MIMIC-IV graph motifs aid timing more than ranking

Two MIMIC-IV papers use 99,527 motifs: graphs mainly improve long-range timing, while literature ranking barely shifts priority.

Two related arXiv papers dated 2026-10-14 describe work on the same frozen, patient-disjoint MIMIC-IV v3.1 catalog of 99,527 branching disease-trajectory motifs. One paper reports a leakage-controlled graph framework that learns train-only admission transitions and retrieves at most 24 endpoint-specific motifs with a time-aware Transformer. Across five seeds, micro-AUPRC rose from 0.9330 to 0.9357 and event-time MAE fell from 676.9 to 657.7 days, a 19.2-day gain concentrated on events beyond 730 days; corrected survival metrics did not outperform a matched history-only discrete-hazard model, and GNN edge scores did not improve forecasting. The companion paper ranks those motifs for literature novelty and clinical priority using BAAI/bge-small-en-v1.5 against 166,972 pre-2023 PubMed references. Positive recombination affected 8.63% of motifs but barely changed global ranking, while the high-semantic subgroup saw a 20.1-day MAE reduction and a 2.37-point sMAPE drop. The papers do not disagree on the motif count; the MAE figures cover different scopes, overall timing versus the high-semantic subgroup, and the authors do not claim causal discovery or clinician-adjudicated novelty.