Application of Machine Learning to Assess Interindividual Variability in Rapid-Acting Insulin Responses After Subcutaneous Injection in People With Type 1 Diabetes

Abstract

Objectives: Circulating insulin concentrations mediate vascular-inflammatory and prothrombotic factors. However, it is unknown whether interindividual differences in circulating insulin levels are associated with different inflammatory and prothrombotic profiles in type 1 diabetes (T1D). We applied an unsupervised machine-learning approach to determine whether interindividual differences in rapid-acting insulin levels associate with parameters of vascular health in patients with T1D.

Methods: We re-analyzed baseline pretreatment meal-tolerance test data from 2 randomized controlled trials in which 32 patients consumed a mixed-macronutrient meal and self-administered a single dose of rapid-acting insulin individualized by carbohydrate counting. Postprandial serum insulin, tumour necrosis factor (TNF)-alpha, plasma fibrinogen, human tissue factor (HTF) activity and plasminogen activator inhibitor-1 (PAI-1) were measured. Two-step clustering categorized individuals based on shared clinical characteristics. For analyses, insulin pharmacokinetic summary statistics were normalized, allowing standardized intraindividual comparisons.

Results: Despite standardization of insulin dose, individuals exhibited marked interpersonal variability in peak insulin concentrations (48.63%), time to peak (64.95%) and insulin incremental area under the curve (60.34%). Two clusters were computed: cluster 1 (n=14), representing increased serum insulin concentrations; and cluster 2 (n=18), representing reduced serum insulin concentrations (cluster 1: 389.50±177.10 pmol/L/IU h^-1; cluster 2: 164.29±41.91 pmol/L/IU h^-1; p<0.001). Cluster 2 was characterized by increased levels of fibrinogen, PAI-1, TNF-alpha and HTF activity; higher glycated hemoglobin; increased body mass index; lower estimated glucose disposal rate (increased insulin resistance); older age; and longer diabetes duration (p<0.05 for all analyses).

Conclusions: Reduced serum insulin concentrations are associated with insulin resistance and a prothrombotic milieu in individuals with T1D, and therefore may be a marker of adverse vascular outcome.

Keywords: apprentissage automatique; diabète de type 1; insulin pharmacokinetics; interpersonal variability; machine learning; pharmacocinétique de l’insuline; type 1 diabetes; variabilité interpersonnelle.