Predicting early Alzheimer’s disease from diffusion tensor imaging using machine learning techniques

Abstract

AbstractBackgroundThe major current challenge in Alzheimer’s disease (AD) is the identification of individuals likely to develop dementia. Non‐invasive imaging techniques that reflect the integrity of the brain’s white matter are potential parameters to study the pathogenesis of AD. Diffusion tensor imaging (DTI) provides tools to analyze AD‐related brain changes despite the large amount of data generated. To process large amounts of data, computational methods capable of finding relationships among large and complex data sets are needed. In this context, machine learning (ML) techniques have been extended to address this issue. This study aimed to evaluate whether DTI features could predict future AD in non‐AD patients using ML techniques.MethodSixty‐two subjects, with mild cognitive impairment and healthy controls, were enrolled and followed for approximately eleven months. At the first assessment, MRI, neuropsychological tests, and cerebrospinal fluid analysis were performed. At the second assessment, ten had progressed to Alzheimer’s dementia. The clinical diagnosis was used to train the ML classification models. Five data sets were generated with different missing data strategies. Feature selection strategies measured by information gain were implemented and compared. Twenty ML algorithms distributed over several mathematical approaches were used in model training. A deep learning model was also implemented for comparison with ML results.ResultThe best model for predicting AD by DTI was Random Forest with kNN missing data imputation and features selected by information gain. This model achieved a test accuracy of 88.72%. The most relevant DTI tract that contributed to the prediction of the model was the fractional anisotropy (FA) of the cingulum ‐ hippocampus . When other categories of data, such as neuropsychological tests, were included in the model, the accuracy hardly changed (90.26%), suggesting that DTI alone is a good predictor of Alzheimer’s disease. When the deep learning model was run, the constructed neural network achieved 92.31% accuracy.ConclusionOur results suggest that DTI measures analyzed by ML models, such as Random Forest, may predict AD in non‐AD patients. Furthermore, DTI measure of cingulum ‐ hippocampus tract may be a sensitive marker of early AD pathology.