Derivation and Validation of ICD-10 Codes for Identifying Incident Stroke

Importance Claims data with International Statistical Classification of Diseases, Tenth Revision (ICD-10) codes are routinely used in clinical research. However, the use of ICD-10 codes to define incident stroke has not been validated against expert-adjudicated outcomes in the US population. Objective To develop and validate the accuracy of an ICD-10 code list to detect incident stroke events using Medicare inpatient fee-for-service claims data. Design, Setting, and Participants This cohort study used data from 2 prospective population-based cohort studies, the Atherosclerosis Risk in Communities (ARIC) study and the Reasons for Geographic and Racial Differences in Stroke (REGARDS) study, and included participants aged 65 years or older without prior stroke who had linked Medicare claims data. Stroke events in the ARIC and REGARDS studies were identified via active surveillance and adjudicated by expert review. Medicare-linked ARIC data (2016-2018) were used to develop a list of ICD-10 codes for incident stroke detection. The list was validated using Medicare-linked REGARDS data (2016-2019). Data were analyzed from September 1, 2022, through September 30, 2023. Exposures Stroke events detected in Medicare claims vs expert-adjudicated stroke events in the ARIC and REGARDS studies. Main Outcomes and Measures The main outcomes were sensitivity and specificity of incident stroke detection using ICD-10 codes. Results In the ARIC study, there were 110 adjudicated incident stroke events among 5194 participants (mean [SD] age, 80.1 [5.3] years) over a median follow-up of 3.0 (range, 0.003-3.0) years. Most ARIC participants were women (3160 [60.8%]); 993 (19.1%) were Black and 4180 (80.5%) were White. Using the primary diagnosis code on a Medicare billing claim, the ICD-10 code list had a sensitivity of 81.8% (95% CI, 73.3%-88.5%) and a specificity of 99.1% (95% CI, 98.8%-99.3%) to detect incident stroke. Using any diagnosis code on a Medicare billing claim, the sensitivity was 94.5% (95% CI, 88.5%-98.0%) and the specificity was 98.4% (95% CI, 98.0%-98.8%). In the REGARDS study, there were 140 adjudicated incident strokes among 6359 participants (mean [SD] age, 75.8 [7.0] years) over a median follow-up of 4.0 (range, 0-4.0) years. More than half of the REGARDS participants were women (3351 [52.7%]); 1774 (27.9%) were Black and 4585 (72.1%) were White. For the primary diagnosis code, the ICD-10 code list had a sensitivity of 70.7% (95% CI, 63.2%-78.3%) and a specificity of 99.1% (95% CI, 98.9%-99.4%). For any diagnosis code, the ICD-10 code list had a sensitivity of 77.9% (95% CI, 71.0%-84.7%) and a specificity of 98.9% (95% CI, 98.6%-99.2%). Conclusions and Relevance These findings suggest that ICD-10 codes could be used to identify incident stroke events in Medicare claims with moderate sensitivity and high specificity.