Read the case study for the business story and results.
We supplied the insulin model. Their team owned the controller.
The deliverable was a pharmacokinetic and pharmacodynamic model, or PK/PD model. It describes exposure and effect over time. We didn't modify the dosing algorithm or integrate the model into the client's simulator.
The client's production model for Novolog supplied a reference and a positive control. Novo Nordisk's published clinical data for the new formulation supplied the evaluation reference. We didn't hold identifiable patient records.
The structure matched the data available.
We used a population model with subcutaneous absorption, disposition, and an effect compartment. A nonlinear mixed-effects approach represented both population behavior and variation between participants.
The structure was kept limited enough for the available data to support its parameters. Model fit and predictive checks considered the pooled profile, individual participant curves, and subgroups.
The first-hour metric had a specific definition.
First-hour model accuracy was one minus relative error in the glucose-infusion-rate area under the curve over the first hour, GIR-AUC (first hour), against the clamp reference. The result rose from ~52% in the client's initial model to ~86% in ours. Their acceptance threshold was 72%, fixed before the work.
Those figures correspond to first-hour prediction error falling from ~48% to ~14%. They aren't time-in-range measurements or clinical outcomes. Acceptance also relied on goodness-of-fit and predictive checks rather than a single percentage.
We checked the evaluation harness before the new result.
While waiting eight days for usable data, we built the fitting and evaluation pipeline and tested it against the existing Novolog reference. When the new data arrived, we applied the agreed checks to the new formulation.
The client reran the evaluation from the handover package. Its VP of Engineering and CTO/CPO accepted the work before forwarding it to Novo Nordisk, which also accepted it.
The handover enabled the next program stage.
Two engineers completed the delivery in ~5 weeks. The model was the final open item for study inclusion, and the company joined inside the deadline. The company retained integration, controller testing, and device responsibilities.
The $14M+ avoided-study figure is the client's leadership estimate of the alternative study. It isn't the cost of our build or cash returned from a study already underway. The engagement's scheduling lesson was to assign backup contacts and escalation paths for partner dependencies before kickoff.