Contribution of Global Amyloid-PET Imaging for Predicting Future Cognition in the MEMENTO Cohort

Abstract

Background and objectives: Global amyloid-PET is associated with cognition and cognitive decline, but most research on this association does not account for past cognitive information. We assessed the prognostic benefit of amyloid-PET measures for future cognition when prior cognitive assessments are available, evaluating the added value of amyloid measures beyond information on multiple past cognitive assessments.

Methods: The French MEMENTO cohort (a cohort of outpatients from French research memory centers to improve knowledge on Alzheimer disease and related disorders) includes older outpatients with incipient cognitive changes, but no dementia diagnosis at inclusion. Global amyloid burden was assessed using positron emission tomography (amyloid-PET) for a subset of participants; semiannual cognitive testing was subsequently performed. We predicted mini-mental state examination (MMSE) scores using demographic characteristics (age, sex, marital status, and education) alone or in combination with information on prior cognitive measures. The added value of amyloid burden as a predictor in these models was evaluated with percent reduction of the mean squared error (MSE). All models were conducted separately for evaluating the added value of dichotomous amyloid positivity status compared with a continuous amyloid-standardized uptake-value ratio.

Results: Our analytic sample comprised 510 individuals who underwent amyloid-PET scans with at least 4 MMSE assessments. The mean age at the PET scan was 71.6 (standard deviation 7.4) years; 60.7% were female. The median follow-up was 4.6 years (interquartile range: 0.9 years). Adding amyloid burden when adjusting for only demographic characteristics reduced the MSE of predictions by 5.08% (95% CI 0.97%-10.86%) and 12.64% (95% CI 3.35%-25.28%) for binary and continuous amyloid, respectively. If the model included 1 past MMSE measure, the MSE improvement was 3.51% (95% CI 1.01%-7.28%) when adding binary amyloid and 8.83% (95% CI 2.63%-16.37%) when adding continuous amyloid. Improvements in model fit were smaller with the addition of amyloid burden when more than 1 past cognitive assessment was included. For all models incorporating past cognitive assessments, differences in predictions amounted to a fraction of 1 MMSE point on average.

Discussion: In a clinical setting, global amyloid burden did not appreciably improve cognitive predictions when past cognitive assessments were available.

Trial registration information: ClinicalTrials.gov Identifier: NCT02164643.

Figure 1. Flowchart of Inclusions and Exclusions and Sample Splitting Into Training and Testing Sets

To avoid predicting from approximately contemporaneous amyloid burden, we predict only cognitive tests at least quarter of a year after PET scan.

Figure 2. Reduction in Mean Squared Error

Percent reduction (95% CI) in the mean squared error with the addition of optimized binary amyloid (A) or continuous amyloid (B) to models including demographic covariates, including demographics plus 1 or 2 past cognitive assessments (CAs), or including demographics, average of past CAs, and slope of change in past CAs (trajectories).

Figure 3. Reduction in Mean Absolute Error

Absolute reduction (95% CI) in the mean absolute error with the addition of optimized binary amyloid (A) and continuous amyloid (B) to models including demographic covariates, including demographics plus 1 or 2 past cognitive assessments (CAs), or including demographics, average of past CAs, and slope of change in past CAs (trajectories). Values are on the original MMSE scale, so an improvement of 0.02 implies that the addition of amyloid data would bring the average prediction 0.02 units closer to the true MMSE score. MMSE = mini-mental state examination.

Figure 4. Predicted Cognitive Scores With and Without Amyloid

Cognitive scores predicted from models with amyloid plotted against cognitive scores predicted from models without amyloid, using (A) binary amyloid or (B) continuous amyloid to improve predictions, by amyloid positivity status. Amyloid-negative individuals are shown in blue and amyloid-positive individuals are shown in red. As expected, incorporating amyloid information into the models leads to slightly lower average predicted scores for amyloid-positive individuals (red dots) and slightly higher average predicted scores for amyloid-negative individuals (blue dots). The visible substructure in the demographic model figure is due to repeated assessments on individuals.