acceptodds
Desk rejected at ICLR 2027

Recurrent Clinical Safety Queues Reverse Multimodal Model Selection

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

At every decision time, a recurrent clinical safety queue spends capacity on new cases and material updates but not on unresolved carryovers; deferral, dismissal, cooldown, and re-entry then determine the next admissible set. Completed-record sorting and next-admissible selection consequently need not commute. On 9,714 patient-episodes and 1,074 adjudicated events from three held-out trials, completed-record discrimination selects gated fusion (AUROC/AUPRC/PPV@top-5%: 0.90/0.47/65.5), while recurrent review selects ChronoGuard. Under the five-selection per-instant boundary, ChronoGuard recognizes events 1.1 days earlier than gated fusion (95% CI, 0.6–1.6), raises severe-event sensitivity by 0.06 (0.02–0.10) and selected-alert precision by 0.07 (0.03–0.11), and removes 0.31 false alerts per monitor-week (0.17–0.45). AdmitLock makes this reversal testable by locking the evidence filtration, scorer-specific admission, recurrent transition, capacity clock, and endpoint registry, then reporting both Pareto orders, selected-episode overlap, and paired queue effects. Only 12–16 of each model's 25 highest-ranked completed-record episodes occur among its first 25 recurrent selections. Against operational key risk indicators, ChronoGuard improves recognition by 3.5 days, severe sensitivity by 0.14, precision by 0.19, and false-alert burden by 0.95 per monitor-week; day-28 restricted mean recognition time falls from 11.1 to 7.4 days. An independently governed four-trial portfolio reproduces both point-estimate leaders after refitting every learned method, and a 228-monitor-week shadow evaluation preserves ChronoGuard's operating profile under live arrivals. Model selection for scarce clinical review should therefore target the stateful sequence that reaches experts, not a ranking formed after records close.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
biomedical and clinical ai, Recurrent Clinical Safety Queues Reverse Multimodal Model Selection

Related papers

Discussion (0)

Sign in to comment.