as Lyme borreliosis), caused by the Borrelia bacterium transmitted to humans by ticks in the genus Ixodes, is the fifth most commonly reported notifiable infectious disease [1] and the most frequently reported vector-borne disease [2] in the United States (US). According to the most recent surveillance reports, approximately 35,000 LD cases were reported to the US Centers for Disease Control and Prevention (CDC) via public health surveillance in 2019 [1]. However, the true burden of LD remains unclear. Because public health surveillance is a passive reporting system, underreporting of true cases exists. Healthcare claims data are another source that can provide information on LD patients. A recent study conducted by CDC researchers using claims data estimated that approximately 476,000 patients were diagnosed and treated for LD in the US annually during 2010–2018 [3], suggesting a remarkably larger clinical and societal burden of LD than estimated via surveillance data. The reliability of claims-based LD estimates, however, depends on the performance of the case-finding algorithms among other factors. To assess the true burden of LD and the effectiveness of the interventions to prevent and treat LD, it is important to use validated well-performing LD case-finding algorithms. We conducted a systematic literature review to identify LD case-finding algorithms used with US healthcare claims data and their validation status.

Methods

Information sources, literature eligibility, search, and selection

We searched PubMed and Embase databases and identified articles published in English since January 1, 2000 (search date: February 20, 2021), using the following search terms (not restricting to article titles or abstracts): (1) “Lyme disease”; and (2) “claim*” or “administrative* data”; and (3) “United States” or “the US*”. We then reviewed the titles, abstracts, full texts, and bibliographies of the articles to select eligible articles, i.e., those describing LD case-finding algorithms used with US administrative healthcare claims data. Fig 1 shows the process of the search and selection of the articles.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Search and selection of articles for the literature review on the US claims-based Lyme disease case-finding algorithms published in English since 2000.

Reference for the flow diagram: Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) Statement (https://doi.org/10.1371/journal.pone.0210242.g001).

https://doi.org/10.1371/journal.pone.0276299.g001

Results

From the database search using the prespecified search strategies, we identified 34 articles from PubMed and 27 articles from Embase. After removing duplicates, 43 articles remained. Of these, 27 articles were excluded because there was no indication of claims-based Lyme disease case-finding algorithms in their titles or abstracts. Of the remaining 16 articles, one article was excluded based upon a full text review. Therefore, 15 eligible articles were included in this literature review (Fig 1). The 43 articles initially identified through the database search are listed in S3 Table with reasons for exclusion for the excluded articles (n = 28).

Table 1 presents summarized information of the LD case-finding algorithms identified in the 15 articles. Seven studies [7–11, 14, 15] used algorithms identifying LD cases from claims data using LD diagnosis codes only (International Classification of Diseases, 9th Revision [ICD-9], 088.81; ICD-10, A69.2, A69.2x). Four studies [3, 12, 13, 16] used LD diagnosis codes and antibiotic dispensing records, and the remaining four studies [6, 17–19] used serologic test order codes (Current Procedural Terminology [CPT], 86617, 86618) in conjunction with LD diagnosis codes and antibiotic dispensing records. Three studies [3, 13, 16] used different algorithms for inpatient and outpatient settings. Only one study (in Tennessee, a low-incidence state for LD) [8] provided algorithm validation results from a medical records review. This review was conducted using medical records from a commercial health insurance plan and included individuals who were reported as LD cases to the Tennessee Department of Health (TDH) and subset of individuals who were not reported to the TDH. That algorithm used a LD diagnosis code alone; its sensitivity was 50% and positive predictive value was 5%. Quantitative or qualitative comparisons of the reported performance of different LD algorithms were not conducted in any of these studies.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Lyme disease case-finding algorithms used with the US administrative healthcare claims data identified in the articles published in English since 2000.

https://doi.org/10.1371/journal.pone.0276299.t001

Discussion

We identified and reviewed 15 articles describing LD case-finding algorithms used with US healthcare claims data. Of these, seven studies used algorithms with LD diagnosis codes only, and the other studies used algorithms with combinations of a LD diagnosis code(s), dispensing of antibiotic medications, and/or a serologic test order code(s). Only one of the studies provided results from algorithm validation [8], which showed that their LD diagnosis code-only algorithm identified only a half of true LD cases, and a large proportion (95%) of the cases identified in the claims data were false positives. The poor performance of this algorithm might be associated with a variety of clinical manifestations of LD, which makes diagnosing LD challenging, a potentially less sophisticated case-finding algorithm, and the accuracy of the claims data, which might not have been sufficiently high for LD case identification. The performance of LD case-finding algorithms might vary depending upon not just the level of sophistication of the algorithms themselves but also the accuracy of LD diagnosis by physicians and of coding in medical records and claims data, as well as geographic regions with varying LD incidences and where risk of a tick bite may vary. Previous studies documented that considerable proportions of LD patients reported through public surveillance lacked an ICD diagnosis code in their medical records [13, 20], and false positive LD cases were also common in medical records [21]. This suggests inaccuracies in medical records, and thus possibly claims data as well.

Six of the 15 articles compared LD incidence estimated using their LD case-finding algorithms with LD incidence from surveillance data [3, 11, 13, 14, 16, 17]. In these studies, claims-based incidence rates were higher than those from surveillance data. For example, one study that used an LD diagnosis code-only algorithm reported that claims-based LD incidence rate in Tennessee during 2000–2009 was >7 times the rate based on surveillance data [11]. Another study that used an algorithm with a combination of diagnosis codes and antibiotic dispensing records reported that claims-based LD incidence rate in the US during 2005–2010 was >11 times the rate estimated by surveillance data [13]. However, because only one algorithm was validated and there was variability in their study designs (e.g., data sources, inclusion/exclusion criteria of individuals, years under study), it is not feasible to determine which case-finding algorithms have better performance or whether there are any patterns in the incidence rates across different algorithms by comparing these incidence rate estimates. Future studies will need to validate LD case-finding algorithms with different designs and assess their performance.

Our findings provide a detailed description of the elements of each LD algorithm and can serve as a useful tool to aid interpretation across existing claims-based estimates. Further, this study is important to spur future research regarding the reliability of LD case finding using administrative claims data and the need to develop and validate claims-based LD case-finding algorithms. We searched only two databases (PubMed and Embase), and though this may be a limitation of our study, these databases are among the largest for biomedical and public health literature searches.

In summary, we found that diverse LD case-finding algorithms have been used with US claims data. Only one of the algorithms that used LD diagnosis code alone was validated, and it did not perform well. Further studies are warranted to assess algorithm performance for different designs and inform the efforts to better assess the burden of LD.

Supporting information