Automatic classification of cognitively normal, mild cognitive impairment and Alzheimer's disease using structural MRI analysis

Early detection of dementia for clinical diagnosis is challenging due to high subjectivity and individual variability in cognitive assessments, as well as the evaluation of protein biomarkers, which are mostly used for staging of Alzheimer's disease. Currently, although there is no effective treatment for Alzheimer's disease, early detection of dementia through magnetic resonance imaging analysis may assist in developing preventive measures to slow disease progression. In this paper, we developed an automated machine learning method for classifying cognitively normal aging, early mild cognitive impairment, late mild cognitive impairment, and Alzheimer's disease individuals. In this study, a total of 1167 whole-brain magnetic resonance imaging scans of individuals who are cognitively normal aging controls, early mild cognitive impairment, late mild cognitive impairment, and patients with probable Alzheimer's disease were obtained from the Alzheimer's Disease Neuroimaging Initiative database. We measured regional cortical thickness of both left and right hemispheres (68 features) using FreeSurfer analysis for each individual, and utilized these 68 features for model building. We further tested scans of individuals to classify them into four groups using various machine learning methods. We found that the cortical thickness feature, based on the non-linear support vector machine classifier with radial basis function, showed the highest specificity (0.77), sensitivity (0.75), F-score (0.72), Matthew's correlation coefficient (0.71), Kappa-statistic (0.69), receiver operating characteristic area under the curve (0.76), and an overall accuracy of 75% in classifying all four groups using ten-fold cross-validation with respect to the clinical scale. In addition, we also predicted the features for classifying all four groups using the support vector regression algorithm. The non-linear support vector machine using a radial basis function kernel showed good accuracy in classifying different stages of dementia. Thus, machine learning methods are useful for radiological imaging tasks such as diagnosis, prognosis, risk assessment, and early detection.