Introduction

The major socioeconomic and healthcare burdens imposed by Alzheimer’s disease (AD) have pushed the focus of clinical research dramatically toward prevention and treatment in pre-symptomatic stages (1,2). This shift has been well-aligned with guidance from the FDA in support of earlier stage therapies and new measurement methodologies for establishing clinically meaningful effects of those therapies (3).

However, recent AD trials have faced significant challenges identifying and enrolling subjects who meet thresholds for AD biomarker positivity but who have not yet experienced notable cognitive deficits. Despite great efforts made to enroll subjects in pre-clinical or early stage AD, trial sponsors have seen screen failure rates as high as 80%, primarily driven by required biomarker thresholds in cognitively normal subjects, leading to significantly prolonged enrollment periods and increased costs (4–9).

This underscores an urgent need for better approaches to pragmatically and cost effectively identify subjects who: 1) are cognitive normal; 2) will decline cognitively within 1–3 years; and 3) are likely to have PET scan positive AD biomarkers. Having such predictive capabilities will accelerate enrollment of specific subjects and will also improve study design for potentially shorter trial durations. The present study focuses on predicting impending cognitive decline in cognitively normal subjects. Predicting positivity for AD biomarkers will be discussed in a future publication.

Many efforts have focused on developing more sensitive cognitive assessment tools (e.g., composite scoring), including PACC (10) and ADCOMS (11). While some recently validated assessment tools can outperform less sensitive tools developed to assess dementia severity, as a group they lack the capability to predict impending cognitive decline in cognitively normal subjects.

One such assessment approach could arise from the application of hierarchical Bayesian cognitive process (HBCP) models to item-response data from a wordlist memory (WLM) test. This approach can generate digital cognitive biomarkers (DCB) that correspond to underlying cognitive processes of encoding, storage, and retrieval into and from various states of learning and memory (Figure 1). Such underlying cognitive processes cannot be directly observed or measured, while DCBs can quantify these processes, providing insights into cognitive function that traditional assessment approaches cannot provide. The details of this HBCP model and the generation of DCBs have been previously discussed elsewhere (12).

Figure 1
figure 1

Hierarchical Bayesian Cognitive Process Model

The HBCP model can quantify underlying cognitive processes that cannot be observed or measured using the traditional approaches, and provide significant insights into how each cognitive process is affected by different conditions. Parameter r corresponds to one-shot learning; a, partial leaning; v, consolidation; b, testing effect; t, immediate retrieval from transient storage; L1, immediate retrieval from durable storage; and L2, delayed retrieval from durable storage

Full size image

In our previous work, HBCP-generated DCBs demonstrated the ability to distinguish groups of individuals with impending cognitive decline from those who would remain cognitively normal, using baseline, item-response data from a WLM test. This study was conducted using Auditory-Verbal Learning Test (AVLT) item response data from the Mayo Clinic Alzheimer’s Disease Patient Registry (13). Subjects, including those with normal cognition at baseline who would progress to amnestic MCI and those who would progress to AD dementia, were compared to those who would not decline. Bayes factor assessment identified notable reductions in Immediate Retrieval from Durable Storage, L1 (BFds = 30.4), and Delayed Retrieval from Durable Storage, L2 (BFds > 100). This study also appeared to identify compensatory increases in One-shot Learning, r (BFds = 3.2); Partial Learning, a (BFds = 10.8); and Consolidation, v (BFds = 13.5). However, subsequent work with our HBCP model did not replicate this apparent compensatory effect (12).

The present study, using baseline data from a novel WLM dataset, was designed to replicate our previous findings of deficits in retrieval DCBs for a group of individuals with impending cognitive decline due to AD, and compared this outcome to those generated by traditional scoring approaches. Replicating these results supports the role of DCBs for accelerating the clinical trial recruitment process and could greatly benefit future decisions about clinical trial study designs.

Methods

We used baseline ADAS-Cog WLM item response data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI: https://www.adni-info.org) database (adni.loni. usc.edu). The ADNI was launched in 2003 as a public-private partnership with the primary goal to test whether serial magnetic resonance imaging, positron emission tomography, other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment and early AD.

From the ADNI dataset, we classified non-decliner subjects (n = 442) as those whose diagnosis remained normal for 3 or more years after normal baseline assessment and decliner subjects (n = 61) as those who developed amnestic MCI or AD dementia within 3 years of normal baseline assessment. Table 1 shows sample characteristics.

Table 1 . ADNI Sample Characteristics
Full size table

Three analytic approaches were compared to demonstrate the relative value of DCBs.

Logistic Regression

Traditional summary scores per ADAS-Cog task were assessed for group differences. Logistic regression modeling was performed with individual subjects’ summary scores for immediate free recall tasks 1 through 3 and for the delayed free recall task, each included as predictors of impending cognitive decline as the outcome.

Bayesian Modeling

A Bayesian model of summary scores was assessed. Gaussian distributions were fitted to individuals’ number of items recalled, from 1 to 10, and uniform distributions for probability of 0 items recalled, on each free recall task.

HBCP Model

The HBCP model was applied to non-decliner and decliner group item response data, aggregated across subjects within each group. The model estimated DCBs across WLM items with the Batchelder multinomial processing tree model of memory for each item’s recall pattern across the four free recall tasks (12).

Results

Logistic regression of summary scores generated β coefficients (Table 2) that did not significantly discriminate between groups, either individually or across the test as a whole.

Table 2 . Summary Score Logistic Regression Analysis
Full size table

Bayes Factor assessment of fitted Gaussian distributions to each free recall task by summary score measurement provided moderate evidence that the groups were measurably equivalent (BFsd = 3.4, 3.1, 2.9, and 1.4, respectively; Figure 2).

Figure 2
figure 2

Bayes Factors for Summary Score Group Difference Parameters

Posterior distributions of Bayesian-modeled summary scores and posterior distributions of mean differences across ADAS-Cog tasks are presented, with Savage-Dickey density ratio Bayes factors calculated for mean differences against prior distributions of no change.

Full size image

The HBCP model produced posterior distributions of group differences (Figure 3). Bayes Factor assessment identified three DCBs with notable group differences: Immediate Retrieval from Durable Storage, L1 (BFds = 11.8, strong evidence), One-shot Learning, r (BFds = 4.5, moderate), and Partial Learning, a (BFds = 2.9, weak).

Figure 3
figure 3

Bayes Factors for HBCP DCB Notable Group Difference Parameters

Posterior distributions of DCB mean differences are presented against prior distributions of no change for three parameters with notable group differences, along with Savage-Dickey density ratio Bayes factors.

Full size image

Discussion

The present study validated our previous findings by demonstrating the HBCP DCBs’ ability to distinguish a group of cognitively normal individuals with impending cognitive decline from a group that would remain cognitively normal. This study also showed DCBs’ advantages over the traditional approach of summary score assessments and their applicability for detection of impending cognitive decline in asymptomatic AD patients.

The HBCP DCBs have an advantage over composite or summary score approaches because of their ability to measure and quantify underlying cognitive processes (Figure 1). Among these processes, only some are affected in the cognitively normal or pre-clinical stages of AD (14), and each is affected differently as the disease progresses (15). The HBCP model can also be applied to any existing, well-validated, WLM test protocol (e.g., AVLT, ADAS-Cog, MCI Screen), so DCBs can be generated on WLM data from past academic studies and clinical trials to examine which processes were improved by particular AD therapies (16) and which were not, even when traditional outcome measures identified no overall differences. This will provide novel insights into efficacy and trial design, potentially targeting different cognitive or disease processes.

This study warrants further development of the HBCP DCBs to predict impending cognitive decline at the individual level and to predict other factors associated with AD, such as the identification of stage progression, the accumulation of biomarkers, and the presence of other cognition-impairing conditions.