Getting worse while waiting: A simulation study of urgency drift and continuous re-triage monitoring in high-volume OPD queues

Background High-volume outpatient departments (OPDs) use static first-come-first-served (FCFS) triage, assessing urgency once at registration with no re-assessment during waiting. Patients may undergo silent clinical deterioration during waits of 60–120 minutes, a phenomenon we term urgency drift—that goes undetected until physician contact. No published study has formally modeled urgency drift prevalence or the potential for AI-assisted detection in OPD queues. Objective To quantify urgency drift frequency and consequences in high-volume OPD queues and compare patient safety outcomes across 4 configurations: (1) static FCFS; (2) single-point AI triage at registration; (3) periodic human re-assessment every 30 minutes; and (4) continuous AI re-triage monitoring. Methods A discrete-event simulation (DES) model was built in Python 3.11/SimPy 4.1, calibrated to Indian district hospital OPD parameters with secondary validation against NHS metrics. Stochastic urgency drift used Weibull functions parameterized from clinical deterioration literature for 5 high-risk presenting categories, as no prospective OPD-specific drift dataset currently exists. Four configurations were run across 50 independent replications of 300 patients each (60,000 patient-episodes). A sensitivity sweep across AI sensitivity (0.60–0.95) and specificity (0.70–0.95) identified the minimum performance threshold. Results Static FCFS produced 27.8 urgency drift events per 100 patients (95% CI 24.8–30.8), all undetected until physician contact (mean time-to-escalation 29.4 minutes). Single-point AI triage (S2) did not reduce undetected drift (36.9/100, P>.05 vs S1). Periodic human re-assessment (S3) reduced undetected drift to 10.9/100 (60.8% reduction). Continuous AI re-triage (S4; sensitivity 0.85, specificity 0.88) reduced undetected drift to 6.0/100 (78.4% reduction), mean time-to-escalation to 8.5 minutes, and false-positive rate to 0.4/100. AI benefit over S3 was maintained at sensitivity ≥0.70. Conclusions Under the model’s assumptions, urgency drift would affect 27.8 per 100 OPD patients per session, entirely undetected under current practice. AI triage at registration provided no protection against drift occurring in the waiting interval, whereas any system that re-assesses patients during the waiting period — whether AI-based or human-delivered — substantially reduced undetected drift. Continuous AI re-triage reduced undetected drift by 78.4% relative to static triage, with negligible modelled false-positive burden, above a minimum sensitivity of 0.70. These findings should be interpreted as hypothesis-generating: they describe the behaviour of a simulation calibrated to published literature rather than measured OPD prevalence, and the reported AI performance values represent a preliminary benchmark for future empirical validation rather than a confirmed regulatory threshold. With this caveat, the study offers one of the first formal attempts to quantify OPD waiting-room deterioration and to propose a candidate performance floor for software as a medical device (SaMD) applicable to both LMIC and universal-coverage healthcare contexts.