The effect of machine learning regression algorithms and sample size on individualized behavioral prediction with...

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Individualized prediction of behavior and cognition using machine learning (ML) regression is a growing field.
• The choice of ML regression algorithm and sample size critically affects prediction accuracy.
• Previous studies have not comprehensively evaluated these factors' impact on prediction performance.

Purpose of the Study:

• To assess the influence of six ML regression algorithms and varying sample sizes on individualized behavioral/cognitive prediction.
• To identify optimal ML algorithms and sample sizes for robust prediction using neuroimaging data.

Main Methods:

• Six ML regression algorithms (OLS, LASSO, Ridge, Elastic-Net, LSVR, RVR) were evaluated.
• Resting-state functional MRI (rs-fMRI) data from the Human Connectome Project (HCP) were used, with whole-brain resting-state functional connectivity (rsFC) and rsFC strength (rsFCS) as features.
• Twenty-five sample sizes, ranging from 20 to 700, were created through sub-sampling.

Main Results:

• LASSO regression with rsFC features and OLS regression with rsFCS features performed significantly worse than other algorithms.
• Prediction accuracy and stability increased exponentially with sample size, irrespective of the algorithm or feature type.
• Findings were robust across different datasets, preprocessing schemes, and cognitive scores.

Conclusions:

• The selection of ML regression algorithm and sample size critically influences individualized prediction performance.
• Increasing sample size consistently enhances prediction accuracy and stability.
• These results provide essential guidance for selecting appropriate ML methods and sample sizes in neuroimaging research.