Prediction of Brain Biological Age and Acceleration of Aging in AD Using Large Datasets of Existing Neuroimaging Data and Deep Learning Approaches
Co-Principal InvestigatorsArthur Toga, PhD
This proposed research seeks to first predict physiological brain age (PBA) for individuals in healthy condition by leverage deep learning-based modeling with large size brain MRI datasets. The preliminary data showed that the current model is very accurate in this task with the mean absolute error (MAE) of 3 years. (i.e., the predicted age and the chronological age of a healthy subject are not different by more than 3 years). The model trained on healthy population will be applied to the cohorts of mild cognitive impairment (MCI) and AD in order to access the acceleration of aging, meaning that the model would show increased age in these cohorts by capturing the faster brain appearance changes on MRI due to MCI or AD. Our proposed deep-learning approach will be able to predict the voxel-wise PBA while mapping it on regional brain anatomy, which will allow the region specific aging pattern in MCI or AD (e.g., increased age of the hippocampus relative to other brain structures) and subsequently classify each patient from healthy or other disease groups (MCI / AD or other dementia group) using the disease specific pattern of PBA map. Finally, we will apply an event-based model (EBM) to the PBA map generated by the proposed deep learning algorithm. This will generate the sequence in which regional PBA(s) become abnormal regarding MCI or AD. In other words, starting from a specific region (e.g., hippocampus) to be abnormal at the earliest stage, one can observe how this brain abnormality propagates to other regions (e.g., prefrontal cortex) incrementally over different stages. The patients belonging to one stage are expected to probabilistically differ in clinical characteristics from those at other stages. New patients will be assigned to one of the stages in EBM based on the similarity of the PBA pattern and this process combined with a mathematical Hazard model will estimate the ‘survival’ probability explaining the remaining days in healthy status prior to the onset of MCI or AD, ultimately allowing disease-specific risk scoring as a clinical tool to be used in routine patient care.
The goal of this project is to predict individual physiological brain age (PBA), taking into account the complex nature of the human aging trajectory, we propose to leverage deep learning-based PBA modeling based upon multicontrast MRI databases of healthy and age-related disease populations.
There are two specific aims:
1. Develop multimodal MRI-based deep-learning methods to predict regional brain age in healthy subjects. To better understand the meaning of high PBA offsets, we will assess the multivariate association of the offset with putative risk variables including cognitive decline, health risk factors (biometrics), genetic influence, predisposed lifestyle and health factors (alcohol use, hypertension, diabetes, etc.).
2. Classification, staging and risk-scoring of dementias based on PBA patterns. Brain regional PBA pattern at a given age combined with putative aging variables is hypothesized to be unique for each of various age-related brain diseases and can be used as a “fingerprint”. Classification based on the voxel-wise PBA pattern together with putative aging variables will diagnose an individual into healthy, MCI, or AD.