Healthcare · Clinical machine learning · 10 weeks
A sepsis warning built from eight numbers every bedside monitor already reports.
Sepsis kills faster the longer it goes unnoticed, and the standard bedside scores catch it late. The question was whether the machines already in the room hold enough signal.
8 inputs
Heart rate, breathing, blood pressure, oxygen, age and sex — no lab work, so it can run on every bed
We started with 1.5 million hourly readings across 43 measures. Twenty-seven of those were more than 90% empty, because most lab panels are drawn rarely — so rather than invent the missing values we dropped them.
What survived was eight numbers: heart rate, breathing rate, three blood-pressure measures, oxygen level, age and sex. That constraint turned into the point of the whole thing. A warning that needs fresh lab results can only run on some beds. One built from continuous vitals can run on all of them.
The model ranked heart rate and breathing rate as the strongest signals, which matches what clinicians already look for. That agreement matters more than the score: a warning that disagrees with the bedside for reasons nobody can explain does not get used.
A critical-care analytics programme, ~40,000-patient de-identified ICU cohort
What it came down to
- Eight continuously-measured inputs, so it works on every bed rather than a few
- Six model types compared before choosing one
- What it flags lines up with what experienced clinicians watch
- Hourly readings processed
1,552,210
Hourly readings processed
- Clinical measures assessed
43
Clinical measures assessed
- Model types compared
6
Model types compared
What this doesn’t tell you
At this setting it catches roughly three in five cases. That is not good enough to act as an alarm on its own, and we said so rather than rounding it up — it is useful for deciding who to check first, not for deciding nobody needs checking.
Built on one public dataset, with no second hospital to test against. Closing the gap needs lab values and trends over time, both of which we scoped rather than claimed.
Clients are described by sector, scale and region only. No names, products or identifiers appear here.
Next step
Have a version of this problem?
Tell us what you are trying to build. You will get an honest read on whether it is worth building at all.
What happens next
- You pick the workflow. We ask what already exists, who touches it, and where it stalls.
- We reply within one business day with a rough shape and a rough timeline — usually smaller than people expect.
- If it makes sense, the next step is a working system in two to six weeks, tested by your own staff.
- You keep your existing tools. Private deployment is available where data cannot leave.
- And if it is not worth building, we say so instead of selling you one.