Validation of medical service insurance claims as a surrogate for ascertaining vitiligo cases

The epidemiology of vitiligo, especially its disease burden on the healthcare system, can be assessed indirectly by analyzing health insurance claims data. Validating this approach is integral to ensuring accurate case identification and cohort characterization. The primary aim of this study was to develop and validate an indirect measure of vitiligo ascertainment using health insurance claims data. These data were used secondarily to identify demographic characteristics, body site involvement, vitiligo subtypes, disease associations, and treatments. This study assessed the validity of identifying vitiligo from billing claims within a Canadian provincial universal health insurance program, versus vitiligo cases accrued from direct medical chart reviews. Claims-based algorithms combining ICD-9-CM diagnostic code 709 with treatment-specific data were derived and tested to identify vitiligo patients. This was compared against cases arising from the manual review of medical records of 606 patient with a diagnostic code for “dyschromia” (ICD-9-CM diagnostic code 709) from January 1 to December 31, 2016. Based on the chart reviews, 204 (33.7%) patients were confirmed to have vitiligo. 42 separate claims-based algorithms combining ICD-9-CM diagnostic code 709 with treatment data specific to vitiligo were modeled and individually tested to evaluate their accuracy for vitiligo ascertainment. One algorithm achieved a sensitivity, specificity, PPV and NPV of 86.8% (95% CI 82.1–91.4), 92.5% (95% CI 90.0–95.1), 85.5% (95% CI 80.7–90.3), and 93.2% (95% CI 90.8–95.7), respectively. There was a 2.2 female-to-male ratio. The most common medical treatments were tacrolimus (74.5%) and topical corticosteroids (54.3%). Hypertension (24.2%) and hypothyroidism (19.6%) were the predominant co-morbidities associated with vitiligo. Health insurance claims data can be used to indirectly ascertain vitiligo for epidemiologic purposes with relatively high diagnostic performance between 85.5 and 93.2%.


Introduction
Vitiligo is a chronic, progressive, autoimmune skin condition characterized by circumscribed and patchy loss of skin pigmentation. It is estimated to affect 0.5-2% of the global population; the burden of disease in Canada is undetermined [1,2]. Vitiligo can be associated with a decreased quality of life as it causes physical disfigurement with known impacts on ethnic identity, self-confidence, and mental health [3].
Validated population-based research is integral to understanding the epidemiology, etiology, natural course, treatment patterns, disease burden, and associated conditions of vitiligo. Health data routinely collected for administrative and clinical purposes are being increasingly used to identify and study dermatologic conditions, including vitiligo [4,5]. Such health data include, but are not limited to, billing codes, diagnostic codes, and prescription 1 3 treatments. This information is considered secondary data in that it is not collected specifically for research; therefore, it necessitates careful validation to ensure appropriate and accurate use.
Vitiligo is typically diagnosed on clinical grounds without ancillary testing. A Wood's lamp can aid in visualizing depigmented patches, and rarely, a skin biopsy is done to confirm the absence of localized melanocytes and epidermal pigment [6]. As opposed to other dermatologic conditions, such as skin cancer, in which pathology-based databases are considered highly reliable for case ascertainment, vitiligo's clinical heterogeneity and infrequent use of formal diagnostic testing can pose challenges to using routinely collected health data to identify patients.
To conduct population-based research on vitiligo using health administrative datasets, surrogates for the diagnosis of vitiligo must be validated. There is a lack of ascertainment methods with validation for these indirect secondary data sources. This limits dermatoepidemiologic research as methodological differences in identifying patients with vitiligo can impact study findings and therefore the meaningful interpretation of results. This challenge has been appreciated in other areas of dermatology as well. A 2018 systematic review by Dizon et al. found that of 59 studies that used routinely collected health data to identify patients with atopic dermatitis, only two described validation of their methods; furthermore, the differences in methodology were associated with up to a threefold variation in prevalence estimates [7].
Many studies have shown that health administrative data can be successfully used as a surrogate for various health outcomes. The International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM), developed by the World Health Organization (WHO), classifies medical conditions and diagnoses using standardized codes [8]. Many studies utilize ICD-9-CM to help validate coding algorithms designed to quantify diseases. With the goal of confirming that patients with a specified diagnostic code truly have the concordant disease, an accurate and acceptable algorithm would need to yield a high positive predictive value that could retrospectively be validated by chart review. Population Data BC (PopData) is a Canadian provincial government data resource that contains one of the world's largest collection of longitudinal health data. Health administrative data obtained through PopData, including physician claims data, have been used to define various skin conditions and several diseases that have been studied by our research team and collaborators [9][10][11]. Many of these conditions have been validated with PPVs > 90%. A cohort study by Zhang et al. used algorithms based on public health insurance claims to ascertain keratinocyte carcinomas with a positive predictive value (PPV) of 76.9% and negative predictive value (NPV) of 99.4% [10]. A study on eczema used algorithms with population-based claims and prescription medication data to ascertain the condition with a sensitivity of 96.4% and PPV of 77.7% [11].
This study sought to investigate indirect vitiligo ascertainment using health insurance claims algorithms based on diagnostic codes and prescribed treatments. In BC, medically required services are covered by the Medical Services Plan of BC (MSP), a universal public insurance plan, and submitted as claims by fee-for-service physicians [13]. BC residents and their dependents, with the exception of active members of the Royal Canadian Mounted Police and the Canadian Armed Forces, are legally required to be enrolled with MSP; accordingly, greater than 98% of BC residents are registered for MSP coverage. MSP claims require an ICD-9-CM diagnostic code for the condition being managed. Treatment data are documented both in patient charts as well as on BC's PharmaNet database, a central data system that contains every prescription dispensed by BC pharmacies. The purpose of this study is to validate vitiligo case finding based on billing and treatment data against a manual patient chart audit from a claims database of dermatologists in metropolitan Vancouver, BC's largest urban center. The study aims to develop ICD-9-CM data algorithms to identify vitiligo cases with high diagnostic performance in a proofof-concept cohort prior to conducting a larger province-wide validation study using cases identified through a populationbased database (i.e., PopData).
The primary aim of this study is to model and test the validity of MSP claims-based algorithms as indirect measures of vitiligo ascertainment in comparison to vitiligo as defined by manual chart review. The secondary aim is to assess the demographics, including age and gender, of those with vitiligo presenting for secondary medical care.

Ethics
This study has been approved by the University of British Columbia Clinical Research Ethics Board (UBC CREB #H18-01813).

Algorithm modeling
The diagnostic code for vitiligo billing follows the WHO's ICD-9-CM definition 709, for "dyschromia." The ICD-9-CM system provides a means of classifying and indexing healthcare diagnoses with applicability in health data statistics. Phototherapy treatment, including both psoralen and ultraviolet A (PUVA) and narrow-band ultraviolet B (nbUVB), is documented within patient charts. Prescribed medications are also recorded within patient charts. These specific treatments and prescription medications constituted the independent variables for algorithm development. Booleanbased diagnostic algorithms were derived using varying combinations of both procedural and medical treatments for vitiligo, such as phototherapy, tacrolimus, pimecrolimus, and topical steroids.

Study cohort and data abstraction
This retrospective study examined the patient charts (both electronic and paper) and MSP claims from a convenience sample of dermatologists who provide consultant-level care for vitiligo patients in Metro Vancouver, British Columbia (BC) in the Vancouver Coastal Health region. MSP is BC's universal insurance plan and claims require submission of an ICD-9-CM diagnostic code for medical services provided. MSP claims were searched to identify patients using ICD-9-CM diagnostic code 709 for the study period of January 1, 2016 to December 31, 2016 ( Fig. 1). There were no age limitations. The patient lists generated 641 cases of 709coded patients. A trained abstractor manually reviewed the complete medical records of 606 patients. 35 charts (i.e., 5.8%) could not be located, and these missing charts were excluded from analysis. A "vitiligo-positive" case referred to a dermatologist's diagnosis of vitiligo documented in the medical record within a 1-year period of the billing service date. No distinction was made between patients diagnosed with vitiligo during or prior to the study period. A "vitiligo-negative" case was defined as the absence of a vitiligo diagnosis in the medical chart within a 1-year period of the billing service date. As the ICD-9-CM diagnostic code 709 is not specific to vitiligo and accounts for all skin conditions that fall under the term "dyschromia", other skin diagnoses, such as hyperpigmentation, melisma, and lentigines, were also documented. For vitiligo-positive cases, if included in the medical chart, data on vitiligo subtype and affected body sites were also abstracted. Additional information on prescribed treatments and medications was recorded for inclusion in the algorithms, as were general demographic information including age and gender. Furthermore, when available in the medical record, non-dermatologic medical conditions were recorded. To maintain quality control, one of the study dermatologists (SK) reviewed a random 10% of all chart abstraction sheets. This study's definition of vitiligo is inclusive of all its subtypes, including non-segmental, segmental, and mixed types.

Data analysis
Data abstracted from the 606 medical charts were manually coded into a securely encrypted computer for statistical analysis using Microsoft Excel (2016) and SPSS Statistics (Version 25). An "algorithm-positive" was defined as a vitiligo Total number of charts unable to be located, excluded (n=35) Fig. 1 Flowchart depicting the method in which data were abstracted. Medical Services Plan (MSP) claims were searched to identify cases with ICD-9-CM diagnostic code 709. Of the 606 cases identified, 204 were positive for vitiligo case identified by one or more of our proposed claims-based data algorithms. An "algorithm-negative" was defined as a vitiligo case missed by the algorithms. A total of 42 claimsbased algorithms combining the ICD-9-CM diagnostic code 709 with procedure and treatment data specific to vitiligo were tested to evaluate the accuracy of vitiligo ascertainment. The sensitivity, specificity, PPV, and NPV for each algorithm were calculated with 95% confidence intervals. The demographics, including age and gender, of vitiligo patients were assessed using descriptive statistics.

Claims-based algorithms for vitiligo ascertainment
A total of 42 separate claims-based algorithms combining ICD-9-CM diagnostic code 709 with treatment data specific to vitiligo were tested to evaluate the accuracy of vitiligo ascertainment. Of these, seven achieved vitiligo ascertainments with adequate sensitivities and PPVs greater than 85% [14]. Algorithms including only phototherapy were very specific (96.5-98.8%), but lacked adequate sensitivity (52.0-55.4%). The inclusion of either tacrolimus or pimecrolimus further increased the specificity, sensitivity, and PPV. Algorithms including or excluding steroids did not prove adequate in their ascertainment of vitiligo. The exclusion of antibiotics and benzoyl peroxide, medications more commonly used in other disorders of dyschromia, significantly increased the PPV.

Discussion
Routinely collected health data are increasingly being deployed for administrative and clinical purposes to study dermatologic diseases, such as vitiligo. As this data source is not generated for research per se, its validation is imperative. The variability in methods for identifying vitiligo patients using this secondary data, as well as the overall lack of validation studies, makes it difficult to meaningfully appraise, reproduce and compare study results, as findings may be 1 3 distorted and inaccurate. Validation research is integral to ensuring patients with vitiligo are accurately identified and misclassification bias is avoided. This study provides validation of various vitiligo definitions based on routinely collected health insurance claims data by comparing these to vitiligo defined by detailed chart audit with respectable Fig. 2 a The distribution of the age at diagnosis of vitiligo. The ages ranged from 1 to 77 years. The mean age at diagnosis of vitiligo was 34 years. b The distribution of body sites affected by vitiligo. The most common areas affected were the face and neck, followed by the trunk, lower extremities, hands, upper extremities, genital region, and feet. Many cases had multiple body sites affected. c The distribution of vitiligo subtypes demonstrated that the majority of cases were classified as non-segmental, which included generalized, acrofacial, and unspecified subtypes. A minority of cases were segmental, focal or mixed  sensitivity, specificity, NPV and PPV. We proposed and tested the validity of 42 claims-based algorithms as potential indirect measures of vitiligo ascertainment in a dermatological dataset. Of those tested, algorithms 27 and 28 were the most diagnostically accurate as they attained the highest sensitivity, specificity, PPV, and NPV. These algorithms included patients prescribed tacrolimus, pimecrolimus, and/ or phototherapy, and excluded patients prescribed either benzoyl peroxide or antibiotics. To our knowledge, there are limited previously published claims-based models for vitiligo ascertainment. The peak age of onset in our study group was 34 years. The current literature has varied results with regards to age at diagnosis [6]. One study from Denmark found that 50% of people develop vitiligo after 40 years of age [15]. A study in India concluded that nearly half of patients present prior to age 20 years, and nearly 70-80% prior to age 30 years [16].
The discrepancy in age at diagnosis is likely due to global differences in accessing healthcare, as well as differences in the age of presentation for distinct vitiligo subtypes. While non-segmental vitiligo can present at any age, segmental vitiligo is more likely to present at a younger age [17].
Of note, while our study found a 2.2 female-to-male ratio, worldwide, men and women are estimated to have equal prevalence rates for vitiligo. Women may be more likely to access treatment for disorders of the skin, possibly due to the more negative social impact for affected women compared to men [6,18].
In keeping with current literature, our study found the face and neck to be the body areas most frequently involved by vitiligo patients seeking medical attention [19]. Our results were similar to that already published for extremity involvement, ranging between approximately 50-60%, as well as acral site involvement, ranging between approximately 70 Fig. 3 The most common medical and procedural treatments for vitiligo. A large majority of those treated medically were prescribed tacrolimus. Other common medications included betamethasone, hydrocortisone, clobetasol, and pimecrolimus. The most common procedural treatment was narrow-band ultraviolet B (nbUVB)  and 75%. Previous studies have suggested that the upper extremities and trunk tend to be more affected after longer disease duration [20]. Differences in distribution patterns according to time were not examined in our study. With regards to subtype, while less than one-quarter of cases in our study documented a specific vitiligo classification, our findings were in keeping with previous studies which demonstrate that non-segmental vitiligo is more common than segmental vitiligo [21,22].
The British Association of Dermatologists and Vitiligo Guideline Subcommittee of the European Dermatology Forum both recommend management of vitiligo with tacrolimus, topical steroids, and narrow-band ultraviolet B phototherapy [23,24]. In our study, nearly all patients with vitiligo were prescribed a topical medical treatment. The most frequently prescribed medications were topical steroids and calcineurin inhibitors. The inclusion of both tacrolimus and/or pimecrolimus increased the diagnostic accuracy of the algorithms in ascertaining vitiligo. Treatment with calcineurin inhibitors may be more specific for vitiligo, because if these patients require longterm treatment, calcineurin inhibitors have a lower risk of side effects, such as skin atrophy, telangiectias, and striae. While topical steroids are also very frequently prescribed as first-line treatment, neither the inclusion nor exclusion of topical steroids in the algorithms could accurately predict vitiligo. This is likely due to the medication's broad set of dermatologic indications, including skin disorders both with and without pigmentation concerns. In addition, as the sample of charts was exclusively from dermatologists, it is possible that those presenting had previously tried and failed steroid therapy with their primary care provider, which would not be documented in this dataset. With regards to phototherapy, while it is a commonly proposed second-line option for vitiligo, it is time-consuming and often difficult for the patient to access and maintain [6,25]. These barriers likely limited the sensitivity of algorithms including phototherapy.
Also in keeping with the previous literature, our results helped to confirm the known links between vitiligo and other medical conditions including hypothyroidism, asthma, allergic rhinitis, and depression. Of the autoimmune diseases, hypothyroidism had the strongest association with vitiligo. An earlier cross-sectional study found that nearly 20% of patients with vitiligo had at least one other comorbid autoimmune disease, with the most common conditions being thyroid disease and alopecia areata [26]. A previous retrospective population-based study confirmed a significant association between vitiligo and multiple autoimmune and atopic diseases, such as alopecia areata, Hashimoto's thyroiditis, myasthenia gravis, systemic lupus erythematosus, and atopic dermatitis [27]. Another study described vitiligo as a risk factor for major depressive disorder, as their results showed that vitiligo patients younger than age 30 years and greater than age 30 years had a 31% and 22% increased risk of developing major depressive disorder compared with the general population, respectively [28]. The comorbidity profiles in relation to vitiligo are likely to vary with age, sex, and region. Further algorithmic-based research could help elucidate vitiligo's co-morbidities and risk factors.  The study's strengths included the diversity of the population and the inclusion of detailed paper and electronic medical charts. Although one of the study's limitation is the inability to include individuals who did not seek medical care, our study focuses on the burden of those accessing the healthcare system; thus, the method takes a population-based approach specifically to those who actively seek medical treatment. Furthermore, as there is no true diagnostic "gold standard", the diagnosis of vitiligo is presumptive, although likely very accurate by trained dermatologists. Also, as this dataset was obtained from dermatology practices, where the prevalence of vitiligo cases and severity of disease are expected to be high, it is possible that the coding validity associated with the proposed claims-based algorithms is less applicable to a primary care provider's patient cohort. There are also the limitations inherent to chart review as compared to other approaches (e.g., prospective population study). Our algorithms' ability to indirectly ascertain vitiligo is a step toward ensuring sound population-based research using universal health administrative datasets. We plan to conduct further studies using Population Data BC that encompasses universal data of patients accessing the healthcare system to study further demographics, treatment patterns, and co-morbidities of vitiligo patients. This will also influence the planning of healthcare and utilization of medical resources.

Conclusion
Health insurance claims data can be used to ascertain vitiligo for epidemiologic purposes with relatively high diagnostic performance.
Author contributions Madelaine Bell collected the data and wrote majority of the manuscript. All the others contributed to the manuscript by writing and revising parts of the manuscript. The data was analyzed by all the authors. Sunil Kalia was the lead author in the study design and supervision.
Funding The study was supported by the Canadian Dermatology Foundation. Dr. Sunil Kalia is supported by the Photomedicine Institute, VGH & UBC Hospital Foundation, and the Michael Smith Foundation for Health Research.