Validity of Nonspecific Stroke Location ICD-10 Subcodes: A Descriptive Study.

introductionWe aimed to describe imaging characteristics in stroke hospitalizations with nonspecific/unspecified vascular region subcodes and to assess for systematic bias in the use of these subcodes.

methodsWe captured first ischemic stroke hospitalizations from 2018 to 2022 at a single stroke center. We reviewed imaging studies to classify a gold standard of vascular region blinded to ICD-10 subcodes in 200 randomly selected hospitalizations: 100 with nonspecific/unspecified subcodes and 100 with specific subcodes oversampled for posterior circulation strokes. We assessed for systematic bias in the use of nonspecific/unspecified subcodes using multilevel logistic regression, with primary provider included as a random intercept. Separate models were applied to the full population of strokes and to those that underwent imaging review.

resultsWe identified 5,234 first ischemic stroke hospitalizations, of which 2,224 (43%) received a nonspecific/unspecified vascular region subcode. Out of the 100 ICD-10 nonspecific/unspecified stroke location cases that underwent imaging review, 85 had acute infarcts in specific locations: 45 anterior circulation, 40 posterior circulation, and 15 with no infarct. Factors associated with the use of nonspecific/unspecified subcodes were low NIHSS scores and non-neurological specialist but not anterior versus posterior vascular distribution. The proportion of variance explained by the models was modest (pseudo-R2 0.16).

conclusionsMost ischemic stroke hospitalizations coded with nonspecific/unspecified ICD-10 vascular region subcodes had imaging-confirmed infarcts in specific vascular regions. These strokes tended to have a lower NIHSS and were overrepresented by posterior circulation lesions. The modest variance explained in the use of nonspecific/unspecified codes indicates that much of the coding is influenced by random variation or unmeasured factors. Future studies in other healthcare systems are needed to verify these findings and evaluate for other predictors. Researchers using these subcodes should recognize the limitations and incorporate sensitivity analyses to evaluate potential bias in results.