Sepsis at minute twelve: early warning is a data-cadence problem

The frustrating truth about sepsis is that the data usually knew first. A small, persistent rise in heart rate. A slow drift in mean arterial pressure. A lactate trend, a white-count drift. None of these fires an alarm on its own — each sits comfortably inside its threshold — and so the bundle starts at the morning round, after the trend has run its course.

A ventilated patient produces more than 1,400 chartable data points a day; a twenty-bed unit, over 28,000 rows. Traditionally, less than five percent of that is meaningfully read before the next clinical decision. Early warning is therefore not a knowledge problem — every intensivist knows the signature — it is a cadence problem. No human can re-read the whole unit every fifteen minutes. Software can.

Composing the quiet signals

The approach that works is composition: read the live stream continuously, compute the trends that individual thresholds ignore, and combine them into a risk band that updates through the critical first day of admission. Done this way, sepsis surfaces four to six hours ahead of clinical declaration — not as a certainty, but as a candidate signal a clinician evaluates while there is still time to act cheaply.

Two design rules keep this safe and trusted:

Advisory only, always. The output is a recommendation with its contributing factors visible. The clinician confirms or overrides; the override is logged. The model’s job is to make the team look earlier — never to decide.

In the chart, not beside it. A risk band that lives in a separate dashboard dies of neglect by week three. Surfaced as a native chart entry — beside the vitals it was computed from — it meets the clinician inside the workflow they already run.

From foresight to bundle

Foresight only matters if the response is fast. The same event that raises the risk band should prime the pathway: lactate, cultures, antibiotics, fluids — checklist open, clock running. In units that work this way, door-to-bundle is measured in minutes, and the audit trail shows it.

The unit that catches sepsis early is not the one with the smartest people awake at 03:00. It is the one whose system reads every row at 03:00 — and knows what a bending trend looks like.