Multi-Site Neuroimage Harmonization for Personalized Brain Disorder Analysis Abstract Predicting the future progression of preclinical Alzheimer's disease (AD) such as subjective cognitive decline (SCD) is essential for drug development and timely intervention to prevent further cognitive decline. Multi-site mul- timodal neuroimaging data, while increasingly employed to augment sample size and improve statistical power for investigating SCD and AD-related disorders (ADRD), are susceptible to inter-site and inter-modality data het- erogeneity caused by differences in scanners/protocols, studied populations, and imaging modalities. Mitigating inter-site data heterogeneity, principled fusion of multimodal data, and precise interpretation of neuroimaging data can reduce bias in subsequent analyses and help avoid erroneous conclusions. In this project, we will develop a set of computational tools, powered by advanced machine learning techniques, for multi-site data harmoniza- tion, multimodal data fusion, and personalized/subject-specific neuroimage interpretation for SCD progression prediction. These tools will be evaluated extensively on 5,300+ subjects with multimodal data (e.g., magnetic resonance imaging, positron emission tomography, and cerebrospinal fluid) involving 79 imaging centers. We propose three aims. In Aim 1, we will develop both feature-level and image-level deep learning frameworks for multi-site data harmonization. Many studies ignore inter-site data heterogeneity by simply assuming a common data source. Our methods will allow feature-level harmonization for precision medicine and image-level harmo- nization targeting a broader range of applications. The developed models will be easy to train via unsupervised learning. In Aim 2, we will develop a framework to effectively fuse multimodal data for subsequent analyses with- out discarding subjects who lack certain modalities. Existing studies usually require modality-complete subjects, limiting their utility in multi-site studies where many subjects may lack one or several modalities due to patient dropouts or failed scans. Our models can be trained with modality-missing subjects, and thus are practical with considerably better adaptability. In Aim 3, we will develop a framework for fast and accurate neuroimage search to facilitate personalized analysis of SCD and ADRD. Interpreting neuroimaging data at the subject level is of- ten challenging due to the ever-increasing amount of imaging information. Our method will help overcome this difficulty by scalable neuroimage search for subject-specific progression prediction of SCD and ADRD.