Diagnosis of Alzheimer’s disease and frontotemporal dementia using FDG‐PET: Application of genetic algorithms

Abstract

Background: Genetic algorithms are methods used in machine learning, which have a proven capability to explore a large space of solutions, deal with very large numbers of input features, and avoid local minima. Diagnosis of Alzheimer's Disease (AD) and Frontotemporal dementia (FTD) is often challenging, and thorough costly assessments are often needed. We hypothesised that the application of machine learning, and specifically genetic algorithms, to 18F-Fluorodeoxyglucose Positron Emission Tomography (FDG-PET) may help in diagnosis by selecting the most meaningful features and automating diagnosis. In this study, we aimed to develop algorithms for three common situations: discrimination between patients with AD or FTD and healthy controls (HC), differential diagnosis between behavioural FTD (bvFTD) and AD, and differential diagnosis between primary progressive aphasia (PPA) variants.Method: Eighty-one patients with bvFTD, 88 patients with AD, 68 patients with PPA, and 39 HC were enrolled. Patients underwent a comprehensive clinical and neuropsychological protocol, and FDG-PET imaging. Genetic algorithms, customised with K-Nearest Neighbor and BayesNet Naives as the fitness function, were developed applied to the PET imaging and compared with Principal Component Analysis. The highest accuracy rate and most relevant features were identified. K-fold cross validation within the same sample and external validation with ADNI samples were performed.Result: Discrimination accuracy of FDG-PET was 92-95% for AD vs HCs, 95-96% for bvFTD vs HCs, 89-90% for AD vs bvFTD, and 90-91% for classification of PPA subtypes. A reduced number of features was achieved, with cutting rates from 62% to 95.69% regarding the total number of variables, and several key brain regions were selected. External validation with ADNI obtained an accuracy of 82.93% for BayesNet Naives algorithm for the differentiation between AD and HC. Conclusion:Genetic algorithms identified the most meaningful features with the minimum set of features, which may be relevant for automated assessment of brain FDG-PET images. Overall, our study contributes to the development of an automated, and optimised diagnosis of neurodegenerative disorders using brain metabolism.