Background:Arterial spin labelling (ASL), greymatter (GM) densities and structural volumes have been shown to be strong predictors of dementia in machine learning classification studies [1,2]. Many of these experiments, however, were performed on single modalities in small 'matched' cohorts using voxels, which introduces noise, increases computation time, and limits interpretability and generalisability. We propose an automated classification pipeline thatworks in the patients' native T1 and ASL MRI spaces, is based on anatomically meaningful regions of interest (ROIs), and iteratively selects predictive biomarkers. Methods: 280 patients with subjective cognitive decline (SCD), mild cognitive impairment (MCI) or probableAlzheimer's disease (AD) were included in the study. For all subjects T1 and ASL MRI as well as basic demographics such as age and sex were available (see Table 1). All subjects' MRI scans underwent brain tissue segmentation and parcellation using geodesic information flows (GIF)  yielding 143 ROIs per subject. Each ROI was associated with mean GMdensity, structural volume and meanASL perfusion.After α-priori exclusion of non-informative regions such as background or skull, 396 features were used for experiments. Support vector machine (SVM) models were used to differentiate pairwise between AD, MCI, and SCD. 50% of all subjects from each class were randomly assigned to either a training or a testing set to create and validate the classifier, and permuted 1000 times. The classifier assigns weights to each feature and our model recursively removes the 20% of features with the lowest average weight over 1000 repetitions. Results: In all three experiments, the highest accuracies can be observed when only 42- 66 out of 396 features are used (see Figure 1). The maximum obtained accuracies are 91.6% for AD vs SCD, 85.3% for AD vs MCI, and 80.1% for MCI vs SCD. The majority of selected features are structural volumes (51%-60%) but ASL also contributes strongly to the classification outcome with ≤33% of biomarkers. An illustration of selected ROIs is shown in Figure 2. Conclusions:Our SVM model provides a stable mechanism to automatically identify multimodal MRI biomarkers relevant to the diagnosis of dementia whilst surpassing the classification accuracy of previous unimodal studies [1,2]. (Figure Presented).