How the Switch from ICD9 to ICD10 Potentially Affects Risk Adjustment Models for the Privately Insured

Tuesday, June 25, 2019
Exhibit Hall C (Marriott Wardman Park Hotel)

Presenter: Randall Ellis

Co-Authors: Bruno Martins; Chenlu Song; Heather Hsu; Jeffrey J Siracuse; Tzu-Chun Kuo; Ying Liu; Arlene Ash

Diagnostic codes are widely used in the US for validating procedures, calculating risk-adjusted payments, tracking trends in disease prevalence and disease-specific mortality, and measuring and rewarding good provider performance. Thus, it is highly desirable for coding patterns to be stable over time, that new diagnoses are incorporated in a manner that does not greatly distort intended uses, and that results are only modestly vulnerable to gaming and coding manipulation. In October 2015 the US finally switched the diagnostic coding system used on insurance claims from the ninth to the much richer tenth revision of the US clinical modification of International Classification of Diseases, (abbreviated here as ICD9 and ICD10), expanding the number of available codes from 14,025 to 69,823. Lacking experience with the new coding system, risk adjustment models for the ACA Marketplace and private premium negotiations, among other uses, have not yet been recalibrated to take advantage of this dramatic change. Existing models merely map new codes into previously defined condition category (CC) clusters used for prediction, leaving payment weights on these existing CCs unchanged.

We will use health insurance claims data from IBM/Watson Truven Commercial Claims and Encounter data on over 20 million enrollees for 2012 to 2017. First, we examine stability over time in monthly levels of coding, in aggregate and for the CCs used in the ACA Marketplace risk adjustment model, to explore concerns about changes in the prevalence of coded conditions with ICD10. Second, we explore whether coding practice has changed more for certain plan types, provider specialties, or groups of procedure codes used by CMS to validate diagnostic codes as having come from clinicians. Third, we present preliminary results on the implications of extending an existing risk adjustment model to incorporate new coding dimensions, such as distinctions in laterality (left, right, unspecified, bilateral), in timing (initial, subsequent, or sequela), and other coding refinements used in ICD10. Fourth, we examine the potential value of separately modeling infants, children and adults (<age 65). We will expand the current work with 2015-2016 claims data, using 2017 Truven data, available in December 2018, to explore seasonality issues during 2012-2017.

Preliminary analyses on the 2015-2016 data find an upward trend but substantial stability in the aggregate level of diagnostic coding before and after the 2015:Q4 switch. Exploratory concurrent year regressions using 810 diagnostic clusters based on the AHRQ Clinical Classification System (CCS) find only modest improvements in simple models that distinguish laterality, timing, and age distinctions. New refinements will be considered.

This research is funded by the Agency for HealthCare Research and Quality (AHRQ) as part of a three-year project 1R01HS026485-01 entitled “Advanced Risk Adjusters and Predictive Formulas for ICD-10 Based Risk Adjustment.” The authors have no conflicts of interest to report.