This web page is a quick look at the Glucose Transform method for estimating insulin action from measured glucose data. The method was first mentioned in the misc.health.diabetes news group as a way to compare Humalog with Novolog, the two competing meal-time synthetic analogue insulins presently on the market. The Glucose Transform is the title of my thesis that is currently in progress. So this method has not yet been published in journals to be peer reviewed.
This example is for 24U of Novolog and was mentioned in misc.health.diabetes to raise awareness of antibody binding delay effects. The first graph shows measured fingerstick BG data as dots before and after the 24U of Novolog was administered abdominally, to a subcutaneous depth of 12 mm. This was the first dose of Novolog for my body, so I should not have any anti-Novolog antibodies at this time. This BG data is then least squares curve fit to a polynomial, the solid line in this first graph. The Bergman Minimal Model differential equation is then used to transform the BG profile into the estimated insulin action profile of the second graph. The following two graphs are placed back to back for easy comparison. The time scales are slightly different and I'll correct this later with a new gif.
Transformed BG data to estimate insulin action.
Initial 70% peak at 1 hour, human insulin antibody binding delays Novolog action until 4 hours, high affinity antibodies release for 28% peak at 6.8 hours. BG data of the first graph is grossly under-fit to smooth out meter jitter, i.e., average out meter noise. This data set was not corrected for urine glucose losses, that are usually too small to bother with. If urine glucose losses were accounted for, the first 70% peak would be smaller and shift out later in time. This would make Novolog worse, as a meal-time insulin, specific to my body. Insulin action ends at about 7.4 hours, 2.4 hours longer then expected from Novo Nordisk published data, (3-5 hours above figure 3 of Novolog product leaflet), for normal subjects with no antibody binding effects. I am a nonsmoker, (smoking slows absorption by vasoconstriction), and the injection site was not chilled to slow absorption.
This is specific to my metabolism, so your mileage may vary. I'm still collecting data sets for both Humalog, (that I've used for 6 years), and Novolog. So it's too early to decide which is better for me. I've never used beef or pork insulins, so the antibodies could only be from synthetic human insulin. I also have slightly faster Novolog data, in fairness to Novo Nordisk. To date, my Humalog data sets have maximum action peaks from 0.5 hour to 1.75 hours, for dose sizes ranging from 12U to 29U. This is consistent with published data from Eli Lilly for lispro activity and demonstrates a smaller antibody binding influence for this preparation. My current insulin therapy uses Humulin R, NPH, ultralente and Humalog daily. My rates of hypoglycemia are less than 1 per month.
It is also important to note that the Minimal Model only has single glucose and insulin compartments for the whole human body. Therefore, the euglycemic glucose-clamp dextrose infusion data of the Biostator, the current method used to measure insulin action, will be slightly different. But the insulin action peak times should closely agree. The method is also extremely dependent on meter strip reaction chemistry. If the strip reaction chemistry does not have enough range to measure hyperglycemic values accurately, the insulin action profile will have gross meter biases when filtered by the differential equation of the Minimal Model. So the BG meter is important to the method. However, if the same meter methodology is used, when estimating BG from these transformed insulin action profiles using the Minimal Model, then the predicted BG should agree closely with measured BG, (unless the strip lot is bad or very different, suggesting the FDA may need to tighten BG meter tolerance requirements to keep up with technology advancements).
The use of these insulin action profiles will be in some future insulin dosing computer model for personal use, after investigating linearity issues of dose summation, by using this Glucose Transform method. Antibody binding will cause summation of insulin effects to be nonlinear, implying linear models without antibody effects, like the AIDA, will be incorrect for personal use. It is therefore important to understand these nonlinear antibody binding effects, specific to an individual patient, prior to constructing an insulin dosing feedback control computer program that will be calibrated to this patient's metabolism. Two interesting parameters that are determined noninvasively, (other than fingersticks), by the method are the CNS glucose uptake rate and initial hepatic glucose flow, as shown in the second graph.
Final notes: The Bergman Minimal Model differential equation is a special linear case of the Riccati family of nonlinear differential equations. As of this moment, I've decided to pull the thesis and keep the algorithm proprietary. I will eventually publish a paper prior to submitting an SBIR proposal to the NIH. But at this time, linux is my bread, butter and hobby. So when the spirit moves or venture capital/funding arrives, the glucose transform method will advance. However, the method was instrumental in determining Novolog (NovoRapid) was too slow, as a prandial insulin, for my metabolism. So I kept the older, but faster Humalog, for my mealtime insulin requirements.