Chart validation of an algorithm for identifying hereditary progressive muscular dystrophy in healthcare claims

This was a retrospective cohort study using medical chart validation to evaluate an ICD-9/10 coding algorithm in a large administrative claims database. The Health Core Integrated Research Database (HIRD®) is a longitudinally integrated database of medical and pharmacy claims and electronic laboratory results from a large commercial health insurer. HIRD contains claims data from 14 commercial health plans geographically dispersed across the United States. An independent institutional review board (IRB) has reviewed the study and granted a waiver of authorization from informed consent to allow access of medical charts.

Patients with a diagnosis of hereditary progressive MDs (ICD-9 359.1 or ICD-10 G71.0) between January 1, 2013 through December 31, 2016 (study period) were identified in HIRD. According to the algorithm, patients were required to have at least two office visits with MD diagnosis, to be males and younger than 18 years at the time of the first MD diagnosis. To maximize the amount of information obtained from patient medical charts to ascertain cases, we sampled patients who had a neurologist, primary care physician (PCP) or pediatrician involved in MD care in the chart study. Patients were excluded if they enrolled in Administrative Service Only plans or were not currently active members of plans.

The diagnosis of MD and specific type (Duchenne, Becker, or other MD) were confirmed by medical chart review (considered the gold standard) by a third-party medical chart abstraction vendor. Two medical chart reviewers (certified nurses with training) performed chart reviews. As part of the quality control process, the first 10 cases selected by the chart reviewers were also reviewed by two of the authors and a clinical expert. The cases were confirmed or reversed based on chart review and further discussion until consensus was reached.

Patient demographic, clinical, and health utilization characteristics were obtained from administrative health insurance claims during the study period. Demographic and clinical characteristics included age, race and ethnicity (obtained during chart review), region, length of health plan eligibility in years, and comorbidity burden such as pediatric complex chronic condition (CCC) index [11], bone health issues (eg, fragility, osteoporosis, or vertebral fractures), impaired growth, puberty delay, and apnea (Additional file 1: Table S1). Healthcare utilization variables included inpatient hospitalizations; outpatient or emergency department (ED) visits; and PCP or specialist visits.

Positive predictive value (PPV) and 95% confidence intervals (CI) were calculated using a 2 × 2 contingence table. The PPV for MD was calculated as the total number of confirmed cases in medical charts divided by the total number of patients with MD found in claims for which we obtained medical records. The percentage of specific MDs was calculated as total number of confirmed specific MD cases in medical charts divided by the total number of general MD cases in medical charts. Patient characteristics were summarized using basic descriptive statistics. All data were analyzed using SAS Enterprise Guide 7.1 (SAS, Cary, NC).

The results of this study demonstrate that the case-finding algorithm can accurately identify patients with MD, primarily Duchenne MD, within a large administrative database. There was only one other study by Soslow et al. that established a claim-based algorithm to identify patients with Becker and Duchenne muscular dystropy. They reported PPV of 77%, but one limitation of the study was that they only looked at hospital encounter data, thus potentially missing a big proportion of patients with MD [12]. Our study was strengthened by the ability to capture events across physician office visits and hospital encounters. Our algorithm, which was constructed using a few items that are easily accessible from claims, achieved a PPV of 95% for MD and 85% for Becker and Duchenne MD. A deeper look into the types of MDs shows that majority of the cases were Duchenne MD (74%), followed by Becker (11%) and other MD types, consistent with known etiologies.

The algorithm achieved similar PPV for Becker and Duchenne MD when compared to the algorithm in Soslow et al.’s study. Yet, the two studies had employed different strategies to optimize the specificity of algorithms. In Soslow et al. study, patients with a change in primary diagnosis to a different 359.x code were excluded. An additional set of clinical characteristics was used to exclude MD other than Becker and Duchenne MD (e.g., patients with early mortality, in ventilatory support, or cardiovascular disease at a young age). Our algorithm required at least two office visits with MD diagnosis because it is widely known that using only one ICD diagnosis code to identify cases could introduce inaccuracy and false positives. Supporting evidence from previous studies has shown that use of two or more ICD diagnosis codes improves the PPV of case-finding algorithms [9, 10].

Using longitudinal claims data, the study also depicted the characteristics of patients with MD. Progressive muscular damage and degeneration in patients with MD manifests in muscular weakness, motor delays, respiratory impairment, and cardiomyopathy. As the disease progresses, the susceptibility to respiratory infection, respiratory compromise, and cardiomyopathy increase. These were confirmed in our study population: the prevalence of cardiovascular diseases increased from 0% among patients younger than 4 years to 67% among patients age 19 years and older. Additionally, the prevalence of respiratory disease increased from 0% among patients younger than 4 years to 27% among those age 19 years and older (data not shown). The overall rate of cardiovascular disease was much higher in our study (49.5% vs. 29.7% in Soslow et al. study), mostly likely due to their exclusion of patients with cardiovascular disease in the young age. While Soslow et al.’s study primarily focused on cardiovascular morbidity, our study also looked at the bone health and endocrine-related comorbidities, which are common in patients with MD and exacerbated by glucocorticoid treatment [5]. We observed that nearly half of our study population had bone-related conditions, and more than 10% had endocrine-related conditions. Using available claim information throughout their health plan coverage period, we were able to capture disease burden as well as healthcare utilization.

Anticipated emergence of genetic and molecular therapies, advances in rehabilitation therapies, and the invention of non-invasive prenatal testing for MDs spark many questions in the care of patients with MD, such as optimum timing for initiation of new therapies and optimal, personalized treatments [5, 6]. Given the small patient pool available, clinical trials will be challenged recruiting populations to investigate these questions. Observational studies, such as those using claims data, have a great amount of utility in advancing the knowledge base of MD care. Claims data reflect the comorbidities, treatment patterns, and the safety and effectiveness of therapies in typical patients seen in real-world settings, providing insight into the actual burden of Duchenne MD.

The study has a few limitations. Firstly, the ICD9/10 diagnosis code available at the time of this study was a single code for general MD and it did not differentiate subtypes. Since claims do not contain adequate clinical information, our case-finding algorithm was unable to distinguish between subtypes. Specific ICD-10 codes for Duchenne/Becker MD and facioscapilohumeral MD are planned to be introduced in October 2018 [13], which are expected to enhance future claims-based research by improving the homogeneity of study populations. Secondly, our study only sampled cases obtained from claims, and thus was unable to report negative predictive value and sensitivity. For rare conditions, the ideal chart review method is impractical given the large amount of medical charts to be reviewed. Alternatively, several sampling strategies to select code negative patients have been implemented by investigators and researchers, but they are prone to bias introduced by disproportional sampling of code positive and code negative patients [14]. Our algorithm picked up 48% of the patients with at least 1 MD diagnosis code. The sensitivity of the algorithm is unknown and needs to be evaluated with further research. Thirdly, the chart validation study sampled cases who were all likely had MD according to the algorithm. As the chart reviewers were not blinded to the study design, bias may have been introduced to move the PPV upward. Fourthly, our study characterized the co-morbidity burdens of patients with MD using claims data. Prevalence may vary by the type of diagnostic instruments used to define disease, however claims-based definition generally doesn’t include such information. Claim-based diagnosis may also subject to claim coding omissions or errors. Lastly, our data were drawn from commercially insured members who may not be representative of members under government or public insurance coverage.