Addition of Genetics to Quantitative MRI Facilitates Earlier Prediction of Dementia: A Non-invasive Alternative to Amyloid Measures

Background: Alzheimer9s disease is a major health problem, affecting ~4.5% of people aged 60 and older in 2016 with over 43 million affected globally. The traditional approach for detection evaluates an individual in the presence of symptoms. However, it has been established that amyloid deposits begin to accumulate years before symptoms begin to appear. With improved technology, there is increased focus on risk reduction, timely diagnosis, and early intervention. Early identification of at-risk individuals may enable patients and their families to better prepare for and reduce the impact of this condition. Methods: We obtained data for patients from two longitudinal retrospective cohorts (Alzheimer9s Disease Neuroimaging Initiative: ADNI and National Alzheimer9s Coordinating Center: NACC), including T1-weighted MRI and genetics data. The polygenic risk score (PRS) used in this study was built based on a published Genome Wide Association Study (GWAS) that identified variants associated with Alzheimer9s disease. Quantitative MRI features were obtained using a 3D U-Net neural network for brain segmentation. Cox proportional hazards (CPH) regression models were used with subjects censored at death or the last evaluation. Time-to-event was defined as the time it takes for an individual who is dementia-free at the baseline MRI to progress to dementia as defined by the criteria described by ADNI. Time-dependent ROC areas under curve (AUCs) were estimated in the presence of censored data. The time-dependent AUCs were compared among models using the Wilcoxon rank sum test for dependent samples. Data was binned into three groups according to survival probability to eight years after baseline and Kaplan-Meier survival analysis was used to estimate the probability of surviving at least to time t. Calibration for both training and validation cohorts was evaluated using the predicted survival probability, splitting samples into five risk groups of equal size based on the predicted survival probability. Findings: We developed a model that predicts the onset of dementia over an eight-year time window in individuals with genetics data and a T1-weighted MRI who were dementia-free at baseline. We then validated the model in an independent multisite cohort. We observed that models using PRS in addition to MRI-derived features performed significantly better as measured by time-varying AUC up to eight years in both the training (p = 0.0071) and validation (p = 0.050) cohorts. We observed improved performance of the two modalities versus MRI alone when compared with more invasive amyloid measures. The combined MRI and PRS model showed equivalent performance to cerebral spinal fluid (CSF) amyloid measurement up to eight years prior to disease onset (p = 0.181) and while the MRI only model performed worse (p = 0.040). Finally, we compared to amyloid positron emission tomography (PET) three to four years prior to disease onset with favorable results. Interpretation: Our finding suggests that the two modalities are complementary measures, in that MRI reflects near-term decline and the addition of genetics extends the prediction scope of quantitative MRI by adding additional long-term predictive power. The proposed multimodal model shows potential as an alternate solution for early risk assessment given the concordance with CSF amyloid and amyloid PET. Future work will include further comparison with amyloid PET (greater than four years) and with CSF (greater than eight years) as additional long-term data becomes available. Also, the model will be evaluated for its clinical utility in the "active surveillance" of individuals who may be concerned about their risk of developing dementia but are not yet eligible for assessment by amyloid PET or CSF. Funding: Human Longevity, Inc Alzheimer9s Disease Neuroimaging Initiative National Alzheimer9s Coordinating Center