Abstract
BACKGROUND: Predicting clinical course of cognitive decline can boost clinical trials' power and improve our clinical decision-making. Machine learning (ML) algorithms are specifically designed for the purpose of prediction; however. identifying optimal features or algorithms is still a challenge. OBJECTIVE: To investigate the accuracy of different ML methods and different features to classify cognitively normal (CN) individuals from Alzheimer's disease (AD) and to predict longitudinal outcome in participants with mild cognitive impairment (MCI). METHODS: A total of 1,329 participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were included: 424 CN, 656 MCI, and 249 AD individuals. Four feature-sets at baseline (hippocampal volume and volume of 47 cortical and subcortical regions with and without demographics and APOE4) and six machine learning methods (decision trees, support vector machines, K-nearest neighbor, ensemble linear discriminant, boosted trees, and random forests) were used to classify participants with normal cognition from participants with AD. Subsequently the model with best classification performance was used for predicting clinical outcome of MCI participants. RESULTS: Ensemble linear discriminant models using demographics and all volumetric magnetic resonance imaging measures as feature-set showed the best performance in classification of CN versus AD participants (accuracy = 92.8%, sensitivity = 95.8%, and specificity = 88.3%). Prediction accuracy of future conversion from MCI to AD for this ensemble linear discriminant at 6, 12, 24, 36, and 48 months was 63.8% (sensitivity = 74.4, specificity = 63.1), 68.9% (sensitivity = 75.9, specificity = 67.8), 74.9% (sensitivity = 71.5, specificity = 76.3), 75.3%, (sensitivity = 65.2, specificity = 79.7), and 77.0% (sensitivity = 59.6, specificity = 86.1), respectively. CONCLUSIONS: Machine learning models trained for classification of CN versus AD can improve our prediction ability of MCI conversion to AD.