Extreme gradient boosting using conventional parameters accurately predicts insulin sensitivity in young and...

Area of Science:

• Metabolic health research
• Biostatistics and machine learning applications

Background:

• Insulin sensitivity (SI) is crucial for metabolic health.
• Accurate estimation of SI is essential for clinical practice and research.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) models in estimating insulin sensitivity (SI).
• To compare ML-based SI estimations using physical indicators alone versus combined with lipid and glucose levels.

Main Methods:

• Employed eight machine learning (ML) methods to estimate SI in 1,268 young and 1,723 middle-aged Japanese individuals.
• Calculated SI using Matsuda index and 1/homeostasis model assessment of insulin resistance as references.
• Compared ML models using physical indicators only, and physical indicators with lipid/fasting glucose levels.

Main Results:

• Extreme gradient boosting models demonstrated the highest correlation with established SI indices.
• Feature importance varied significantly based on age and glucose tolerance status.
• ML-derived SI estimates outperformed traditional lipid-related estimates in correlation strength.

Conclusions:

• Machine learning enables SI estimation from physical indicators and clinical data.
• Developing universal SI estimates using conventional parameters is challenging.
• Further validation in diverse populations is required for robust SI estimation.