Shift
AI triage
Presets
Playback
Model checks
The science
What is simulated
One clinician sees patients one at a time (a single-server queue). Patients arrive at random (Poisson process, rate λ per hour) and each consultation takes an exponentially distributed time with mean m minutes. The two wards below run the same patients with the same consultation times; only the order of service differs. FIFO serves the longest-waiting patient. AI priority serves the waiting patient with the highest AI urgency score; the score is the patient's hidden true urgency plus Gaussian noise N(0, σ2). A consultation is never interrupted (non-preemptive priority). A patient is urgent when true urgency > 0.75 (about one in five).
The mathematics
Utilisation is ρ = λ m. For FIFO (M/M/1) the mean wait is Wq = ρ m / (1 − ρ). For two non-preemptive priority classes (Cobham 1954) with class-1 load σ1 = ρ p and mean residual work W0 = λ E[S²]/2 = λ m², the class waits are W1 = W0/(1 − σ1) and W2 = W0/((1 − σ1)(1 − ρ)). Kleinrock's conservation law says that with equal service times the load-weighted mean wait is the same as FIFO: p W1 + (1 − p) W2 = ρ m/(1 − ρ). Prioritising does not make the queue shorter; it moves waiting from urgent patients to routine ones.
Independent numerical check
The first check simulates 1.2 million patients headlessly at ρ = 0.8 with a perfect urgent/routine classifier and compares the simulated mean waits with the formulas above (they agree to about 1 %). The other checks are structural: both policies finish the shift at the same minute (work conservation) and a noise-free AI orders urgent patients first (AUC = 1).
Honest simplifications
Real triage (ESI, Manchester) uses vital signs, several clinicians, resus bays, re-triage and patients who leave. Here there is one server, exponential consultations, a synthetic urgency score and a single noise parameter for "AI accuracy"; the 30-minute threshold is illustrative. The shift charts use one random shift (small samples are noisy, so the right-hand chart averages 30 shifts per noise value). This is a queueing lesson, not a clinical tool and not medical advice.