Background: Alzheimer's disease (AD) is the most common form of dementia. While neuropathological changes pathognomonic for AD have been defined, early detection of AD prior to cognitive impairment in the clinical setting is still lacking. Pioneer studies applying machine learning to magnetic-resonance imaging (MRI) data to predict mild cognitive impairment (MCI) or AD have yielded high accuracies, however, an algorithm predicting neuropathological change is still lacking. The objective of this study was to compute a prediction model supporting a more distinct diagnostic criterium for AD compared to clinical presentation, allowing identification of hallmark changes even before symptoms occur. Methods: Autopsy verified neuropathological changes attributed to AD, as described by a combined score for Aβ-peptides, neurofibrillary tangles and neuritic plaques issued by the National Institute on Aging - Alzheimer's Association (NIAA), the ABC score for AD, were predicted from structural MRI data with RandomForest (RF). MRI scans were performed at least 2 years prior to death. All subjects derive from the prospective Vienna Trans-Danube Aging (VITA) study that targeted all 1750 inhabitants of the age of 75 in the starting year of 2000 in two districts of Vienna and included irregular follow-ups until death, irrespective of clinical symptoms or diagnoses. For 68 subjects MRI as well as neuropathological data were available and 49 subjects (mean age at death: 82.8 ± 2.9, 29 female) with sufficient MRI data quality were enrolled for further statistical analysis using nested cross-validation (CV). The decoding data of the inner loop was used for variable selection and parameter optimization with a fivefold CV design, the new data of the outer loop was used for model validation with optimal settings in a fivefold CV design. The whole procedure was performed ten times and average accuracies with standard deviations were reported. Results: The most informative ROIs included caudal and rostral anterior cingulate gyrus, entorhinal, fusiform and insular cortex and the subcortical ROIs anterior corpus callosum and the left vessel, a ROI comprising lacunar alterations in inferior putamen and pallidum. The resulting prediction models achieved an average accuracy for a three leveled NIAA AD score of 0.62 within the decoding sets and of 0.61 for validation sets. Higher accuracies of 0.77 for both sets, respectively, were achieved when predicting presence or absence of neuropathological change. Conclusion: Computer-aided prediction of neuropathological change according to the categorical NIAA score in AD, that currently can only be assessed post-mortem, may facilitate a more distinct and definite categorization of AD dementia. Reliable detection of neuropathological hallmarks of AD would enable risk stratification at an earlier level than prediction of MCI or clinical AD symptoms and advance precision medicine in neuropsychiatry.
Keywords: Alzheimer’s disease; MRI; machine learning; neuroimaging; neuropathology.