Machine learning for neuropsychological assessment of Alzheimer’s disease and behavioral variant frontotemporal dementia

Abstract

AbstractBackgroundDiagnosis of Alzheimer’s disease and behavioral variant frontotemporal dementia is often challenging. In spite of comprehensive clinical and cognitive assessments, the use of biomarkers is usually needed. We aimed to develop machine learning based models for the diagnosis of AD and bvFTD using only cognitive testing. These techniques may allow selecting the most relevant tests for an optimized neuropsychological diagnosis.MethodWe included 329 participants: 171 patients with AD, 72 patients with bvFTD, and 87 healthy controls (HC). All patients met the current diagnostic criteria, had a neuroimaging compatible with FDG‐PET, and had at least two years of follow‐up confirming the diagnosis. A comprehensive neuropsychological protocol was performed. Evolutionary algorithms were developed, including NSGA‐II. F1‐score was calculated, as a measure of accuracy. Algorithms were developed for the following classification problems between AD vs FTD, AD vs HC, FTD vs HC, and AD/FTD vs HC.ResultsDifferentiation between FTD vs HC, AD vs HC and FTD/AD vs HC reached an F1 superior to 90%. Differentiation between FTD vs AD was slightly inferior (F1 80‐85%). Test selected by the NSGAII algorithm for AD vs FTD were as follows: Symbol Digit Modalities Test, Stroop test, Trail Making Test, Corsi blocks, and ACE‐III. For FTD vs HC, the algorithm selected Free and Cued Selective Reminding Test, Trail Making Test, Rey Figure (copy type), Boston Naming Test, and ACE. For AD vs HC, the following test was selected: Rey Figure, Trail Making Test, Corsi test, and verbal fluency.ConclusionsWe have developed a machine‐learning approach to perform feature selection and modeling of neuropsychological test scores, with a high level of discrimination between groups. Our study suggests the interest of applying machine learning for an optimized use of cognitive tests and to improve the interpretation of neuropsychological assessment. . These algorithms allow maximizing the diagnostic capacity, selecting the tests with best characteristics, and the automation, which may be useful for diagnosis.