• explainable AI
• fMRI
• harmonization
• machine learning
• multi-center
• Pearson
• SHAP
• wavelet

This thesis concerns the study of brain functional connectivity using machine learning (ML) models aimed to the investigation of Autism Spectrum Disorders (ASD) using MRI data.
Data are collected in different medical centers, thus they suffer the so-called batch effect.
This effects should be removed through harmonization procedures to avoid biases during the learning process of ML algorithms.
Two harmonization procedures are implemented and compared, one analytical and the other based on a deep neural network.
From fMRI data, functional connectivity is obtained by using a Pearson-based correlation measure and a time-frequency analysis by means of wavelet transforms of brain areas time series.
The two approaches have been compared to assess which one allow a deep neural network to better differentiate between control subjects and ASD patients.
Explainable Artificial Intelligence method, namely the SHAP approach have been employed to decode the outputs of machine learning models.
With this analysis it was possible to identify altered functional connections and determine what brain areas are mainly involved in ASD.