Simultaneous modeling of Alzheimer's disease progression via multiple cognitive scales

Abstract

Analyzing the progression of Alzheimer's disease (AD) is challenging due to lacking sensitivity in currently available measures. AD stages are typically defined based on cognitive cut‐offs, but this results in heterogeneous patient groups. More accurate modeling of the continuous progression of the disease would enable more accurate patient prognosis. To address these issues, we propose a new multivariate continuous‐time disease progression (MCDP) model. The model is formulated as a nonlinear mixed‐effects model that aligns patients based on their predicted disease progression along a continuous latent disease timeline. The model is evaluated using long‐term follow‐up data from 2152 participants in the Alzheimer's Disease Neuroimaging Initiative. The MCDP model was used to simultaneously model three cognitive scales; the Alzheimer's Disease Assessment Scale‐cognitive subscale, the Mini‐Mental State Examination, and the Clinical Dementia Rating scale—sum of boxes. Compared with univariate modeling and previously proposed multivariate disease progression models, the MCDP model showed superior ability to predict future patient trajectories. Finally, based on the multivariate disease timeline estimated using the MCDP model, the sensitivity of the individual items of the cognitive scales along the different stages of disease was analyzed. The analysis showed that delayed memory recall items had the highest sensitivity in the early stages of disease, whereas language and attention items were sensitive later in disease.

Estimating the longitudinal trajectory of cognitive function measurement using short‐term data with different disease stages: Application in Alzheimer's disease

“…To characterize the longitudinal trajectory of cognitive function measures, it is ideal to conduct a large cohort study over a couple of decades that periodically collects data of participants with normal cognition (NC) with baseline characteristics of interest. However, except for a few studies with long-term follow-up data of 10 to 20 years, [10][11][12] most studies have involved participants at different disease stages (eg, NC, mild cognitive impairment [MCI], and AD) with cognitive function measures collected over only a few years. [13][14][15][16] Establishing a statistical methodology that estimates the underlying longitudinal trajectory of cognitive function measures using short-term follow-up data is challenging.…”

“…For instance, we can easily modify the individual components of the four steps (eg, variables included in the mixed-effects model or types of fractional polynomial functions) without changing the remaining components to optimize the trajectory estimation. Compared with the global likelihood-based modeling approaches, 12,21 the proposed method also does not require a complex nonlinear mixed-effects model with multiple parameters and consequently avoids problems with convergence to maximum likelihood estimates, which are often encountered in limited sample sizes. However, the proposed method may induce potential biases in the trajectory estimation because the random variation of MMSE is passed through several steps of the proposed method.…”

“…While the methods of Donohue et al 19 and Li et al 20 transform the multiple cognitive outcomes to the corresponding quantiles using a weighted transformation prior to modeling, Raket 21 directly modeled the nonlinear disease progression using non‐transformed data and studied the univariate modeling of ADAS‐Cog scores with covariate adjustment. More recently, Kühnel et al 12 extended this univariate framework to multivariate outcomes on an original scale. These methods are based on the global modeling approaches using nonlinear mixed‐effects models 12,19‐21 including random effects (ie, time‐shift variation and additive effect on a cognitive scale) and can be beneficial, especially when using long‐term follow‐up data (eg, $$$ \ge $$$10 years), such as the Alzheimer's Disease Neuroimaging Initiative (adni.loni.usc.edu).…”

“…More recently, Kühnel et al 12 extended this univariate framework to multivariate outcomes on an original scale. These methods are based on the global modeling approaches using nonlinear mixed‐effects models 12,19‐21 including random effects (ie, time‐shift variation and additive effect on a cognitive scale) and can be beneficial, especially when using long‐term follow‐up data (eg, $$$ \ge $$$10 years), such as the Alzheimer's Disease Neuroimaging Initiative (adni.loni.usc.edu). Meanwhile, the application of these methods to short‐term follow‐up data would be questionable, although in principle they could be applied to short‐term follow‐up data with sufficient overlap between patient‐level trajectories to piece together a disease trajectory.…”

Disease Progression in Multiple System Atrophy—Novel Modeling Framework and Predictive Factors

“…A similar framework has been used to analyze the progression of Alzheimer's disease in several cohorts where it has been shown to make state-of-the-art predictions of individual patient progression along the Alzheimer's disease continuum, predictions of future decline, and provide novel insights into the evolution of biomarkers with disease progression. [22][23][24] In this paper, we present a disease-progression model framework for MSA progression and analyze progression patterns of patients in the European MSA (EMSA) natural history study. 14 We show how the method can predict a natural, personalized staging of patients and population-based mean progression trajectories of clinical scores along the disease continuum.…”