Functional assessment improves discrimination of clinical and biomarker-defined Alzheimers disease
TL;DR
In a contemporary, diverse cohort, most established dementia risk algorithms had limited power in discriminating amyloid positivity, pTau-217 positivity, and current cognitive status despite acceptable calibration.
Abstract
INTRODUCTION: Accurately identifying risk of Alzheimers Dementia (AD) is essential for supporting people living with symptoms in clinical settings, as well as recruiting adults in prospective medical research. Various algorithms have been created to calculate AD risk based on evidenced risk factors which are weighted toward a total score. As daily life conditions determining risk change at scale, it remains unclear how effective gold standard algorithms remain in modern cohorts.
Methods
In the Bio-Hermes-001 diverse cohort, we assessed algorithm discrimination and calibration in six outcomes: classifying AB PET binary outcome (negative N = 603, positive N = 342); phosphorylated tau-217 binary outcome (pTau-217 negative N = 166, positive N = 469); participants with Healthy Cognition (N = 417) from probable AD (N = 272); HC from Mild Cognitive Impairment (N = 312), HC from pooled MCI or AD; and MCI from AD. Approximately a third of the cohort are individuals from populations typically underrepresented in dementia research (HC: 19%; MCI: 24%; AD: 33%).
Results
Hosmer-Lemeshow tests and Brier score demonstrate acceptable calibration of all algorithms except the oldest algorithm. However, Receiver Operating Characteristic (ROC) curves and the associated area under the curve (AUC) estimates evidenced that in this cohort only the BDSI exceeded conventional thresholds for good discrimination (.8 AUC in HC-AD classification, with AUC approximately .7 in the other clinical, AB PET, and pTau-217 comparisons). When the functional item is removed from the BDSI score, it remains acceptably calibrated, but DeLong tests reflect statistically significant reduction in discriminatory performance for all group comparisons. The two earliest published algorithms were only chance-level accurate.
Discussion
In a contemporary, diverse cohort, most established dementia risk algorithms had limited power in discriminating amyloid positivity, pTau-217 positivity, and current cognitive status despite acceptable calibration. Including a functional measure markedly improved discrimination across both clinical and biomarker-defined outcomes, suggesting that proximal indicators of cognitive vulnerability are critical for identifying individuals with underlying AD-related pathology.