Healthcare policy · Modelling · 3 weeks

We ran 540 versions of the model. The best one was 0.42% better. We said stop.

The job was predicting how long patients stay in hospital. The honest finding was that the data had nothing more to give.

+0.42%

Total improvement from 540 model runs — the point at which more work stops being worth paying for

A first pass got close to the achievable accuracy almost immediately. We then ran 540 variations looking for more. The best of them was 0.42% better than where we started.

Looking at what the model was actually using explained why. Most of its accuracy came from simply knowing which country a figure belonged to — it was recognising national baselines, not learning anything transferable. The only factor with real explanatory weight was how much imaging equipment a system had per person.

So the deliverable was the recommendation to stop. Another month of tuning would have produced a slightly better number and no better decision, and we would rather say that than bill for it.

A health-system capacity study, 32-country public panel

What it came down to

The ceiling was the finding, and finding it early is worth more than a better score
Imaging capacity per person was the one factor that genuinely mattered
We recommended stopping instead of selling more work

Model versions tested: 540
Model versions tested
Improvement over the untuned baseline: +0.42%
Improvement over the untuned baseline
Countries in the panel: 32
Countries in the panel

What this doesn’t tell you

Country-level public data cannot say anything about an individual hospital. It answers a policy question, not an operational one.

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.