UCSF

Alzheimer’s Disease (AD) represents the most frequent (60-80%) subtype of dementia and is one kind of progressive spectrum disorder without effective treatment so far. In the last decades, great efforts from all over the community have been made on the early diagnosis of AD at its preclinical stage, Mild Cognitive Impairment (MCI). Recently, a series of machine learning studies have successfully constructed several computational models in predicting conversion of cognitive impairment but seldom foresee beyond 4 years. Thanks to Alzheimer’s Disease Neuroimaging Initiative (ADNI) database, in this study we extracted cognition feature from several clinical outcomes. We then took advantage of structural MRI data and one Network Diffusion Model (NDM) raised by our group for subject-specific prediction of future cognition features. One supervised classification neural network was trained with ground-truth baseline and time-of- interest data but applied with predicted future cognition features. This established machine learning framework has demonstrated descent sensitivity and specificity in prediction of MCI-to- AD conversion (0.890 ± 0.083 and 0.923 ± 0.045) and healthy control (HC)-to-AD conversion (0.900 ± 0.074 and 0.744 ± 0.154) 5 years post baseline. To the best of our knowledge, we are the very first groups working on long-term prediction of AD spectrum conversion from both HC and MCI.