Cognitive Profiling Related to Cerebral Amyloid Beta Burden Using Machine Learning Approaches

Abstract

Background: Cerebral amyloid beta (Aβ) is a hallmark of Alzheimer's disease (AD). Aβ can be detected in vivo with amyloid imaging or cerebrospinal fluid assessments. However, these technologies can be both expensive and invasive, and their accessibility is limited in many clinical settings. Hence the current study aims to identify multivariate cost-efficient markers for Aβ positivity among non-demented individuals using machine learning (ML) approaches. Methods: The relationship between cost-efficient candidate markers and Aβ status was examined by analyzing 762 participants from the Alzheimer's Disease Neuroimaging Initiative-2 cohort at baseline visit (286 cognitively normal, 332 with mild cognitive impairment, and 144 with AD; mean age 73.2 years, range 55-90). Demographic variables (age, gender, education, and APOE status) and neuropsychological test scores were used as predictors in an ML algorithm. Cerebral Aβ burden and Aβ positivity were measured using ¹⁸F-florbetapir positron emission tomography images. The adaptive least absolute shrinkage and selection operator (LASSO) ML algorithm was implemented to identify cognitive performance and demographic variables and distinguish individuals from the population at high risk for cerebral Aβ burden. For generalizability, results were further checked by randomly dividing the data into training sets and test sets and checking predictive performances by 10-fold cross-validation. Results: Out of neuropsychological predictors, visuospatial ability and episodic memory test results were consistently significant predictors for Aβ positivity across subgroups with demographic variables and other cognitive measures considered. The adaptive LASSO model using out-of-sample classification could distinguish abnormal levels of Aβ. The area under the curve of the receiver operating characteristic curve was 0.754 in the mild change group, 0.803 in the moderate change group, and 0.864 in the severe change group, respectively. Conclusion: Our results showed that the cost-efficient neuropsychological model with demographics could predict Aβ positivity, suggesting a potential surrogate method for detecting Aβ deposition non-invasively with clinical utility. More specifically, it could be a very brief screening tool in various settings to recruit participants with potential biomarker evidence of AD brain pathology. These identified individuals would be valuable participants in secondary prevention trials aimed at detecting an anti-amyloid drug effect in the non-demented population.

Keywords: Alzheimer’s disease; amyloid beta deposition; cognitive profiling; machine learning; neuropsychological assessment.

FIGURE 1

Multivariate patterns of demographic information and cognitive measures predicting amyloid positivity in mild (A), moderate (B), and severe group (C). ADAS13, Alzheimer’s disease assessment scale; APOE, ApoE 𝜀4 positivity; AVLT, Rey auditory verbal learning test; AVLT _B, Rey auditory verbal learning test list B; ANRAT; American national adults reading test; BNT, Boston naming test, CLOCK, clock drawing task, CLOCK_Copy; clock drawing task copy; LM_del, logical memory delayed recall; MMSE, mini-mental state examination; MoCA, Montreal cognitive assessment scale.

FIGURE 2

Classification accuracy as indexed by the receiver-operating characteristic (ROC) curves and their area under the curve (AUC) on the training set. Mild, mild change group (CN+SMC+EMCI); Moderate, moderate change group (SMC+EMCI+LMCI); Severe, severe change group (EMIC+LMCI+AD).

FIGURE 3

Classification accuracy as indexed by the receiver-operating characteristic (ROC) curves and their area under the curve (AUC) on the testinkg set. Mild, mild change group (CN+SMC+EMCI); Moderate, moderate change group (SMC+EMCI+LMCI); Severe, severe change group (EMIC+LMCI+AD).

FIGURE 4

Distributions of the Aβ retention (SUVR) across groups.