xMICD Creates Explainable ICD Code Representations for EHRs

Effectively representing International Classification of Diseases (ICD) codes from Electronic Health Records (EHRs) is a challenge in clinical risk prediction. Existing methods often force a trade-off: grouping-based representations are interpretable but can lose information, while embedding-based representations offer high predictive power but lack transparency. xMICD (Explainable Representation of Multiple ICD Codes) bridges this gap. It constructs patient representations by integrating clinically meaningful diagnostic groupings with semantic similarity derived from pre-trained ICD embeddings. Instead of simple binary group assignments, xMICD uses similarity-based relative assignments, creating features that show how closely a patient's diagnoses align with various clinical groups. This approach yields predictive performance on par with advanced embedding methods like ICD2Vec, while ensuring that each feature dimension corresponds to an understandable clinical concept, thus enhancing interpretability.