Feasibility Study: Machine Learning in Neurodegenerative Disorders, Alzheimer’s Disease

Abstract

Introduction: Clinical decision support systems (CDSSs), powered by machine learning and artificial intelligence, have demonstrated potential in clinical diagnosis and intervention for neurological and psychiatric disorders. Considering the importance of early detection and intervention of Alzheimer’s disease (AD), this study aims to explore the potential of a data-driven non-knowledge-based machine learning CDSS for predicting AD diagnoses in individuals. In non-knowledge-based CDSSs, no prior knowledge about AD or any other disorder impacts the decision-making of classification models Method: In this study, publicly available data of 14037 data points collected by the Alzheimer’s Disease Neuroimaging Initiative were used for model training and testing. Binary classification and multiclassification machine learning were applied, and results from six mainstream classification models were analyzed. Results: The binary classification models (AD diagnosis present or absent) gave accuracies around 0.92-0.93, and the multiclassification models gave accuracies around 0.85-0.87. Logistic regression model (binary classification) had the highest overall hit rate (0.93). This model maintained this hit rate when only features with over 90% non-empty data are available. Discussion: Binary classification models are more reliable for diagnosing AD than multiclassification models. The high hit rates of the logistic regression model (binary classification) on generally available data implicate its feasibility. Conclusion: There is strong potential for a complete machine learning-based CDSS to aid in AD diagnoses in the future