(C-201) Validation of a claims algorithm for identifying patients with Duchenne muscular dystrophy in US administrative claims with genetic laboratory data

Background: Dystrophinopathies are a spectrum of heterogenous X-linked muscle diseases including Duchenne muscular dystrophy (DMD) and Becker muscular dystrophy (BMD) with DMD considered more rapidly progressive. While administrative claims can be a robust data source for real-world evidence studies, claim codes differentiating DMD and BMD do not currently exist.

Objectives: To assess accuracy of administrative claims-based algorithms for identifying patients with DMD using a novel tokenized database.

Methods: This retrospective cohort study combined deidentified, tokenized data from the Inovalon closed administrative claims database (claims) with consolidated genetic laboratory (lab) data. The study sample included US males with an ICD-10 diagnosis code for DMD/BMD (G71.01) between September 19, 2016 – December 31, 2023, ≥ 6 months continuous enrollment in claims, and linkable to the lab data as patients with a genetic lab test for DMD/BMD. The index date was defined as first G71.01 diagnosis code. A set of published DMD claims algorithms (Schrader, 2023) with 3 levels (broad, narrow, restrictive) was assessed in claims and compared against the reference standard, DMD-like genotype result in the lab data. Confusion matrices were used to assess performance of the algorithms through positive predicted value (PPV), sensitivity, specificity, negative predicted value, and kappa statistics with 95% CIs. A priori thresholds of 0.8 for PPV and 0.6 for Kappa were specified to prompt exploratory analysis with classification trees and random forest methods to improve the published algorithms if these thresholds were not met.

Results: A total of 5,682 patients in claims met inclusion criteria; 361 were linked with the lab data and had interpretable lab results. The median age at index prior to and post linkage was 16 (interquartile range, 9-26) and 8 (5-16) years, respectively. PPV was 71.4 (95% CI: 66.3-76.5), 78.0 (72.8-83.2), and 79.7 (74.4-85.1), in the broad, narrow, and restrictive algorithms, respectively. All algorithm levels had kappa statistics < 0.45. A random forest fit using a 70-30 training-test split indicated the “loss of ambulation (LOA) by age 12” criterion was unimportant. Improved algorithms, which removed the “LOA by age 12” criterion yielded a satisfactory PPV of 81.1 (75.9-86.4) and 83.5 (78.2-88.8) for the narrow and restrictive algorithms, respectively.

Conclusions: This work demonstrates the challenge in accurately identifying patients with DMD from administrative claims data where specific claims code is not present and exemplifies a strong case for data tokenization for more accurate patient identification. This study yielded an improved algorithm in an individual claims source, advancing the ability to precisely identify patients with DMD.