Systematic evaluation of machine learning algorithms for neuroanatomically-based age prediction in youth

Area of Science:

• Neuroscience
• Computer Science
• Developmental Psychology

Background:

• Machine learning (ML) models applied to structural magnetic resonance imaging (sMRI) data can estimate brain age, reflecting biological brain maturity.
• Accurate brain-age estimation is crucial in youth due to dynamic age-related brain changes.
• Systematic comparisons of ML algorithm performance for youth brain-age estimation are lacking.

Purpose of the Study:

• To evaluate and compare the accuracy and computational efficiency of 21 ML algorithms for estimating brain age in youth using sMRI data.
• To identify the most effective ML approaches for quantifying brain age in developing individuals.

Main Methods:

• Utilized sMRI data from 2105 typically developing individuals (aged 5-22 years) across five cohorts.
• Trained and tested 21 ML algorithms, including parametric, nonparametric, Bayesian, linear, nonlinear, tree-based, and kernel-based models.
• Performed sensitivity analyses on factors like parcellation, feature selection, cross-validation, outlier handling, and sample size.

Main Results:

• Tree-based models and nonlinear kernel algorithms demonstrated comparable accuracy and computational efficiency.
• Extreme Gradient Boosting (MAE: 1.49 years), Random Forest Regression (MAE: 1.58 years), and Support Vector Regression with RBF Kernel (MAE: 1.64 years) were the top-performing models.
• Linear algorithms generally performed poorly, with Elastic Net Regression as a notable exception.

Conclusions:

• The study provides a guide for selecting optimal ML methodologies for brain-age estimation in youth.
• Nonlinear and tree-based ML models are recommended for accurate and efficient brain-age quantification in this demographic.
• Findings can inform future research on neurodevelopmental trajectories and brain health in children and adolescents.