Alzheimer's Disease Classification Confidence of Individuals using Deep Learning on Heterogeneous Data

Abstract

Making accurate diagnosis of Alzheimer's disease (AD) is crucial for effective treatment and management. Although deep learning has been applied to AD classification, it is typically performed at group level, the data used are not sufficiently heterogeneous and comprehensive, and decision confidence is not evaluated at individual (single patient) level. This paper proposed a more practical deep learning approach that not only detects AD stages of individuals, but also provides its corresponding confidence estimation. In particular, in addition to a convolutional neural network (CNN), we incorporated a softmax confidence metric based on the network's output activity to evaluate its classification confidence. Further, we applied this approach to a heterogeneous and comprehensive data that comprised cognitive and functional assessments, tau-PET and MRI neuroimaging, medical/family history, demographic, and APoE genotype. Importantly, we utilised leave-one-out cross-validation to train the CNN and classify an individual's healthy control, mild cognitive impairment or AD state, while concurrently estimating each output decision's confidence. We showed that, over different confidence softmax temperature values, CNN could attain classification accuracies at 83-85% for the three classes while having robust confidence scores of 78-83%. Further improvement in confidence breakdown was achieved using the optimal temperature value in confidence evaluation, with higher confidence scores for correct than error decisions. Overall, the computed classification confidence of an individual may aid clinicians and other stakeholders in understanding the reliability of the model's decision outcome and offer better trust. The implication of this work may extend to other classification applications, in which the confidence level of a single deep learning-based decision can be evaluated.