Overcoming Site Variability in Multisite fMRI Studies: an Autoencoder Framework for Enhanced Generalizability of...

Area of Science:

• Neuroimaging
• Machine Learning
• Data Harmonization

Background:

• Multisite functional magnetic resonance imaging (fMRI) data harmonization is essential for generalizable machine learning (ML) models.
• Traditional methods like ComBat may fail to capture complex, non-linear site variations and risk data leakage.
• This can limit the reliability and clinical applicability of ML models trained on harmonized data.

Purpose of the Study:

• To propose and evaluate Autoencoders (AEs) as a novel method for harmonizing multisite fMRI data.
• To leverage AEs' non-linear representation learning to reduce site-specific effects while preserving biological features.
• To address data leakage issues inherent in traditional statistical harmonization techniques.

Main Methods:

• Developed and implemented a framework utilizing various Autoencoder (AE) architectures (AE, SAE, TAE, DAE) for fMRI data harmonization.
• Evaluated the framework on the Autism Brain Imaging Data Exchange I (ABIDE-I) dataset (1,035 subjects, 17 centers).
• Employed leave-one-site-out (LOSO) cross-validation to assess performance against baseline methods.

Main Results:

• All AE variants demonstrated statistically significant improvements over the baseline (p < 0.01).
• Mean accuracy improvements ranged from 3.41% to 5.04% in LOSO cross-validation.
• AEs effectively reduced site-specific variability while preserving neurobiological features.

Conclusions:

• Autoencoders offer a powerful, non-linear approach to harmonize multisite fMRI data.
• This method enhances the robustness and reproducibility of downstream neuroimaging analyses.
• The proposed AE framework effectively mitigates data leakage, improving ML model reliability for clinical use.