Yeseul Jeon ¹ Jeong-Jae Kim ¹ SuMin Yu ² Junggu Choi ¹ Sanghoon Han ^1*

• ¹ Yonsei University, Seoul, Seoul, Republic of Korea
• ² Duke University, Durham, North Carolina, United States

Functional magnetic resonance imaging (fMRI) data is characterized by its complexity and high-dimensionality, encompassing signals from various regions of interest (ROIs) that exhibit intricate correlations. Analyzing fMRI data directly proves challenging due to its complex structure. Additionally, analyzing resting-state fMRI data poses extra challenges for direct analysis, since patterns may not be easily identifiable without specific tasks. Nonetheless, ROIs' interconnections provide essential information about brain activity and exhibit unique characteristics among different groups. To address this, we propose a cutting-edge interpretable fusion analytic framework that facilitates the identification and understanding of ROI connectivity disparities between two groups, thereby revealing their unique ROI features. Our novel approach encompasses three key steps. Firstly, we construct ROIs functional connectivity networks (FCNs) to effectively manage resting-state fMRI data. Secondly, by employing the FCNs, we utilize a Self-Attention Deep Learning Model (Self-Attn) for binary classification, generating an attention distribution that encodes group differences. Lastly, we employ a Latent Space Item-Response Model (LSIRM) to extract group representative ROI features, visualizing these features on the group summary FCNs. We validate the effectiveness of our framework by analyzing four types of cognitive impairments, showcasing its capability to identify significant ROIs contributing to the differences between the two disease groups. This innovative and interpretable fusion analytic framework has significant potential to enhance our understanding of cognitive impairments and could lead to more targeted therapeutic interventions.