Background/Purpose: Delayed diagnosis and treatment of ankylosing spondylitis (AS) contribute to the economic, physical and psychological burden on patients and their caregivers. The objective of this analysis was to refine a previously developed predictive mathematical model for AS¹ based on features observed in the medical histories of patients with and without a diagnosis of AS to aid in the earlier identification of AS.

Methods: This retrospective cohort study used administrative claims data from > 182 million patients in the Truven Health MarketScan® Commercial and Medicare Supplemental Databases from January 2006 to September 2015 (Segment 1) and October 2015 to November 2016 (Segment 2). The AS population in Segment 1 included all patients with ≥ 2 diagnoses of AS (ICD-9-CM 720.0) by rheumatologists ≥ 30 days apart who had ≥ 12 months of continuous enrollment prior to first AS diagnosis. Control patients were matched by age, sex, enrollment period and geographic location and were randomly selected from the same database. Mutual information was used to identify features that differentiated AS from the control population; select features were then used as inputs in development of a suite of predictive models using data from Segment 1.¹ The optimized predictive model was then tested by observing whether patients predicted to have AS in Segment 1 subsequently received an ICD-10-CM AS diagnosis code (M45.x or M08.1) in Segment 2.

Results: In the initial study, 3 iterations of the predictive risk model (Models 1, 2 and 3a/3b) were developed using data from patients with ≥ 1 ICD-9-CM AS diagnosis during Segment 1, with each subsequent model iteration using additional AS-specific queries to reflect a more real-world situation.¹ A 2-stage model (Models 4 and 5) was built using patients with ≥ 2 ICD-9-CM AS diagnoses during Segment 1 to improve performance. Model 4 identified patients with AS from the general control population in the database (Figure 1A). To further improve precision, Model 5 was built using 50,000 random patients in Segment 1 who scored above a 0.5 in Model 4 (i.e., were more like patients with AS) as a new control population (Figure 1B). Models 4 and 5 were then combined to identify new patients with AS in Segment 2; of the ≈ 20 million patients who tracked across Segment 1 and Segment 2, 742 patients had an ICD-10-CM AS diagnosis in Segment 2 who did not have an ICD-9-CM diagnosis in Segment 1.

Conclusion: Predictive models for AS diagnosis were developed and refined in this US administrative claims database. Work is ongoing to compare the sensitivity and positive predictive value of these predictive models versus more traditional linear regression models, and future research will validate the model in a separate, commercially insured population database.

Reference:

1.)   Garges C, et al. Presented at the ISPOR 22nd Annual International Meeting; May 20-24, 2017; Boston, MA; [Poster PRM85].