The result
In about five weeks, we delivered a model of Novo Nordisk’s new pump-optimized insulin formulation. The insulin-pump company entered Novo Nordisk’s feasibility study inside the deadline.
First-hour model accuracy rose from about 52% to 86%, exceeding the client’s 72% acceptance bar. This metric measured how closely the model reproduced the insulin’s effect over the first hour, using GIR-AUC₀–₁ₕ against the clinical reference.
The bottleneck
To enter the study, the company needed a model that could show how its dosing algorithm would handle the new insulin. Its specialist advisor had retired, and the team’s first attempt fell short of the acceptance bar.
We checked the diagnosis before building. The missing insulin model was the constraint, and the client’s existing dosing algorithm was sound.
How we built and tested it
We agreed on the acceptance bar before the new-insulin data arrived. We built the fitting pipeline and evaluation harness, then tested the harness against Novolog, where the answer was already known.
When the new data arrived, we evaluated the model against both the pooled clinical profile and individual curves, including subgroup checks. Two engineers delivered the work in about five weeks, including an eight-day wait for partner data.
What we handed over
We handed the insulin model and evaluation evidence to the client’s algorithm team. The client could rerun the evaluation against its own acceptance bar.
Our deliverable was the insulin model. The client owned simulator integration and the dosing algorithm. Meeting its acceptance bar didn’t constitute regulatory approval of a device.
Free Expert Assessment
See what’s standing between you and your goal.
Tell us your goal. Your Expert Assessment identifies the likely bottleneck, the evidence behind it, and what to do next.
Get my Expert Assessment