Impact of diagnosis code grouping method on clinical prediction model performance: a multi-site retrospective observational study

Abstract Objective The primary purpose of this work is to systematically assess the performance trade-offs on clinical prediction tasks of four diagnosis code groupings: AHRQ-Elixhauser, Single-level CCS, truncated ICD-9-CM codes, and raw ICD-9-CM codes. Materials and Methods We used two distinct datasets from different geographic regions and patient populations and train models for three prediction tasks: 1-year mortality following an ICU stay, 30-day mortality following surgery, and 30-day complication following surgery. We run multiple commonly-used binary classification models including penalized logistic regression, random forest, and gradient boosted trees. Model performance is evaluated using the Area Under the Receiver Operating Characteristic (AUROC) and the Area Under the Precision-Recall Curve (AUCPR). Results Single-level CCS, truncated codes, and raw codes significantly outperformed AHRQ-Elixhauser ICD grouping when predicting 30-day postoperative complication and one-year mortality after ICU admission. The performance across groupings was more similar in the 30-day postoperative mortality prediction task. Discussion Single-level CCS groupings represent aggregations of raw codes into meaningful clinical concepts and consistently balance interoperability between ICD-9-CM and ICD-10-CM while maintaining strong model performance as measured by AUROC and AUCPR. Key limitations include experimentation across two datasets and three prediction tasks, which although were well labeled and sufficiently prevalent, do not encompass all modeling tasks and outcomes. Conclusion Single-level CCS groupings may serve as a good baseline for future models that incorporate diagnosis codes as features in clinical prediction tasks. Code and a compute environment summary are provided along with the analyses to enable reproducibility and to support future research.