Verifying and Refining Early Statuses in Alzheimer’s Disease Progression: A Possibility from Deep Feature Comparison

Abstract

Abstract Background: Defining effective early status of Alzheimer’s disease (AD) could be challenging, due to complexity in linking early and late stages in the progression and the internal biological heterogeneity within same status. We explore whether it is possible to verify and refine candidature early statuses in the AD progressions by comparing the neurological features learned by deep learning models. Methods: We collect functional magnetic resonance imaging (fMRI) data from 432 subjects, including 79 healthy controls (HCs), 109 amnestic mild cognitive impairments (aMCIs), 39 non-amnestic MCIs (naMCIs), 98 subjective cognitive declines (SCDs) and 107 ADs. We train graph convolutional networks (GCNs) based on multiscale functional networks to accurately classify aMCI from naMCI and HC from MCIs. The trained models are applied to AD and SCD groups to suggest the neural feature similarity (as the ratio of predictions) among the statuses and identify clinically informative subpopulations. The corresponding demographics, cognitive assessments, T1, and PET images are used to provide supportive evidences for sub-divided populations based on the model decisions. Results: The GCN model achieves 89.2±1.9% and 83.7±3.1% accuracies in aMCI-vs-naMCI and HC-vs-MCI classifications. The aMCI-vs-naMCI classification model identifies 71.8% of the AD subjects as aMCI. The HC-vs-MCI classification model suggests that 73.5% of the SCDs are MCI, in which 88.8% are further diagnosed as “aMCI” by the aMCI-vs-naMCI classifier. The analyses based on T1 and PET images suggests that the aMCI-like AD exhibits more globally elaborated Aβ depositions, severer glucose metabolism reduction and grey matter atrophy than naMCI-like AD after contrasted with clinical HCs. MCI-like SCD shows more reduction of glucose metabolism than HC-like SCD, baselined by clinical HCs. Further, aMCI-like SCD exhibits slightly elaborated Aβ while naMCI-like SCD shows none when compared to clinical HCs. MCI-like SCD has lower executive ability than HC-like SCD. aMCI-like SCD shows lower memory ability than naMCI-like SCD. Conclusions: This study suggests an overall neurological similarity among SCD, aMCI and AD from perspective of brain dynamics, and verifies the position of SCD and aMCI in the AD progression. Further, it offers a potentially refined progression progress, consisting of aMCI-like SCD, aMCI, and aMCI-like AD dementia. Clinical Trials Registration: The data collection has been registered as “ChiCTR2000036842”.