Cross-cohort generalizability of deep and conventional machine learning for MRI-based diagnosis and prediction of Alzheimer’s disease

This work validates the generalizability of MRI-based classification Alzheimer’s disease (AD) patients and controls (CN) to an external data set task prediction conversion AD in individuals with mild cognitive impairment (MCI). We used a conventional support vector machine (SVM) deep convolutional neural network (CNN) approach based on structural MRI scans that underwent either minimal pre-processing or more extensive into modulated gray matter (GM) maps. Classifiers were optimized evaluated using cross-validation Disease Neuroimaging Initiative (ADNI; 334 AD, 520 CN). Trained classifiers subsequently applied predict ADNI MCI (231 converters, 628 non-converters) independent Health-RI Parelsnoer Neurodegenerative Diseases Biobank set. From this multi-center study representing tertiary memory clinic population, we included 199 patients, 139 participants subjective decline, 48 converting dementia, 91 who did not convert dementia. AD-CN GM maps resulted similar area-under-the-curve (AUC) for SVM (0.940; 95%CI: 0.924–0.955) CNN (0.933; 0.918–0.948). Application yielded significantly higher performance (AUC = 0.756; 0.720-0.788) than 0.742; 0.709-0.776) (p<0.01 McNemar’s test). In validation, was slightly decreased. For AD-CN, it again gave AUCs (0.896; 0.855–0.932) (0.876; 0.836–0.913). MCI, performances decreased both 0.665; 0.576-0.760) 0.702; 0.624-0.786). Both CNN, outperformed minimally processed images (p=0.01). Deep performed equally well their only when cohort. expect validation contributes towards translation learning clinical practice.