ICU telemetry · alarm storm · throughput
Alarm Storm
Ten ICU patients alarm at once — a full alarm storm. Five triage teams race to clear the alarms, each one triaging a patient at its own measured latency. The fast paths clear the whole storm; the slow ones can barely keep up. This is a throughput race: in a storm, the model that can't triage fast is useless, no matter how accurate it is.
Five triage teams, one storm
each clears alarms at its measured latencyDecision stream
every triage: patient · confidence · latency · outcomeHow this race works
The storm is the workload
Ten patients alarm at once. Each triage team must clear them one at a time — a real ICU alarm storm is a throughput problem, not a single-decision problem.
Throughput = shift ÷ latency
A team clears a patient every measured latency. In a fixed shift, the number it clears is the shift length divided by its latency. Fast path → many; slow path → almost none.
Accuracy doesn't help here
In a storm, a model that takes a second to triage one patient can't keep up at all. The spread in throughput is the story — not the accuracy.
Shorten the shift to test
Drag the shift length. At 800 ms the fast paths clear the whole storm and the slowest clears almost nothing; at 2000 ms even the slow autoregressive path catches up a little.