Resample aggregating improves the generalizability of connectome predictive modeling.

Area of Science:

• Neuroimaging
• Machine Learning
• Cognitive Neuroscience

Background:

• Developing reliable brain-behavior relationship models is a key neuroimaging goal.
• Data-driven predictive models are popular but prone to overfitting, hindering generalization.
• Cross-validation estimates performance but optimal out-of-sample application remains unclear.

Purpose of the Study:

• To propose and evaluate an ensemble learning method (resample aggregating) for robust brain-behavior modeling.
• To investigate the use of aggregated models for estimating fluid intelligence (fIQ) from fMRI functional connectivity (FC).
• To assess model performance within-sample and out-of-sample across different datasets.

Main Methods:

• Utilized two large, open datasets: Human Connectome Project (HCP) and Philadelphia Neurodevelopmental Cohort (PNC).
• Employed the Connectome Prediction Modelling (CPM) framework to build aggregated and non-aggregated fIQ models using HCP data.
• Evaluated models on held-out HCP data and out-of-sample on PNC data across various test-train splits.

Main Results:

• Resample aggregated models demonstrated superior performance for fIQ estimation, both within-sample and out-of-sample.
• Significant variability in feature selection was observed within-sample.
• Out-of-sample performance was notably improved by the aggregated modeling approach.

Conclusions:

• Ensemble learning via resample aggregating offers a more generalizable approach to brain-behavior modeling.
• Robust feature selection is crucial for enhancing the cross-sample performance of CPM-based models.
• The proposed method advances reliable prediction of cognitive abilities from neuroimaging data.