NIRCa: An artificial neural network-based insulin resistance calculator

Area of Science:

• Pediatric Endocrinology
• Computational Biology
• Diabetes Research

Background:

• Direct measurement of insulin sensitivity in children with type 1 diabetes is challenging.
• Accurate assessment of insulin resistance is crucial for managing type 1 diabetes in pediatric populations.

Purpose of the Study:

• To develop novel machine learning (ML) methods for estimating insulin resistance.
• To create accurate ML-based models for predicting glucose disposal rate (GDR) in children with type 1 diabetes.

Main Methods:

• Utilized hyperinsulinemic hyperglycemic clamp studies to determine GDR in 315 pediatric patients.
• Developed and compared artificial neural networks (ANN) and multivariate adaptive regression splines (MARSplines) models against a reference model.
• Validated model performance using training and independent testing sets.

Main Results:

• The ANN model demonstrated superior accuracy, achieving a median error of 0.6% and R² of 0.66.
• ANN predictions were within ±20% error for 75% of cases, outperforming MARSplines and the reference model.
• The reference model showed moderate performance (R² = 0.26), with poorer accuracy in younger children.

Conclusions:

• Developed an optimized GDR estimation model using ANN for research and clinical applications.
• The ANN-based tool offers a reliable method for assessing insulin sensitivity in pediatric type 1 diabetes.
• This ML approach simplifies and enhances the evaluation of insulin resistance in children.