Systematic Literature Review of Machine Learning Methods in Insulin Secretion Model Analysis

Abstract

— Endogenous insulin secretion ( U N ) plays a critical role in maintaining glucose homeostasis. Pathological changes in U N enable early detection of metabolic inefficiency prior to the onset of diabetes mellitus (DM). Numerous researches have been carried out to establish the most effective method for assessing the participant’s glycemic state by identifying their U N profile. In contrast to insulin sensitivity ( SI ), there is no gold standard for U N profile. Thus, the deconvolution of C-peptide measurements is used in the majority of research to identify the U N profile. Due to the fact that C-peptide and insulin are co-secreted equimolarly from pancreatic β -cells, the latter method is shown to be accurate. Although studies have shown that the machine learning-based strategies can yield very positive outcomes in other areas of DM diagnosis, there is currently little research that employing machine learning for quantifying the U N profile to enable early diagnosis of metabolic dysfunction. Hence, the main objective of this study is to conduct a thorough search on machine learning-based modelling strategies that were used to identify the individual-specific U N profile through the development of a U N model. Additionally, this study will investigate whether the data acquired from the U N model can be used to quantify a person’s metabolic condition (either normal, pre-diabetic or T2D). The literature search turned up prospective studies linking machine learning and U N in its search and analysis. Meta-analyses summarize the available data and highlight various methodological stances. Thus, the exploratory of machine learning classification and regression technique can be portrayed in 3 different scenarios during the identification of U N profile. The 3 scenarios are: the study of insulin secretion through analyzing the insulin sensitivity, the study of U N without taking into considerations or in-depth study of U 1 and U 2 , and the study of insulin secretion using deconvolution of plasma C-peptide concentrations. It is evident that while Decision Tree (DT) is ideal for the first scenario, Random Forest (RF) is the better option for the other two scenarios. Further optimization can be implemented with the use of these techniques under supervised learning to improve diagnosis and comprehend the pathogenesis of diabetes, particularly in U N .