Training a neural network for estimating the homeostatic model of insulin resistance

Abstract

Introduction: the evaluation of the homeostatic model is a test very close to the gold standard (euglycemic clamp).
Objective: train a multilayer perceptron-type neural network to determine the homeostatic model of insulin resistance.
Methodology: analytical and cross-sectional study. The learning of the neural network was carried out from a database of 2004 Venezuelan adults. Subsequently, 4,363 Mexican adults were added to the database of the National Health and Nutrition Survey (ENSANUT). The variables were homeostatic model of insulin resistance (HOMA2-IR), basal insulin, and basal glucose. Multilayer perceptron-type neural networks were used.
Results: the training of the neural network model had a relative error of 0.003, while in the test it was 0.005. For qualitative HOMA2-IR, the percentage of incorrect predictions was 0.60 % in training, and 0.70 % in testing. After learning the model, insulin and basal glucose values from 4363 Mexican adults were added, observing that the HOMA2-IR values generated by multilayer perceptron maintained the efficiency of the model, obtaining a coefficient of determination R2 of 0.983, which implies that 98 % of the variation in HOMA2-IR values can be explained by HOMA2-IR values obtained using multilayer perceptron.
Conclusions: the multilayer perceptron-type neural network gives results virtually identical to those obtained using the HOMA2-IR calculator. The implementation of this algorithm can be beneficial as a tool that is easy to implement in primary and specialized care systems and in hospital environments.