CARing combines compositional medical semantic IDs with a coverage reward to reach the highest recall in next-visit diagnosis prediction on MIMIC-III/IV (R@30 46.04%/46.52%)
The work proposes CARing, which compresses disease names and ICD ontology paths into compositional medical Semantic IDs (SIDs) via residual quantization, grounds them through six alignment tasks plus teacher-generated semantic-enrichment corpora and reasoning activation so a single autoregressive model reasons before emitting a diagnosis SID, and optimizes multi-label coverage with a coverage reward and multi-positive supervision; on MIMIC-III and MIMIC-IV it exceeds all EHR-trained baselines in weighted score and attains the highest recall at every reported cutoff, with R@30 of 46.04% and 46.52% in reasoning mode.