Each vital sign is generated by an Ornstein-Uhlenbeck process relaxing toward a set point μ with physiological noise:
dX = θ(μ − X)·dt + σ·√dt·N(0,1)
An "Inject Deterioration Event" ramps μ for HR/RR/temperature up and SpO₂ down over ~40 simulated seconds via a smoothstep curve — modelling a sepsis-like decompensation reaching the wearable's sensors continuously, not just at ward-round check intervals.
The AI early-warning score aggregates NEWS2-style banded sub-scores (0–3 per vital, thresholds below) into a base score, then adds a trend term — the smoothed rate of change of that base score — which is the actual advantage of continuous AI monitoring over a periodic manual score:
R_base = Σ subscore(HR) + subscore(SpO₂) + subscore(RR) + subscore(Temp)
R_AI = clamp(100 · (R_base/12 + κ·dR_base/dt), 0, 100)
t_crit = (100 − R_AI) / max(slope, ε) [minutes to projected critical]
Because R_AI reacts to acceleration, it crosses an alert threshold earlier than a static NEWS2 cutoff would — the same principle real wearable early-warning platforms (used for sepsis and post-surgical deterioration) rely on. Sensor confidence models motion artefact / signal-quality dropout, which briefly widens the noise term σ and can itself suppress the score's reliability.
- Baseline sliders — set the patient's resting physiological set point.
- Inject Deterioration Event — starts a 40s (sim-time) decompensation ramp; watch the AI core in the 3D view shift from green to red before any single vital alone would trigger a hard alarm.
- Speed — compresses/stretches simulated time relative to real time.