Quantifying performance of machine learning methods for neuroimaging data.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Machine learning (ML) is increasingly used for neuroimaging data analysis.
• Standard ML algorithms struggle with neuroimaging data's high dimensionality and low signal-to-noise ratio.
• The comparative performance of ML regression algorithms on neuroimaging data remains unclear.

Purpose of the Study:

• To quantify the performance of various ML regression algorithms on neuroimaging data.
• To investigate the impact of sample size, feature set size, and predictor effect size on algorithm performance.
• To evaluate the contribution of embedded feature selection and bootstrap aggregation (bagging) to model performance.

Main Methods:

• Examined five ML regression methods: Gaussian Process Regression, Multiple Kernel Learning, Kernel Ridge Regression, Elastic Net, and Random Forest.
• Utilized both real and simulated MRI data.
• Compared ML algorithms against standard multiple regression, incorporating feature selection and bagging.

Main Results:

• Algorithm performance varied significantly based on sample size, feature set size, and effect size.
• Elastic Net, Kernel Ridge Regression, and Gaussian Process Regression performed well with large effect sizes.
• Elastic Net showed accurate predictions for small effect sizes, but only with sample sizes > 400. Random Forest offered moderate performance across all sample sizes for small effect sizes.
• ML techniques enhanced prediction accuracy in multiple regression.

Conclusions:

• Empirical evidence demonstrates differential performance of ML algorithms on neuroimaging data.
• Algorithm selection should consider sample size, feature count, and effect size.
• ML techniques, particularly Elastic Net, offer promising predictive capabilities for neuroimaging analysis.