Predicting the optimal basal insulin infusion pattern in children and adolescents on insulin pumps

Abstract

Objective: We aimed at developing and cross-validating a mathematical prediction model for an optimal basal insulin infusion pattern for children with type 1 diabetes on continuous subcutaneous insulin infusion therapy (CSII).

Research design and methods: We used the German/Austrian DPV-Wiss database for quality control and scientific surveys in pediatric diabetology and retrieved all CSII patients <20 years of age (November 2009). A total of 1,248 individuals from our previous study were excluded (dataset 1), resulting in 6,063 CSII patients (dataset 2) (mean age 10.6 ± 4.3 years). Only the most recent basal insulin infusion rates (BRs) were considered. BR patterns were identified and corresponding patients sorted by unsupervised clustering. Logistic regression analysis was applied to calculate the probabilities for each BR pattern. Equations were based on both independent datasets separately, and probabilities for BR patterns were cross-validated using typical test patients.

Results: Of the 6,063 children, 5,903 clustered in one of four major circadian BR patterns, confirming our previous study. The oldest age-group (mean age 12.8 years) was represented by 2,490 patients (42.18%) with a biphasic dawn-dusk pattern (BC). A broad single insulin maximum at 9-10 p.m. (F) was unveiled by 853 patients (14.45%) (mean age 6.3 years). Logistic regression analysis revealed that age, to a lesser extent duration of diabetes, and partly sex predicted BR patterns. Cross-validation revealed almost identical probabilities for BR patterns BC and F in the two datasets but some variation in the remaining two BR patterns.

Conclusions: Reconfirmation of four key BR patterns in two very large independent cohorts supports that these patterns are realistic approximations of the circadian distribution of insulin needs in children with type 1 diabetes. Prediction of an optimal pattern a priori can improve initiation and clinical follow-up of CSII in children and adolescents. In addition, these BR patterns represent valuable information for insulin-infusion algorithms in closed-loop CSII.

Figure 1

Left panel: Data heat map based on unsupervised hierarchical average linkage clustering of the most recent BRs of 6,063 pediatric patients with type 1 diabetes treated with CSII. The patients have been sorted by clustering along the y-axis, while the time course of the BRs is displayed from left to right over a period of 4 × 24 h on the x-axis for visualization of the differences of patterns and circadian rhythms. Increasing red intensity represents increasing insulin infusion rates, while increasing green intensity represents decreasing insulin infusion rates. Blackish colors reflect BRs near an individual’s mean BR. Clustering identifies the most similar BRs and sorts them right next to each other. The right margin of the heat map depicts the four leading BR patterns of the dataset, named F, AG, BC, and D. Right panel: mean BRs of all patients clustering in pattern F, AG, BC, or D and the variation from mean ± SD BR per pattern (y-axis) are displayed (mean BR = 1). (See also Supplementary Table 1.) The x-axis represents a 24-h interval from 0000 h to 2300 h.

Figure 2

Calculation of probabilities for typical patients of being treated with a BR pattern F (A), AG (B), BC (C), and D (D), respectively. The y-axis represents the probability for each of the four patterns in percent. (See also Supplementary Table 3A–D.) Age and duration of diabetes (in parenthesis) are given on the x-axis. ●, girls (dataset 1); ○, boys (dataset 1); ■, girls (dataset 2); □, boys (dataset 2).