Abstract
Background
While multiple biomarkers of Alzheimer's disease (AD) and mild cognitive impairment (MCI) predict cognitive decline over time, their predictive value at each of the disease stages remains unclear.
Objective
To examine how the predictive power of biomarkers changes over time.
Methods
We dynamically ranked a comprehensive set of multimodal biomarkers-including
APOE
genotype, medical history, structural MRI, FDG-PET, amyloid PET, CSF markers, metabolic measures, and neuropsychiatric tests-from baseline to 30 months in ADNI participants. Feature importance for each timepoint was estimated using random forests and features were clustered based on their patten of importance over time. As a post-hoc, the top features over time were input into a long short-term memory (LSTM) model for future ADAS-13 prediction.
Results
In 981 participants (751 MCI, 230 AD), clustering of top feature importances over time revealed three trajectories in MCI (stable, early-only, and declining) and four in AD (early-only, stable, early-declining, and late-increasing). The top important features over time were cognitive scores, CSF Aβ
42
, tau, imaging biomarkers (FDG-PET hypometabolic convergence index, temporal/parietal cortical thickness), metabolic measures (serum albumin, glucose), and apolipoproteins and omega-3. Optimized LSTM models achieved peak R
2
= 0.86 (RMSE = 3.90) in MCI and R
2
= 0.78 (RMSE = 5.86) in AD using as few as 10 features.
Conclusions
Early-stage prognosis relies on CSF Aβ
42
, p-tau181, and tau; short-term decline is best predicted by FDG-PET; structural MRI shifts from hippocampal/entorhinal to parietal/network regions over time; metabolic markers remain consistently informative; and ADAS/MMSE gain value with advancing global decline.