Bayesian workflow for the investigation of hierarchical classification models from tau-PET and structural MRI data across the Alzheimer’s disease spectrum

Background Alzheimer’s disease (AD) diagnosis in its early stages remains difficult with current diagnostic approaches. Though tau neurofibrillary tangles (NFTs) generally follow the stereotypical pattern described by the Braak staging scheme, the network degeneration hypothesis (NDH) has suggested that NFTs spread selectively along functional networks of the brain. To evaluate this, we implemented a Bayesian workflow to develop hierarchical multinomial logistic regression models with increasing levels of complexity of the brain from tau-PET and structural MRI data to investigate whether it is beneficial to incorporate network-level information into an ROI-based predictive model for the presence/absence of AD. Methods This study included data from the Translational Biomarkers in Aging and Dementia (TRIAD) longitudinal cohort from McGill University’s Research Centre for Studies in Aging (MCSA). Baseline and 1 year follow-up structural MRI and [ 18 F]MK-6240 tau-PET scans were acquired for 72 cognitive normal (CN), 23 mild cognitive impairment (MCI), and 18 Alzheimer’s disease dementia subjects. We constructed the four following hierarchical Bayesian models in order of increasing complexity: (Model 1) a complete-pooling model with observations, (Model 2) a partial-pooling model with observations clustered within ROIs, (Model 3) a partial-pooling model with observations clustered within functional networks, and (Model 4) a partial-pooling model with observations clustered within ROIs that are also clustered within functional brain networks. We then investigated which of the models had better predictive performance given tau-PET or structural MRI data as an input, in the form of a relative annualized rate of change. Results The Bayesian leave-one-out cross-validation (LOO-CV) estimate of the expected log pointwise predictive density (ELPD) results indicated that models 3 and 4 were substantially better than other models for both tau-PET and structural MRI inputs. For tau-PET data, model 3 was slightly better than 4 with an absolute difference in ELPD of 3.10 ± 1.30. For structural MRI data, model 4 was considerably better than other models with an absolute difference in ELPD of 29.83 ± 7.55 relative to model 3, the second-best model. Conclusion Our results suggest that representing the data generating process in terms of a hierarchical model that encompasses both ROI-level and network-level heterogeneity leads to better predictive ability for both tau-PET and structural MRI inputs over all other model iterations.