Network-level enrichment provides a framework for biological interpretation of machine learning results

Area of Science:

• Neuroscience
• Computational Biology
• Biostatistics

Background:

• Machine learning (ML) is widely used for identifying brain connectivity biomarkers.
• Current ML research often prioritizes prediction accuracy over biological interpretability.
• Inconsistent ML implementation can reduce model accuracy in neuroimaging studies.

Purpose of the Study:

• Introduce a network-level enrichment approach for connectome-wide analysis.
• Integrate brain system organization to link brain connectivity with behavior.
• Enhance biological interpretability of ML models in neuroimaging.

Main Methods:

• Utilized linear support vector regression (LSVR) models.
• Examined resting-state functional connectivity networks and chronological age.
• Compared network-level associations using raw LSVR weights against forward and inverse models.

Main Results:

• Failure to account for shared family variance inflated prediction performance.
• K-best feature selection via Pearson correlation decreased accuracy and reliability.
• Raw LSVR weights yielded network associations differing from forward/inverse model findings.

Conclusions:

• Network enrichment is valuable for biological interpretation in neuroimaging ML.
• Accounting for shared variance and appropriate feature selection are critical.
• The proposed approach offers crucial insights for applying ML to neuroimaging data.