Early detection of amyloid-related changes in memory among cognitively unimpaired older adults with daily digital testing

Abstract

Objective:

To determine whether assessing learning over days reveals Alzheimer’s disease(AD)-biomarker related declines in memory consolidation that are otherwise undetectable with single timepoint assessments.

Methods:

Thirty-six (21.9%) cognitively unimpaired older adults (aged 60–91) were classified with elevated $\text{β}$-amyloid ($\text{Aβ}+$) and 128 (78%) were $\text{Aβ}-$ using Positron Emission Tomography with ^11CPittsburgh Compound-B (PiB). Participants completed the Multi-Day Boston Remote Assessment for Neurocognitive Health (BRANCH) for 12 min/day on personal devices (i.e., smartphones, laptops) which captures the trajectory of daily learning of the same content on three repeated tests (Digit Signs, Groceries, Face-Name). Learning is computed as a composite of accuracy across all three measures. Participants also completed standard in-clinic cognitive tests as part of the Preclinical Alzheimer’s Cognitive Composite-(PACC-5) with 123 participants undergoing PACC-5 follow-up after 1.07(SD=0.25) years.

Results:

At the cross-section, there were no statistically significant differences in performance between $\text{Aβ}+/-$ participants on any standard in-clinic cognitive tests (e.g., PACC-5) or on Day 1 of Multi-Day BRANCH. $\text{Aβ}+$ participants exhibited diminished 7-day learning curves on Multi-Day BRANCH after 4 days of testing relative to $\text{Aβ}-$ participants (Cohen’s d=0.49, 95%CI=0.10–0.87). Diminished learning curves were associated with greater annual PACC-5 decline (r=0.54, p<0.001).

Interpretation:

Very early $\text{Aβ}-$related memory declines can be revealed by assessing learning over days, suggesting that failures in memory consolidation predate other conventional amnestic deficits in AD. Repeated digital memory assessments, increasingly feasible and uniquely able to assess memory consolidation over short time periods, have the potential to be transformative for detecting the earliest cognitive changes in preclinical AD.

INTRODUCTION

The ability to detect and track subtle Alzheimer’s disease (AD)-related cognitive decrements at the preclinical stage of disease has been a significant challenge for the field. Standard cognitive assessments, administered in the clinic at a single timepoint, exhibit small and tenuous associations with AD biomarkers of amyloid and tau amongst cognitively unimpaired(CU) older adults^1,2 if they are observed at all. Despite this, CU older adults harboring AD biomarkers decline on standard assessments, when observed over longer intervals, are at greater risk for progression to dementia^3,4. While it is possible that subtle decrements are only observable just prior to the onset of impairment, it is more likely that standard cognitive instruments and annual administration schedules are insufficiently sensitive to capture the earliest stages of a slowly progressive neurodegenerative disease.

A few converging lines of evidence suggest that decrements in memory consolidation, a process whereby a temporary, labile memory is transformed into a more stable, lasting form, may be a very early cognitive indicator of preclinical AD. For example, diminished practice effects (lack of expected improvement on repeated testing) are related to AD pathology among CU older adults^5,6 particularly for memory tests which involve re-testing using the exact same stimuli⁷. More specifically, recent work showed that decrements in learning on monthly testing of the same, but not alternate face-name pairs, was associated with elevated amyloid⁷, hinting that failures in memory consolidation could underlie the association between diminished practice effects and preclinical AD. In related work, diminished memory retention after one week, but not after 30 minutes, was associated with presymptomatic autosomal dominant AD⁸ and APOE ε4⁹ among CU adults. The commonality between these studies is in temporal design, meaning that memory for the same content is retested (with and without re-exposure) after the passage of time. This study design allows for the time-dependent processes involved in memory consolidation to effectively (or ineffectively) unfold over the course of hours¹⁰ at the synaptic level and days to weeks¹¹ at the systems level.

Building on these promising findings, we were interested in capturing memory consolidation using daily testing over one week. Among CU older adults with otherwise normal performance on traditional measures, we sought to determine whether individual differences in learning curves were related to underlying AD pathology as well as predictive of future decline. By capturing learning for 7 consecutive days, we may produce a more psychometrically robust estimate of learning than afforded by a single re-testing session, a critical concern in preclinical AD, where cognitive effects are subtle.

As such, we examined multi-day learning curves (MDLC) using a previously validated web-based platform-the Boston Remote Assessment for Neurocognitive Health (BRANCH)¹². CU older adults with known amyloid status were asked to complete MDLC BRANCH, which includes three tests with identical stimuli presented daily, for 7 days. We hypothesized that diminished learning curves, representing early deficits in memory consolidation, would be associated with elevated amyloid and prospective cognitive decline on standard measures.

METHODS

Participants

164 CU participants ages 60–91 were recruited from three affiliated cohorts: 98 participants from the Harvard Aging Brain Study (HABS; 2P01AG036694–11-Sperling, Johnson) 33 from the Instrumental Activities of Daily Living Study (IADL; R01AG053184-Marshall) and 33 from the Subjective Cognitive Decline study (SCD; 1R01AG058825–01A-Amariglio). Study procedures were conducted in accordance with human subjects’ protections and the study protocol was approved by the Mass General Brigham IRB. All participants underwent informed consent. Exclusion criteria included: history of alcoholism, drug abuse, head trauma, or current serious medical/psychiatric illness. Given that participants were recruited at various stages of their participation in ongoing longitudinal studies with annual assessments, they were classified as CU by either study entry criteria or via a multidisciplinary consensus¹³ depending on which source of information was most proximal to their completion of BRANCH. Study entry criteria included a Clinical Dementia Rating (CDR) global score=0, Mini-Mental Status Exam (MMSE) >25, and Logical Memory Delayed Recall (LMDR) scores above education-adjusted cutoffs (≥9 for 16+ years of education, ≥5 for 8–15 years of education). Participants were brought to multidisciplinary consensus if they had a global CDR≥0.5 and/or performance falling 1.5 standard deviations below the sample mean on any individual domain-specific cognitive composite score; if they did not fall below any of these cutoffs, they were considered CU.¹³ While a subset of participants in the current sample have a CDR=0.5 (n=6), they were retained in the study because they were deemed CU via multidisciplinary consensus.

Standard Clinical and Cognitive Assessments

Participants completed in-clinic cognitive assessments including the Preclinical Alzheimer Cognitive Composite (PACC-5)^14,15, which includes LMDR, Free and Cued Selective Reminding Test (FCSRT), MMSE, Digit Symbol Substitution Test(DSST) and category fluency. Participants also completed the Buschke 6-trial Selective Reminding Test (6-SRT)¹⁶. We examined MDLC BRANCH in relation to the PACC-5 and memory measures including LMDR, FCSRT free recall, and 6-SRT total recall. In-clinic assessments were completed within 97 days of Multi-Day BRANCH. 123 participants were followed annually with the PACC-5 for up to 2 years.

Quantification of Amyloid Burden

Participants underwent Positron Emission Tomography (PET) with ^11CPittsburgh Compound-B (PiB) to quantify amyloid burden¹⁷. PET scans were completed within 1.03(SD=1.36) years of Multi-day BRANCH. PiB images were acquired using a 60-min dynamic acquisition on a Siemens ECAT HR+ PET scanner. PET images were co-registered to corresponding T1 images using Freesurfer-based(v6) structural regions of interest mapped into native PET space using SPM12. PiB is expressed as the distribution volume ratio (DVR), with a cerebellar gray reference region. A global cortical aggregate was calculated for each participant for the target region, comprising frontal, lateral temporal and retrosplenial regions. Participants were dichotomized into low($\text{Aβ}-$) versus high($\text{Aβ}+$) groups (cut-off-1.185)¹⁸.

MDLC Procedure.

MDLC were collected using Multi-Day BRANCH^12,19 which includes three tasks (described below) captured once per day for seven days(Fig 1). Prior to daily testing, participants attested to completing tasks independently without recording stimuli/responses with the goal of advancing research. Participants specified the time of day they preferred to take the test and were notified at that time. Participants were also asked to rate the enjoyability of tasks each day on a Likert scale.

FIGURE 1.

Capturing Learning Curves using the Multi-Day Boston Remote Assessment for Neurocognitive Health (BRANCH)

The Digit Signs Test is modeled on the DSST²⁰, a measure of processing speed with an associative memory component, in that performance is faster if pairs are memorized. Participants are shown a key of 6 street signs paired with digits must indicate “yes” or “no” whether a series of digit-sign pairs are correct. The outcome is number of correct pairs completed within 120 seconds.

The Groceries Test is an associative memory measure²¹. Participants are asked to remember a price paired with a pictured grocery item. Following a delay, participants identify the correct price among counterbalanced incorrectly paired and partially novel price/grocery distractor pairs. The outcome is number of correct responses.

In FNAME⁷, participants are asked to remember a series of face-name pairs. Following a delay, the participant is shown each face and asked to select the first letter of the name paired with that face (first letter name recall). Next, they are asked to identify the correct name via multiple choice (target name, re-paired same-sex name, same-sex foil name) (face-name memory). The outcome is the average number of correct responses for first letter name recall and face-name memory combined.

Computing the MDLC:

To account for the different learning curve shapes (i.e., learning on Digit-Sign is linear whereas learning on FNAME is logarithmic) and the potential impact of ceiling effects (i.e., 23% and 8% of participants performed at ceiling levels-100% accuracy for 2 or more administrations-on the FNAME and Groceries tests respectively), we computed an AUC for each task to capture both the rapidity with which an individual learns as well as the total accumulation of content (explained in detail elsewhere).¹⁹ In contrast with the more common use of an AUC to be used in classification models (e.g., ROC analyses), use of an AUC in the current context allows us to produce a summary metric for the overall proportion of information learned using a general formula from integral calculus. Additionally, to account for an individual’s starting point (Day 1 performance), we computed a scaled AUC which equals AUC/AUC_max where AUC_max is the maximum value of the AUC obtained if the participant scored at the maximum value from the second through the final test administration. AUC values (herein referred to as “learning curves”) were computed for an equally weighted composite across the three tasks as well as for each test.

Statistics

Statistical analyses were completed using R(v4.0.3). Statistical significance was set at p<0.05.

Linear regression analyses were performed to determine the association between learning curves (composite, individual tests) and $\text{Aβ}+/-$ while correcting for demographics (i.e., age, sex, and education). Cohen’s d effect-sizes were used to evaluate the strength of $\text{Aβ}+/-$ group differences using the following benchmarks: small (d=0.2), medium (d=0.5), and large (d=0.8)²². Comparable analyses were performed to determine whether differences between $\text{Aβ}+/-$ groups were similarly observable on Day 1 of Multi-Day BRANCH (both the composite and individual tests) as well as on standard in-clinic measures.

We investigated whether decrements on learning curves were associated with annual change on the PACC5 among those with longitudinal testing. Individual PACC5 slopes were extracted using linear mixed effects models correcting for demographics, and their interaction with time. Pearson’s correlations were used to investigate the association between PACC5 slopes and learning curves.

Sensitivity analyses.

To determine whether a shortened administration schedule could reduce participant burden, we examined whether $\text{Aβ}+/-$ performance differences were retained when reducing testing to 5, 4, or 3 days. Additionally, we repeated the primary analyses excluding top performers (i.e., those who reached ceiling defined as 100% accuracy on two or more days on FNAME and Groceries).

RESULTS

Participant Characteristics

Participants included 164 individuals (mean age=74.3) of whom 36(21.9%) were classified as $\text{Aβ}+$ and 128 classified as $\text{Aβ}-$(Table 1). The $\text{Aβ}+$ group was older but there were otherwise no group differences in terms of sex, years of education, race, MMSE, or global CDR score.

TABLE 1.

Demographic Characteristics

	Overall (n=164)	$\text{Aβ}+$ (n=36)	$\text{Aβ}-$ (n=128)	Difference between $\text{Aβ}+$/$\text{Aβ}-$
Age	74.3 (7.31)	76.9 (5.80)	73.6 (7.55)	t=−2.83, p=0.006 95%CI[−5.66, −0.98]
Global CDR (0, 0.5)	(158, 6)	(34, 2)	(124, 4)	χ2 =0.03, p=0.854
MMSE	28.90 (1.26)	28.5 (1.7)	29.00 (1.10)	t=1.66, p=0.105 95%CI[−0.11,1.10]
LMDR	16.1 (3.87)	15.7(4.81)	16.2 (3.57)	t=0.56, p=0.579 95%CI[−1.25,2.22]
Sex (Female, Male)	(105, 59)	(20, 16)	(85, 43)	χ2 =1.00, p=0.316
Race				χ2 =0.82 p=0.936
Caucasian	146	33	113
Black	12	2	10
Asian	4	1	3
Native American	1	0	1
Other	1	0	1
Years of Education	16.6 (2.45)	16.6 (2.11)	16.7 (2.55)	t=0.11, p=0.914 95%CI[−0.79,0.88]
Continuous Aβ (DVR in Global Composite)	1.16 (0.21)	1.47 (0.26)	1.08 (0.04)	t=−8.95, p<0.001 95%CI[−0.48,−0.30]

Note: Mean and SD are reported unless otherwise noted, CDR=Clinical Dementia Rating, MMSE=Mini Mental Status Exam, LMDR=Logical Memory Delayed Recall

BRANCH Feasibility

Multi-day BRANCH was completed over a mean of 8.1 days (Max=14 days) with no differences between $\text{Aβ}+/-$ groups regarding days to completion (t=−0.46, 95%CI[−1.96,1.23] p=0.647). There were no $\text{Aβ}+/-$ groups differences regarding completion time (t=−0.58, 95%CI[−0.54,0.29] p=0.561) with $\text{Aβ}+$ individuals requiring a mean of 12.23 minutes and $\text{Aβ}-$ participants requiring 12.11 minutes. The majority (n=151, 92%) of participants completed all 7 days. Among the 13 participants with incomplete data, 10 were included in analyses as they completed at least 3 days of assessments, allowing for the computation of a learning curve. There were no differences between completers/non-completers regarding clinical severity on the MMSE (mean difference 1.3 points, p=.08) or global CDR (mean difference 0.02 points, p=.58); however, we did observe a slightly higher proportion of $\text{Aβ}+$ participants among non-completers (χ² =6.48, p=0.011).

Participants reported increasing task enjoyability by day (B=0.50, 95%CI[0.46,0.53] p<0.001). 45.7% of participants completed multi-day BRANCH on a smartphone while the remaining participants used a laptop/desktop. Smartphone users tended to be younger but there were otherwise no other demographic or BRANCH performance differences between groups based on device used.

Performance on Standard Cognitive Assessments

There were no statistically significant differences in performance between $\text{Aβ}+/-$ groups on standard in-clinic cognitive testing (Table 2) including the PACC-5, FCSRT Free Recall, LMDR, or 6-SRT Total Recall, although some measures exhibited trend-level differences (i.e., PACC-5, 6-SRT Total Recall).

TABLE 2.

Linear Regression Models Showing Performance between $\text{Aβ}+$ and $\text{Aβ}-$ Groups on Standard Singe Timepoint Cognitive Assessments

	PACC-5			FCSRT Free Recall			Logical Memory Delayed Recall			6-SRT Total Recall
Predictors	Est.	CI	p	Est.	CI	p	Est.	CI	p	Est.	CI	p
Intercept	0.40	−1.42, 2.21	0.666	2.66	−0.91, 4.40	0.003	−2.97	−4.80, −1.14	0.002	1.02	−0.74, 2.79	0.254
Age	−0.02	−0.05, −0.00	0.038	−0.04	−0.06, −0.02	0.001	0.02	−0.00, 0.04	0.070	−0.02	−0.04, −0.00	0.050
Sex [M]	−0.53	−0.86, −0.20	0.002	−0.53	−0.85, −0.21	0.001	−0.33	−0.66, 0.01	0.056	−0.62	−0.94, −0.30	0.001
Educ.	0.09	0.03, 0.16	0.005	0.03	−0.03, 0.09	0.303	0.09	0.03, 0.16	0.004	0.05	−0.01, 0.11	0.116
Aβ status [+]	−0.29	−0.67, 0.08	0.125	0.03	−0.34, 0.39	0.887	−0.16	−0.54, 0.22	0.417	−0.32	−0.69, 0.05	0.089
R2 / R2 adj	.143 / .122			.166/.145			.084/.061			.152/.130

Note: n=164; PACC5= Preclinical Alzheimer’s Cognitive Composite 5, FCSRT=Free and Cued Selective Reminding Test, 6-SRT= 6-trial Selective Reminding Test, regression coefficients are standardized

Performance on Learning Curves using Multi-Day BRANCH

There were no differences between $\text{Aβ}+/-$ groups when examining composite or individual measure performance on Day 1 of the learning curve (Table 3). However, $\text{Aβ}+$ showed diminished learning curves (composite score) relative to $\text{Aβ}-$ with a medium effect size of Cohen’s d=0.49 (95%CI[0.10,0.87] (Table 3, Fig 2). This pattern of less robust learning among $\text{Aβ}+$ was observed across each individual test (Figure 2). However, the difference between $\text{Aβ}+/-$ was not statistically significant for FNAME whereas it was significantly different for Digit Sign and the Groceries Test (Table 3).

TABLE 3.

Linear Regression Models Showing Performance between $\text{Aβ}+$ and $\text{Aβ}-$ Groups on Day 1 (top) and Learning Curves using the Multi-Day Boston Remote Assessment for Neurocognitive Health (BRANCH) (bottom)

	Composite Day 1			Digit Sign Test Day 1			Groceries Test Day 1			FNAME Day 1
Predictors	Est.	CI	p	Est.	CI	p	Est.	CI	p	Est.	CI	p
Intercept	2.35	0.64, 4.07	<0.001	1.58	0.59, 2.57	0.002	−0.46	−1.80, 0.87	0.496	−0.60	−1.69, 0.48	0.272
Age	−0.04	−0.06,−0.02	<0.001	−0.04	−0.05, −0.03	<0.001	−0.00	−0.02,−0.01	0.566	−0.02	−0.03 −0.00	0.007
Sex [M]	−0.45	−0.76,−0.13	0.005	0.05	−0.14, 0.23	0.616	−0.17	−0.42, 0.07	0.165	−0.39	−0.59,−0.19	<0.001
Educ.	0.05	−0.01, 0.11	0.132	0.04	0.00, 0.07	0.033	0.00	−0.05, 0.05	0.983	0.04	−0.00, 0.07	0.062
Aβ status [+]	−0.17	−0.53, 0.19	0.350	−0.14	−0.34, 0.07	0.194	−0.09	−0.37, 0.19	0.531	−0.01	−0.23, 0.22	0.963
R2 / R2 adj	.163 / .142			.240 / .221			.022/.003			.153 / .131
	Composite Multi-Day Learning Curve			Digit Sign Test Multi-Day Learning Curve			Groceries Test Multi-Day Learning Curve			FNAME Multi-Day Learning Curve
Predictors	Est.	CI	p	Est.	CI	p	Est.	CI	p	Est.	CI	p
Intercept	3.18	1.61, 4.76	<0.001	4.46	2.91, 6.01	<0.001	2.10	0.47, 3.74	0.012	1.69	0.03, 3.36	0.046
Age	−0.05	−0.07, −0.03	<0.001	−0.07	−0.09, −0.05	<0.001	−0.03	−0.05, −0.01	0.001	−0.03	−0.05, −0.01	0.007
Sex [M]	−0.66	−0.95, −0.37	<0.001	0.07	−0.22, 0.35	0.649	−0.69	−0.99, −0.39	<0.001	−0.88	−1.19, −0.58	<0.001
Educ.	0.06	0.00, 0.11	0.045	0.05	−0.00, 0.11	0.065	0.05	−0.01, 0.10	0.107	0.04	−0.02, 0.10	0.153
Aβ status [+]	−0.36	−0.69, −0.03	0.034	−0.37	−0.70, −0.04	0.028	−0.43	−0.78, −0.09	0.015	−0.13	−0.48, 0.22	0.478
R2 / R2 adj	.323 / .305			.310 / .293			.255 / .236			.248/.229

Note: n=164; FNAME=Face-Name Associative Memory Exam, regression coefficients are standardized

FIGURE 2.

Decrements in Learning Curves (Multi-Day BRANCH) Between $\text{Aβ}+$ and $\text{Aβ}-$ Cognitively Unimpaired Older Adults

Note: n=164; FNAME=Face-Name Associative Memory Exam, raw (uncorrected) data is shown, y axis refers to percent of items correct

Systematically reducing the number of days over which learning curves were computed, $\text{Aβ}+/-$ group differences emerged on the fourth testing day (Table 4) indicating four administrations were needed to observe group differences. Overall results were comparable when top performers were removed (Table 4).

TABLE 4.

Sensitivity Analysis: Reducing the Multi-Day BRANCH Learning Curve to 5, 4, and 3 days (A) and Removing Top Performers (B)

A
	3-day Multi-Day BRANCH Composite			4-day Multi-Day BRANCH Composite			5-day Multi-Day BRANCH Composite
Predictors	Estimate	CI	p	Estimate	CI	P	Estimate	CI	p
(Intercept)	0.05	−0.13 – 0.23	0.578	0.09	−0.09 – 0.27	0.311	0.09	−0.08 – 0.27	0.297
Aβ Group [$\text{Aβ}+$]	−0.33	−0.74 – 0.08	0.110	−0.59	−0.99 – −0.20	0.003	−0.58	−0.97 – −0.20	0.003
R2 / R2 adjusted	0.017 / 0.010			0.053 / 0.047			0.054 / 0.048
	3-day Multi-Day BRANCH Composite			4-day Multi-Day BRANCH Composite			5-day Multi-Day BRANCH Composite
Predictors	Estimates	CI	p	Estimates	CI	p	Estimates	CI	p
(Intercept)	3.09	1.43 – 4.75	<0.001	2.92	1.30 – 4.54	<0.001	2.95	1.35 – 4.55	<0.001
Age	−0.05	−0.07– −0.03	<0.001	−0.05	−0.07 – −0.03	<0.001	−0.05	−0.07 – −0.03	<0.001
Sex [M]	−0.58	−0.89 – −0.28	<0.001	−0.67	−0.97 – −0.38	<0.001	−0.68	−0.98 – −0.39	<0.001
Education	0.05	−0.01 – 0.11	0.118	0.05	−0.00 – 0.11	0.071	0.06	0.00 – 0.11	0.046
Aβ Group [$\text{Aβ}+$]	−0.11	−0.48 – 0.25	0.539	−0.35	−0.70 – −0.00	0.048	−0.34	−0.67 – 0.00	0.052
R2 / R2 adjusted	0.246 / 0.226			0.292 / 0.274			0.306 / 0.288

B
	Composite Multi-Day Learning Curve			Digit Sign Test Multi-Day Learning Curve			Groceries Test Multi-Day Learning Curve			FNAME Multi-Day Learning Curve
Predictors	Est.	CI	p	Est.	CI	p	Est.	CI	p	Est.	CI	p
Intercept	2.63	.91, 4.36	0.003	4.46	2.91, 6.01	<0.001	2.02	.46, 3.57	0.012	1.77	−.16, 3.70	0.072
Age	−0.04	−.06, −.02	<0.001	−0.07	−.09, −.05	<0.001	−0.03	−.05, −.01	0.001	−.02	−.05, −.00	0.030
Sex [M]	−0.51	−.82, −.20	0.002	0.07	−.22, .35	0.649	−0.62	−.90, −.33	<0.001	−.75	−1.09, −.40	<0.001
Educ.	0.04	−.03, .10	0.244	0.05	−.00, .11	0.065	0.03	−.03, .08	0.327	.01	−.06, .08	0.837
Aβ status [+]	−0.40	−.76, −.05	0.027	−0.37	−.70, −.04	0.028	−0.36	−.69, −.03	0.031	−.11	−.51, .28	0.570
Observations	116			164			122			148
R2 / R2 adj	.300/ .274			.310 / .293			.209/ .182			.243/.221

Note: n=164; regression coefficients are standardized

Associations between Learning Curves and annual change on the PACC-5

123 participants completed at least 1 follow-up in-clinic cognitive assessment (mean age 73.6 (SD=7.22), 65% female, mean years of education 16.5 (SD=2.53), 20% $\text{Aβ}+$, mean follow-up=1.07 years (Max=2 years)). A diminished learning curve at baseline was associated with greater subsequent annual decline on the PACC-5 (r=.54, 95%[0.40, 0.65], p<.001)(Fig 3). Weaker associations were observed between annual PACC-5 change and either BRANCH Day 1 (r=0.36, 95%CI[0.19, 0.50], p<.001) or baseline PACC-5 (r=0.39, 95%CI[0.22, 0.55], p<.001).

FIGURE 3.

Association between a diminished learning curve (Multi-Day BRANCH) collected over 7 days and subsequent cognitive decline (PACC-5) over 1.07 years of follow-up

Note: n=123

DISCUSSION

Our findings suggest that assessing learning over repeated evaluations can reveal disease-relevant decrements in memory during preclinical AD that are less readily observed using single timepoint assessments. More specifically, we found that a diminished learning curve for the same information presented for 7 days (12 minutes/day of unsupervised web-based testing on personal devices) was associated with elevated amyloid whereas this association was absent using single timepoint (Day 1) or standard cognitive assessments. This same pattern was observed across three different memory tasks (with the strongest effects on Digit Signs and Groceries Tests) and was associated with longitudinal decline on standard assessments, suggesting that a diminished MDLC may be a harbinger of decline in the setting of otherwise normal cognitive performance. These results provide insights into the nature of the earliest detectable memory decrements in AD. That is, decrements in memory consolidation, made measurable by assessing learning over repeated daily evaluations, may predate other conventional deficits in learning and recall. Furthermore, our results provide a practical paradigm for identifying those with preclinical AD at greatest risk for short-term decline, individuals who would be ideal candidates for interventions.

Multi-day learning curves and memory consolidation

Failures in initial memory encoding differentiate individuals with dementia versus MCI²³, whereas failures in retention after a 30 minute delay distinguish between MCI and normal aging²⁴. Our results suggest that failures in memory consolidation rather than initial encoding or retention may be the earliest observable memory changes in preclinical AD given that $\text{Aβ}-$related decrements in memory were not observed on standard (i.e., LMDR, FCSRT) or challenging (i.e., 6-SRT) learning and recall tasks administered at a single timepoint. Instead, differentiation between $\text{Aβ}+/-$ only became evident with repeated daily testing. Because memory consolidation is a process that requires the passage of time²⁵, it is unable to be probed with a single timepoint assessment, regardless of the sensitivity of the measure.

There are multiple avenues by which Aβ and related AD pathology may have both direct and indirect effects on memory consolidation. Aβ is known to impair synaptic plasticity²⁶, which is critical to learning. For example, in rodent models, injection of human brain-derived Aβ into rats after avoidance training impaired recall of that training at 48 hours²⁷. Interestingly, the adverse effect of Aβ on recall was not significant at 24 hours and required a re-exposure to the training condition, suggesting that Aβ did not obliterate the memory, but may instead have made it more susceptible to decay or extinction. We similarly observed that significant trends towards differences in learning curves between $\text{Aβ}+/-$ groups emerged after several days.

In addition to synaptic plasticity, it may also be that $\text{Aβ}-$related decrements in learning becomes more evident with re-learning over days because of the critical role of sleep for memory consolidation. Sleep is disrupted in symptomatic AD with recent work suggesting sleep changes may be occurring in presymptomatic stages with potentially bidirectional effects. That is, Aβ may reduce sleep quality with reduced sleep quality, in turn, accelerating Aβ accumulation²⁸. An important study showed that elevated Aβ among normal older adults was associated with diminished slow wave activity during NREM sleep (a stage of sleep critical for memory consolidation) and that the extent of this reduced activity was associated with worse overnight memory retention²⁹.

Finally, tau deposition has a more anatomically direct impact on memory versus Aβ, with tau tangles impacting the integrity of MTL functioning. With ongoing acquisition of tau PET in this sample, we will, in future studies, be able to determine whether memory consolidation deficits, measured with diminished learning curves, are the first $\text{Aβ}-$related sign of impaired synaptic function that portends a tau-associated amnestic syndrome typical of symptomatic AD³⁰.

Multi-day learning curves: individual test results

Interestingly, $\text{Aβ}+/-$ effects were most robust on a speeded measure with an associative memory component (Digit-Signs) versus on frank associative memory measures (FNAME, Groceries). We hypothesize that the absence of a performance ceiling on Digit-Signs, in contrast with FNAME/Groceries, where participants may be able to achieve 100% accuracy, may have conferred measurement benefits. Aligned with this, removing top performers in our sensitivity analyses improved estimates. Alternatively, recent work has also shown that declines in processing speed may be a very early indicator when amyloid is accumulating, with associative memory processes impacted at later stages of tau accumulation.³¹ Thus, Digit-Signs may be particularly sensitive because it comes a speeded task with an associative memory component. Finally, some of our participants may have become habituated to the FNAME test, as many have encountered variations of FNAME during other parent-study related assessments, and thus may have developed their own unique strategies. Regardless, the similar pattern of findings across individual tests indicate that repeated memory testing reveals vulnerabilities in the ability to acquire specific episodic content among $\text{Aβ}+$ individuals.

Multi-day learning curves, practice effects and other repeated and digital assessments

Our findings are also relevant to a larger body of work showing diminished practice effects signal cognitive vulnerability⁵. While clinical trials actively eschew practice effects, as they can impede the ability to capture cognitive decline,, the MDLC capitalizes on factors known to improve performance including shorter retest intervals³², focus on memory³³, and, most importantly, repeated stimuli⁷. While practice effects can be related to both familiarity with test strategies and familiarity with stimuli, it seems that the latter may be most relevant for AD. For example, the present study was motivated, in part, by our previous findings which showed that decrements in learning of the same face-name pairs, but not alternate face-name pairs, on monthly testing was associated with $\text{Aβ}+$⁷. Other experimental psychology paradigms, which leverage repeated stimuli, have reported similar results³⁴. These findings are in keeping with the notion that failures on repeated testing with the same stimuli hint at taxed memory consolidation processes.

BRANCH is one among several digital cognitive assessments whose development has been accelerated by several converging factors including the necessity for remote assessment resulting from social distancing requirements of COVID-19, a need for scalable assessment approaches in the context of early detection and preclinical AD, and the increasingly widespread uptake of digital devices among older adults. Recent studies using web/app based cognitive testing with older adults have shown that they are highly feasible^35–37 and that they exhibit high concordance with in-clinic cognitive testing, ^12,36,38 with several showing correlations with AD biomarkers (see Öhman et al 2021 for a review).³⁹ These digital assessments vary in approach with some offering more traditional measures in a digital format, others emphasizing the interrogation of specific cognitive processes, and many leveraging the digital format for repeated testing.

Notably, the MDLC differs from other types of repeated sampling of cognition such as “ecological momentary assessment” and “burst design.” In these paradigms, cognition is sampled multiple times within and across days and averaged in order to achieve a more robust and stable estimate⁴⁰. Reducing measurement variability may thus improve the detection of change over time whereas the MDLC leverages improvement in performance day by day.

Associations between diminished learning curves and prospective cognitive decline

Generally, cross-sectional cognitive performance, even in the context of AD biomarker positivity, is not a robust predictor of risk for short-term clinical progression, particularly in preclinical stages of disease where rates of decline are highly heterogenous. Our findings of a moderate association between MDLC and prospective PACC-5 decline are aligned with other studies showing the predictive utility of diminished practice effects to risk for cognitive decline^41,42. Cognitive paradigms which can help identify CU at greatest risk for short-term cognitive progression (and those most eligible for secondary prevention/treatment) will be critical as more accessible and cost-effective AD biomarkers (i.e., blood-based biomarkers) make identification of preclinical AD more widespread⁴³. However, replication in much larger and more representative samples is needed to determine whether pairing diminished MDLC with biomarker positivity amongst CU can improve the prediction of risk for imminent clinical progression at an individual level.

Limitations and Future Directions

The racial breakdown (~10.4% from under-represented groups) as well as the high education level of our sample is not aligned with the demographics of broader US population at-risk for AD^44,45. Additionally, because assessments are completed unsupervised, there are factors that cannot be experimentally controlled such as the presence of interruptions or unauthorized use of memory aides. While it is possible to counter this with technology (i.e., recording the participant taking the test via web-video), the benefits of these strategies are outweighed by data privacy concerns. To counter this challenge, we asked participants to attest to completing the task unaided each day for the sake of advancing research. Despite this attestation, it is possible that some participants used memory aides. However, in a sensitivity analysis, removal of top performers (i.e., those who hit ceiling) on FNAME/Groceries did not alter our findings. Another limitation of this study is the relatively small sample size which may have left us underpowered to observe differences in performance between $\text{Aβ}+/-$ groups on traditional measures. In addition to obtaining a more racially and ethnically diverse sample, it will be critical to increase the sample size to determine the extent to which this type of paradigm may be informative at the individual level for diagnosis and prognosis.

In addition to assessing MDLC in relation to tau, future work will determine whether repeating the MDLC over time with new sets of stimuli provides a more sensitive marker to track cognitive decline. For example, it may be that this paradigm is useful in detecting a change in synaptic integrity in response to a therapeutic over a shorter interval, facilitating brief, adaptive, and early phase clinical trials.

Conclusions

Consistent with many other cross-sectional studies in preclinical AD^1,2, standard single timepoint cognitive assessments did not differentiate between $\text{Aβ}+/-$ groups. Observing associations between biomarkers and cognition in preclinical AD is a methodological challenge given that, by restricting samples to those with a minimum level of cognitive performance, we both exclude individuals whose underlying biomarker positivity is having a demonstrable clinical impact as well as limit our ability to observe correlations given a restricted range of possible cognitive scores. However, our findings that $\text{Aβ}+$ participants exhibited diminished learning curves over seven days of testing, alongside other recent work,^7,9,34 suggest that $\text{Aβ}+/-$ differences are demonstrable with non-standard single timepoint assessments. Importantly, these results suggests that failures in memory consolidation predate other conventional amnestic deficits in AD. Implementing these promising assessments into future research may improve early detection and tracking of cognitive decline in preclinical AD.

Summary for Social Media.

1. If you and/or a co-author has a Twitter handle that you would like to be tagged, please enter it here. (format: @AUTHORSHANDLE) @PappKate @rivkae @BRANCH_Harvard

2. What is the current knowledge on the topic? Early detection of subtle cognitive decline in Alzheimer’s disease is challenging to measure.

3. What question did this study address? Can multi-day cognitive paradigms reveal previously undetectable memory decrements related to preclinical Alzheimer’s disease?

4. What does this study add to our knowledge? Otherwise difficult to detect memory decrements are revealed by assessing learning over days, suggesting subtle failures in memory consolidation predate other memory changes at the earliest stages of Alzheimer’s disease.

5. How might this potentially impact on the practice of neurology? (one to two sentences) In the future, neurologists may be able to monitor the cognition of their patients remotely and more frequently, facilitating earlier diagnosis and tracking of treatment responsiveness.

Data Availability

To use BRANCH in academic research, please contact the corresponding author.