Poster presentation from ADA2019 San Fransisco. This post show the poster and abstract from American Diabetes Association Conference 2019.
The 500-rule for the insulin to carbohydrate ratio (ICR) and 100-rule for insulin sensitivity factor (ISF) are easy to use but unfortunately do not provide good ratio estimates for many patients. Analysis of downloaded CGM and pump data can give good estimates but require a lot of experience from the user and is time consuming. We have developed an algorithm (Diassist), based on clinical experience and signal processing, which automatically estimates ICR and ISF, and investigated this method in a clinical context.
Pump and CGM downloads over 2 weeks were used for analysis. The physician interpreted these and compared with the Diassist results. Diassist calculated only valid meals and corrections, i.e. when blood glucose and carbohydrates had been entered by the patient.
8 patients with poor glucose control participated with 3-5 visits over 2-8 months.
Mean (±SD) age was 12.7 ± 1.8 (range 10.4-16.5) yrs, diabetes duration 8.0±1.1 yrs. Insulin dose was 0.78±0.17 (0.6-1.1) units/kg/24 hrs (% basal 40.8±13.0 (23-58)). The tool estimated ICR and ISF over 2 weeks and proposed to increase or decrease the settings based on actual glucose excursions. To validate the robustness of the method, cross validation with two data sets was used. Data was divided into 2 sets per patient; running Diassist on the two sets and comparing the results.
The physician followed the Diassist recommendations for ICR titration in 73% of the time slots (4-5/day, 58 out of 79), increasing or decreasing by 10-25%. In one occasion the ICR was changed in the opposite direction to the recommendation.
In 98 %, the ICR was lowered (more insulin for the meal), in 2% it was increased. HbA1c was not changed (61±4.4 vs. 57±6.3 mmol/mol, 7.7±0.4 vs. 7.4±0.6%).
The quota of carbohydrate and boluses used in analysis were 67% and 58% of the logged data.
Compared to methods used today, Diassist decreased both time and complexity of the estimation task. The proposed method shows promising results for usability, robustness and safety.