Estimating Injury Severity From ICD-10 Codes: Agreement, Misclassification and Analytic Implications

The Injury Severity Score (ISS) measures trauma severity using Abbreviated Injury Scale (AIS) scores, but ISS calculation is resource-intensive, typically requiring manual abstraction. Because International Classification of Diseases (ICD) codes are ubiquitous, ICD-to-AIS methods could automate AIS/ISS estimation. This study compares the accuracy of ICD-based tools (Association for the Advancement of Automotive Medicine, [AAAM] ICD-to-AIS crosswalk and International Classification of Diseases Programs for Injury Categorization in R [ICDPIC-R]) in estimating AIS and ISS.

This retrospective cohort study included adult (18-89 y) trauma patients from 2018 to 2023 Trauma Quality Program (TQP) Participant Use Files (PUFs). AIS body region severity scores were classified as originally defined (AIS_br8) or used for ISS calculation (AIS_br6). Injury severity was derived from three sources: TQP PUF dataset, AAAM ICD-to-AIS map, and ICDPIC-R package. PUF gold standard agreement was assessed using Cohen's weighted kappa and intraclass correlation.

In 5,793,422 patients, the AAAM crosswalk showed variable AIS_br8-PUF agreement (Cohen's weighted kappa: 0.73-0.95), peaking in lower extremities and lowest agreement in the spine. ICDPIC-R lacks AIS_br8 scores, which limits direct comparison. For AIS_br6-PUF, AAAM demonstrated poor to excellent agreement for low-severity scores (1-3), but poor to fair agreement for high-severity scores (4-6). ICDPIC-R showed poor to fair agreement across most body regions (Cohen's weighted kappa: 0.04-0.58). Both tools misestimated cohort sizes by as much as -41% underestimation and +22% overestimation.

ICD-based tools showed variable accuracy for estimating AIS, particularly for high-severity injuries and cohort identification. These important limitations should be considered when using derived AIS/ISS for risk adjustment and cohort selection.