The Relationship Between Learning Slopes and Alzheimer’s Disease Biomarkers in Cognitively Unimpaired Participants with and without Subjective Memory Concerns

Abstract

Objective:

Learning slopes represent serial acquisition of information during list-learning tasks. Although several calculations for learning slopes exist, the Learning Ratio (LR) has recently demonstrated the highest sensitivity towards changes in cognition and Alzheimer’s disease (AD) biomarkers. However, investigation of learning slopes in cognitively unimpaired individuals with subjective memory concerns (SMC) has been limited. The current study examines the association of learning slopes to SMC, and the role of SMC in the relationship between learning slopes and AD biomarkers in cognitively unimpaired individuals.

Method:

Data from 950 cognitively unimpaired participants from the Alzheimer’s Disease Neuroimaging Initiative (aged 55 to 89) were used to calculate learning slope metrics. Learning slopes among those with and without SMC were compared with demographic correction, and the relationships of learning slopes with AD biomarkers of bilateral hippocampal volume and β-amyloid pathology were determined.

Results:

Learning slopes were consistently predictive of hippocampal atrophy and β-amyloid deposition. Results were heightened for LR relative to the other learning slopes. Additionally, interaction analyses revealed different associations between learning slopes and hippocampal volume as a function of SMC status.

Conclusions:

Learning slopes appear to be sensitive to SMC and AD biomarkers, with SMC status influencing the relationship in cognitively unimpaired participants. These findings advance our knowledge of SMC, and suggest that LR – in particular – can be an important tool for the detection of AD pathology in both SMC and in AD clinical trials.

INTRODUCTION

Subjective memory concerns (SMC; Steinberg et al., 2013) refer to self- or collateral-reported concerns of memory difficulties. SMC may exist either with or without objective neuropsychological deficits, and individuals who are cognitively unimpaired but reporting SMC have received increased research focus. For example, while not reaching thresholds for objective cognitive impairment necessary for a diagnosis of mild cognitive impairment (MCI; Albert et al., 2011), cognitively unimpaired individuals with SMC display worse memory and cognitive functioning (Park et al., 2019; Sohrabi et al., 2019), and worse daily functioning (Ryu et al., 2016), than their peers without SMC. Similarly, cognitively unimpaired individuals with SMC are reported to have higher amyloid burden (Miebach et al., 2019), higher tau burden (Buckley et al., 2017), and smaller hippocampal volumes (Liang et al., 2020) compared to those without memory concerns. In addition, longitudinal studies show that individuals with SMC are at risk of developing objective cognitive deficit and a subsequent diagnosis of MCI and/or dementia (Mitchell et al., 2014; Ronnlund et al., 2015). Therefore, SMC may be regarded as the first clinical manifestation along the Alzheimer’s disease (AD) continuum (Kiselica, 2021; Kiselica et al., 2021), and has been conceptualized as being Stage 2 of the recently developed National Institute of Aging (NIA) – Alzheimer’s Association (AA) Research Framework for the diagnosis of AD (“ATN model”; Jack et al., 2018). However, individuals with SMC are a heterogenous group with various etiologies, and further studies are needed to examine aspects of memory that are sensitive to SMC with underlying AD pathology.

Learning slopes reflect serial acquisition of information during list-learning tasks (Gifford et al., 2015; Hammers et al., 2021). They are associated with encoding abilities, and facilitate improved retention of incoming information (Hammers, Spencer, et al., 2022). While several calculations for learning slopes exist, the Learning Ratio (LR) has recently had the highest sensitivity towards changes in cognition and AD biomarkers such as amyloid, tau, and hippocampal atrophy (Hammers, Suhrie, Dixon, Gradwohl, Archibald, et al., 2021). In fact, LR has recently been shown to be successful in discriminating among participants within the NIA-AA Research Framework (A+T+ vs. A+T− vs. A-T−; Hammers, Kostadinova, et al., 2022). As will be explained in more detail in the Methods, LR represents the proportion of information acquired after the first trial of a multiple-trial learning test, relative to the amount of information remaining to learn after trial 1 (Spencer et al., 2020). For example, if an individual learns 3 words on the first trial of a 12-word list learning task and 9 words on a later trial, they have learned 6 additional words out of a possible 12 additional words. LR would therefore be 0.50, or more commonly represented as 50% (Spencer et al., 2020). This calculation is in contrast to the more traditionally examined learning slope calculation (the Raw Learning Score or RLS; Benedict, 1997) that reflects the difference in performance between trial 1 and the final trial – regardless of the amount of information left to learn after trial 1. Although research into LR has been promising, investigation of learning slopes in cognitively unimpaired individuals with SMC is limited.

The first aim of the study is to examine whether individuals with SMC have deficiencies in learning slope. This is particularly relevant because recent research has suggested that SMC is associated with both temporal (Striepens et al., 2010) and frontal lobe-mediated processes (Garrido-Chaves et al., 2021). Owing to the necessary incorporation of several cognitive demands – including episodic-memory and working memory/attention (Gifford et al., 2015) – in the acquisition of information, the literature has identified that the hippocampus (Hammers, Suhrie, Dixon, Gradwohl, Archibald, et al., 2021), ventrolateral prefrontal (Gifford et al., 2015), and dorsolateral prefrontal lobes (D’Esposito et al., 1999) have involvements in learning slope functioning. As such, patients with SMC may be particularly susceptible to deficiencies in learning slope. Additionally, recent findings (Thomas et al., 2020; Thomas et al., 2018) have suggested that learning process scores – including the learning slope calculation “Learning Over Trials” (LOT) that will be explained in greater detail in the Methods – can identify participants with subtle cognitive decline. Subtle cognitive decline is a similar – but distinct – condition along the early continuum of the NIA-AA Research Framework whereby patients experience very slight changes to their cognitive status, which is not of the magnitude required to necessitate “impairment” (Jessen et al., 2020). Consequently, we hypothesized that among cognitively unimpaired participants, learning slope performances derived from the Rey Auditory Verbal Learning Test (RAVLT; Schmidt, 1996) and the Word Recall subtest from the Alzheimer’s Disease Assessment Scale – Cognitive Subscale (ADAS-Cog; Rosen et al., 1984) will be predictive of SMC status. The second aim is to look at the role of SMC in the relationship between learning slopes and AD biomarkers. It is anticipated that poor performance in learning slopes, especially in LR, will be more tightly associated with greater amyloid deposition or smaller hippocampal volume in the presence of SMC.

Overall, should our hypotheses be correct, our results would provide documentation that learning slopes are sensitive to subtle decline in participants with memory concerns who have not yet demonstrated objective impairments in traditional testing. This has applications to diagnosis and decision-making in the clinic (and research studies) as learning slopes may identify subtle changes in cognition before normative-based cognitive testing identifies impairment. Additionally, this would highlight the importance of SMC in the relationship between learning and AD biomarkers, and may shed light on the utility of learning slope measurement to capture SMC with underlying AD etiology in clinical trials of preclinical AD.

METHODS

Participant data for the current project from ADNI’s multi-center longitudinal study (http://adni.loni.usc.edu). ADNI (ADNI1, 2004; Weiner et al., 2017) was originally launched as a public-private partnership in 2003, with extensions (ADNIGO, 2009) and renewals occurring over time (ADNI2, 2011; ADNI3, 2016). The primary scientific goals of ADNI include examining progression of MCI and early AD using magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessments. Please see www.adni-info.org for up-to-date information. Written informed consent was obtained from study participants or authorized representatives, and Institutional Review Board approval has been obtained for each multi-center site within the ADNI consortium.

Data were available for 2366 ADNI participants enrolled across all protocols as of January 7^th, 2022. Participant data collection began in 2005 with enrolled participants being followed for up to 180 months. ADNI inclusion criteria include: baseline age between 55 and 90; fluency in either English or Spanish; the availability of a reliable study partner; possessing ≥6 years of education; the absence of significant head trauma, neurologic disease, or depression; and stability on allowed medications. For the current study, 546 participants were excluded for not having β-amyloid data at their baseline visit. Additionally, 870 participants were excluded for not meeting cognitively unimpaired diagnostic criteria using actuarial methods (Bondi et al., 2014; Jak et al., 2009) – this may have been due to having missing cognitive data necessary to make a diagnosis, or because of actuarial diagnosis of MCI or dementia (as will be described later in the Methods). Consequently, 950 participants were included in the present study. Please see Figure 1 for a schematic representation of the current study’s participant utilization from ADNI.

Figure 1.

Flow diagram of participants recruited into the current study from the total sample of ADNI participants.

Actuarial Diagnostic Classification and Subjective Memory Concern Status

Classification of participants as cognitively unimpaired was undertaken using a modified version of Jak/Bondi and colleagues’ (Bondi et al., 2014; Jak et al., 2009) actuarial model of diagnosis for MCI. This method was used because ADNI’s diagnostic classification has been found to over-pathologize age-related memory changes (Duff, 2021). To apply this classification model, raw scores on domains of memory, language, and executive functioning were converted to age-, education-, and sex-adjusted normative scores using published normative data from the National Alzheimer’s Coordinating Center neuropsychological battery (Shirk et al., 2011; Weintraub et al., 2018). The memory domain included Logical Memory (LM) I and II (“Story A”) from the Wechsler Memory Scale – Revised (WMS-R; Wechsler, 1987), the language domain included Category Fluency – Animals (Morris et al., 1989), and either the Boston Naming Test (Kaplan et al., 1983) or Multi-Lingual Naming Test (MINT; Gollan et al., 2012), and the executive functioning domain included Trail-Making Test Parts A and B (Reitan, 1992). In our study, participants with an ADNI diagnosis of normal cognition or MCI were classified as having an actuarial diagnosis of MCI if the following criteria were met: the presence of (1) impaired scores (>1 SD below the normative mean) on both measures within one or more cognitive domains (i.e., memory, language, or executive function); (2) one impaired score (>1 SD below the normative mean) in all three cognitive domains; or (3) a score on the Functional Activity Questionnaire (FAQ; Pfeffer et al., 1982) ≥ 9. If no actuarial criteria were met, then the participants were classified as being cognitively unimpaired. Using these criteria, 950 participants were classified as being cognitively unimpaired based on actuarial criteria. Note that no RAVLT or ADAS-Cog performances contributed to actuarial classification.

SMC status in ADNI was based on participants and caregivers being asked at baseline if the participant displayed memory difficulties – regardless of the actual objective results of testing. Specifically, ADNI protocols include a line to be indicated as “Yes/No” by the rater pertaining to the participant having “a memory complaint by subject or study partner that is verified by a study partner”. Of the 950 cognitively unimpaired participants included in the study, 364 denied subjective memory concerns – defined here as being Cognitively Normal (CN) – and 586 expressed subjective memory concerns (constituting the SMC group). As will be described in greater detail below, data from inventories of subjective cognitive complaints are included to ensure appropriateness of the categorization of SMC group status in ADNI.

Procedure

Neuropsychological and Behavioral Measures

All participants underwent a clinical and neuropsychological battery at a baseline visit consistent with ADNI protocols. The following neuropsychological and clinical measures were relevant for the current study:

• RAVLT is a 15-item auditorily-presented verbal list-learning task that is presented across five trials. Total Recall is the total number of words accurately recalled across trials 1-5 (score range=0-75), and Delayed Recall is the number of words accurately recalled after approximately 20 minutes of delay (score range=0-15). For both Total and Delayed Recall, higher raw values indicate better performance. Individual trial raw data were used to calculate learning slope performances, as described below.

• LM I and II from the WMS-R assess auditory memory for short stories. LM I examines immediate memory, and LM II examines recall after a 20-30 minute delay. Note that ADNI-3 protocol administers “Story A” only, resulting in the scoring range being a raw score of 0 – 23 for both LM I and II. Higher values indicate better performance.

• Word Recall from the ADAS-Cog is a 10-item auditorily-presented verbal list-learning task that is presented across three trials. Immediate Recall in the current study is calculated as the number of words accurately recalled across trials (score range=0-30), and Delayed Recall score is the number of accurately recalled words following a 10-minute delay (score range=0-10). This scoring reflects a deviation from test developer’s protocols for “Question 1” and “Question 4” of the ADAS-Cog, which allows permits higher values to indicate better performance. The current scoring system has been used in previous research on subtests of the ADAS-Cog (Hammers et al., 2022), and results in immediate and delayed raw scores that are in the same direction as the other memory measures in the study.

• Additional neuropsychological test measures used in ADNI-3 that are relevant to the current study will not be described here given their common knowledge with most dementia clinicians and researchers. Readers are encouraged to review ADNI protocols for descriptions and psychometric properties of these neuropsychological measures. Additional measures include American National Adult Reading Test (AMNART; Grober & Sliwinski, 1991), MoCA (Nasreddine et al., 2005), the 15-item Geriatric Depression Scale (GDS; Sheikh & Yesavage, 1986), the Everyday Cognition – Patient (ECogPT) and – Informant (ECogINF) Memory Index and Total score (Farias et al., 2008), and the Cognitive Change Index (CCI) Memory Index (first 12 questions) and Total score (Rattanabannakit et al., 2016). As described above, Category Fluency – Animal Fluency, Boston Naming Test, MINT, Trail Making Test Parts A and B, and FAQ were used in the application of the actuarial diagnosis model. Higher scores indicate better performance for all tasks except Trail Making Test Parts A and B, GDS (cutoff for depression >5), ECogPT, ECogINF, and CCI.

AD Biomarkers

The classification of bilateral hippocampal volume and β-amyloid positivity in the current study was conducted following standardized ADNI protocols (ADNI1, 2004; ADNI2, 2011; ADNI3, 2016; ADNIGO, 2009) from the participant’s baseline visit. The following procedures were undertaken:

• Bilateral hippocampal volume (mm³) measurements were obtained by Magnetic Resonance Imaging (MRI). Briefly, two T1-weighted 1.5T scans were obtained for each participant, with the ADNI MRI Core (Mayo Clinic) selecting the highest quality scan for incorporation into the study. Scans were collected using sagittal volumetric magnetization-prepared rapid gradient echo sequencing (3D MP RAGE), using a number of fixed parameters. Processing was conducted via FreeSurfer versions 4.3 (ADNI), 5.1 (ADNI-2 and ADNI-GO), and 6.0 (ADNI-3). Additional ADNI MRI protocol information is available at http://adni.loni.usc.edu/data-samples/mri.

• β-amyloid positivity status was obtained from amyloid-Positron Emission Tomography (PET) and lumbar puncture for cerebral spinal fluid (CSF). Specifically, available amyloid-PET positivity results using radioligands ¹¹C-labeled Pittsburgh Compound B (whole brain standardized uptake value ratio [SUVR] cutoff ≥1.50 normalized to whole cerebellum), ¹⁸F-Florbetapir (SUVR ≥1.11), or ¹⁸F-Florbetaben (≥1.08) were prioritized. Amyloid-PET results were available for 809 of the 950 participants in the current study. For the remaining participants, β-amyloid peptide 42 (Aβ42) positivity was used from CSF using Elecsys immunoassays utilizing the fully automated Roche cobas e 601 analyzer (cutoff <1098 pg/mL; Schindler et al., 2018); Roche Diagnostics, Indianapolis, IN). Please see ADNI protocols for greater details about amyloid-PET and amyloid-CSF methods in ADNI, https://adni.loni.usc.edu/data-samples/data-types/pet/.

Calculation of Learning Slopes

Learning slopes were derived from performance on learning trials of both the RAVLT and Word Recall from the ADAS-Cog. See algorithms below for the mathematical descriptions of the learning slope metrics. RLS scores were computed as the highest number of items learned on trials 2 through the final trial, relative to Trial 1 (Benedict, 1997; Brandt & Benedict, 1997). Learning Over Trials (LOT) scores were calculated as the sum of Trials 1 through the final trial, minus the value of Trial 1 multiplied by the number of trials (Morrison et al., 2018). Please note that there were five trials presented for the RAVLT, and three trials presented for Word Recall. LR scores were calculated as the difference in performance between the highest trial score [of Trials 2 through the final trial] and Trial 1 in the numerator, and the difference between the total points available for a trial and Trial 1 performance in the denominator (Spencer et al., 2020). The “Total Points Available for a Trial” for RAVLT is 15 and Word Recall subtest is 10. The specific algorithms are as follows:

$$\mathbf{RLS}=\left(\text{Highest performance on Trials 2 through Final Trial}-\text{Trial 1 performance}\right)$$

$$\mathbf{LOT}=\left(\text{Sum of Trials 1 through Final Trial}\right)-\left(\text{Number of Trials * Trial 1 performance}\right)$$

$$\mathbf{LR}=\frac{\left(\text{Highest performance on Trials 2 through Final Trial}-\text{Trial 1 performance}\right)}{\left(\text{Total Points Available for a Trial}-\text{Trial 1 performance}\right)}$$

Data Analysis

For the primary analyses examining whether learning slopes can predict SMC status, a series of hierarchical linear regression analyses were conducted with demographic variables in Step 1 and the respective learning slope in Step 2. Similar analyses were conducted for traditional measures of learning and memory. Akaike information criterion (AIC) and Bayesian information criterion (BIC) values are reported for regression models to facilitate goodness-of-fit comparisons across models, with lower values indicating a better model fit. To determine the appropriateness of covariates in these analyses, independent samples t tests were conducted between continuous demographic variables (e.g., age, education) and SMC group, and chi-square analyses were conducted between categorical demographic variables (e.g., sex, and ethnicity) and group.

For the examination of interaction between learning slope performance and SMC group on AD biomarkers (bilateral hippocampal volume and β-amyloid positivity), a modified series of hierarchical linear regression analyses were conducted with demographic variables, SMC status, and learning slope metric in Step 1. Step 2 included the interaction between SMC status and learning slope – following centering of the learning slopes. Additionally, to ensure appropriateness of the categorization of SMC group status in ADNI, supplemental analyses were undertaken to examine (1) differences between groups for ADNI-administered measures of subjective cognitive complaints (ECog and CCI) using independent samples t tests, and (2) comparisons between learning slope performance and ADNI-administered subjective cognitive complaints (ECog and CCI) when pooled across groups using partial linear regression.

Measures of effect size were expressed as Cohen’s d (t tests), r² values (hierarchical linear regression), and φ (Phi; chi-square). To protect against multiple comparisons, a Holm-Bonferroni method of adjustment of the two-tailed alpha level was undertaken for all primary analyses.

RESULTS

Demographics

This study included 950 cognitively unimpaired participants, classified as CN (n=364) or SMC (n=586) participants (Table 1). The mean age was 71.49 (SD=6.7) years old, with the total sample having an average of 16.41 (SD=2.5) years of education. Age was different between the two groups (p<.001, d=0.25), with the SMC group being younger than the CN group. Slightly over half the sample was female (54.7%), and most were non-Hispanic white participants (88.7%). No differences existed between groups for education (p=.53, d=0.06), sex (p=.95, φ=−.003), or ethnicity (p=.09, φ=.06). The sample had a high level of estimated verbal intelligence (M=120.13; SD=8.1), and low symptoms of self-reported depression on the GDS (M=1.09; SD=1.3). Additionally, differences were observed between groups for global cognitive status, based on the MoCA (p<.001, d=0.32), with the SMC group performing worse than the CN group. Further, when comparing our groups on endorsements of cognitive concerns on the ECog (Patient and Informant versions) and the CCI, the SMC group possessed significantly higher scores (reflecting greater endorsement of concerns) on all measures – including ECog-PT Memory Index, ECog-PT Total, ECog-INF Memory Index, ECog-INF Total, CCI-Memory, and CCI-Total (ps<.001) – with effect sizes ranging from d=0.78 to 1.53 for the memory-specific indexes and d=0.66 to 1.43 for the total scores. Expanded statistics are available from the first author upon request.

Table 1.

Demographic, neuropsychological, and behavioral variables for the diagnostic groups and total sample. All values are Mean (Standard Deviation; range) unless listed otherwise.

Variable	Cognitively Normal	Subjective Memory Concerns	Total Sample
n	364	586	950
Age (years) 1	72.52 (6.3)	70.85 (6.9)	71.49 (6.7)
Education (years)	16.48 (2.6)	16.37 (2.5)	16.41 (2.5)
Sex (% female)	54.9%	54.6%	54.7%
Race (% non-Hispanic white)	86.5%	90.1%	88.7%
Geriatric Depression Scale 1	0.70 (1.1)	1.34 (1.4)	1.09 (1.3)
AMNART Verbal Intelligence	120.75 (7.9)	119.75 (8.2)	120.13 (8.1)
MOCA 1	26.02 (2.4)	25.12 (2.9)	25.41 (2.8)

Note: AMNART = North American Adult Reading Test, MOCA = Montreal Cognitive Assessment,

¹Denotes significant difference between groups, p<.001.

Relationship Between Learning Slopes/Traditional Memory Measures and SMC

Based on the aforementioned demographic results, age was used as a covariate in the subsequent hierarchical regression analyses. Table 2 displays the means and SDs for learning slopes and traditional memory measures for CN and SMC participants, with SMC participants consistently performing worse on learning scores than CN participants. As seen in Table 3, regression models including age and learning slope significantly predicted SMC status for most learning slope metrics, including: Word Recall LR, RAVLT LR, RAVLT RLS, and RAVLT LOT. No differences were observed for Word Recall RLS or LOT analyses. This effect remained when accounting for baseline self-reported depression (r² change scores for Word Recall LR and RAVLT LR, RAVLT RLS, and RAVLT LOT=.01 each, ps=.001-.02). When comparing AIC and BIC values between LR derived from RAVLT with RAVLT RLS and LOT – or LR derived from Word Recall with Word Recall RLS and LOT Word Recall – the AIC and BIC values for LR metrics were smaller than those of RLS and LOT, suggesting better model goodness-of-fit at predicting SMC status.

Table 2.

Learning slope and memory scores for the diagnostic groups and total sample. All values are Mean (Standard Deviation; range) unless listed otherwise.

Variable	Cognitively Normal	Subjective Memory Concern	Total Sample
n	364	586	950
RAVLT LR	0.69 (0.2; 0.00 – 1.00)	0.64 (0.2; 0.00 – 1.00)	0.66 (0.2; 0.00 – 1.00)
RAVLT RLS	6.40 (2.2; 0.0 – 11.0)	6.12 (2.3; 0.0 – 12.0)	6.23 (2.3; 0.0 – 12.0)
RAVLT LOT	18.37 (7.4; −3.0 – 38.0)	17.31 (7.7; −4.00 – 36.0)	17.72 (7.6; −3.00 – 46.0)
Word Recall LR	0.68 (03; −1.00 – 1.00)	0.63 (0.3; −1.00 – 1.00)	0.65 (0.3; −1.00 – 1.00)
Word Recall RLS	2.94 (1.4; −1.0 – 7.0)	3.02 (1.3; −1.0 – 7.0)	2.99 (1.4; −1.0 – 7.0)
Word Recall LOT	4.84 (2.7; −5.00 – 12.0)	4.97 (2.5; −3.00 – 14.0)	4.92 (2.6; −5.0 – 14.00)
RAVLT Total Recall	45.92 (10.1; 18.0 – 71.0)	43.38 (10.6; 18.0 – 70.0)	44.35 (10.5; 18.0 – 71.0)
RAVLT Delayed Recall	7.86 (3.9; 0.0 – 15.0)	6.84 (4.2; 0.0 – 15.0)	7.23 (4.1; 0.0 – 15.0)
LM Immediate Recall	14.25 (3.1;3.0 – 22.0)	12.67 (3.2; 3.0 – 23.0)	13.27 (3.3; 3.0 – 23.0)
LM Delayed Recall	13.30 (3.2; 0.5 – 22.0)	10.64 (3.6; 0.0 – 23.0)	11.66 (3.7; 0.0 – 23.0)
Word Recall Immediate Recall	21.50 (3.8; 10.0 – 30.0)	20.22 (4.0; 3.0 – 30.0)	20.71 (3.9; 3.0 – 30.0)
Word Recall Delayed Recall	7.30 (1.8; 0.0 – 10.0)	6.56 (2.1; 0.0 – 10.0)	6.85 (2.0; 0.0 – 10.0)

Note: RAVLT = Rey Auditory Verbal Learning Test, LR = Learning Ratio, RLS = Raw Learning Score, LOT = Learning Over Trials, Word Recall = Word Recall subtest from the Alzheimer’s Disease Assessment Scale – Cognitive subscale.

Table 3.

Hierarchical regression of learning slopes and traditional memory measures predicting subjective memory concern status (n = 950).

	Model F(df), p, r 2	Incremental r2 change, p	AIC, BIC Value
RAVLT LR
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: RAVLT LR	F(2, 947) = 15.44, p<.001, r2 = .03	r2 = .02, p<.001	AIC=1239.16, BIC=1253.73
RAVLT RLS
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: RAVLT RLS	F(2, 947) = 10.08, p<.001, r2 = .02	r2 = .006, p=.015	AIC=1249.62, BIC=1264.18
RAVLT LOT
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: RAVLT LOT	F(2, 947) = 11.03, p<.001, r2 = .02	r2 = .008, p=.005	AIC=1247.75, BIC=1262.32
Word Recall LR
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: Word Recall LR	F(2, 947) = 13.39, p<.001, r2 = .03	r2 = .02, p<.001	AIC=1243.14, BIC=1257.71
Word Recall RLS
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: Word Recall RLS	F(2, 947) = 7.53, p<.001, r2 = .016	r2 = .001, p=.35	AIC=1254.64, BIC=1269.21
Word Recall LOT
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: Word Recall LOT	F(2, 947) = 7.39, p<.001, r2 = .01	r2 = .001, p=.43	AIC=1254.92, BIC=1269.48
RAVLT Total Recall
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: RAVLT Total Recall	F(2, 947) = 20.24, p<.001, r2 = .04	r2 = .02, p<.001	AIC=1229.87, BIC=1244.44
RAVLT Delayed Recall
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: RAVLT Delayed Recall	F(2, 947) = 18.83, p<.001, r2 = .04	r2 = .02, p<.001	AIC=1229.70, BIC=1244.26
LMI
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: LMI	F(2, 947) = 36.72, p<.001, r2 = .07	r2 = .06, p<.001	AIC=1198.68.16, BIC=1213.25
LMII
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: LMII	F(2, 947) = 76.28, p<.001, r2 = .14	r2 = .12, p<.001	AIC=1127.74, BIC=1142.31
Word Recall Immediate Recall
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: Word Recall Immediate Recall	F(2, 947) = 24.80, p<.001, r2 = .05	r2 = .03, p<.001	AIC=1221.13, BIC=1235.70
Word Recall Delayed Recall
Step 1: Age	F(1, 948) = 14.17, p<.001, r2 = .01	--
Step 2: Word Recall Delayed Recall	F(2, 947) = 29.18, p<.001, r2 = .06	r2 = .04, p<.001	AIC=1212.82, BIC=1237.79

Note: RAVLT = Rey Auditory Verbal Learning Test, LR = Learning Ratio, RLS = Raw Learning Score, and LOT = Learning Over Trials, Word Recall = Word Recall subtest from the Alzheimer’s Disease Assessment Scale – Cognitive Subscale, LMI = Logical Memory Immediate Recall, LMII = Logical Memory Delayed Recall, AIC = Akaike Information Criterion, BIC = Bayesian Information Criterion.

Additionally, supplementary analyses indicate that learning slope scores correlate significantly and negatively with the ECog Memory Index (rs= −.16 to −.18, ps<.001 for RAVLT/Word Recall LR and ECog-PT, and rs=−.16 to −.27, ps<.001 for RAVLT/Word Recall LR and ECog-INF) and CCI (rs=−.12 to −.19, ps=.001-.02 for LR and CCI) scores.

Similarly, regression models including age and traditional learning and memory measures significantly predicted SMC status for all measures (see Table 3), including: RAVLT Total Recall, RAVLT Delayed Recall, LM I, LM II, Word Recall Immediate Recall, and Word Recall Delayed Recall. In a direct comparison of AIC and BIC values between RAVLT LR and Word Recall LR with their respective traditional memory measures (e.g., RAVLT LR versus RAVLT Total Recall, Word Recall LR versus Word Recall Immediate Recall), the AIC and BIC values for LR metrics were larger than those of the traditional memory measures (despite roughly the same r² change values), suggesting better model goodness-of-fit by the traditional scores at predicting SMC status.

Relationship with AD Biomarkers

Table 4 displays the results of hierarchical linear regressions predicting bilateral hippocampal volume from learning slope performance and SMC status. Across all learning slopes (RAVLT LR, RLS, LOT, and Word Recall LR, RLS, LOT), models including age, sex (given the known relationship with hippocampal volume; Perlaki et al., 2014), SMC status, and learning slope significantly predicted bilateral hippocampal volume (ps<.001, r²s=0.20-0.21). The interaction term between SMC and the respective learning slope was significant for Word Recall LR and RAVLT LR (ps=.002-.009, r² changes=.006-.008), but not for RAVLT RLS or LOT (after correcting for multiple comparisons; ps=.017-.036, r² changes=.004-.005) or Word Recall RLS or LOT (ps=.49-.70, r² changes=.000). Figure 2 shows that while the SMC group exhibited a consistent positive relationship between learning slope and bilateral hippocampal volume for RAVLT LR and Word Recall LR, the CN group displayed either no relationship or a negative relationship between slope and hippocampal volume.

Table 4.

Hierarchical regression of learning slopes and subjective memory complaint status predicting bilateral hippocampal volume (n = 950).

	Model F(df), p, r 2	Incremental r2 change, p
RAVLT LR
Step 1: Age, Sex, SMC, RAVLT LR	F(4, 935) = 63.15, p<.001, r2 = .21	--
Step 2: SMC x RAVLT LR Interaction	F(5, 934) = 52.21, p<.001, r2 = .22	r2 = .006, p=.009
RAVLT RLS
Step 1: Age, Sex, SMC, RAVLT RLS	F(4, 935) = 61.81, p<.001, r2 = .21	--
Step 2: SMC x RAVLT RLS Interaction	F(5, 934) = 52.21, p<.001, r2 = .21	r2 = .004, p=.036
RAVLT LOT
Step 1: Age, Sex, SMC, RAVLT LOT	F(4, 935) = 61.97, p<.001, r2 = .21	--
Step 2: SMC x RAVLT LOT Interaction	F(5, 934) = 50.97, p<.001, r2 = .21	r2 = .005, p=.017
Word Recall LR
Step 1: Age, Sex, SMC, Word Recall LR	F(4, 935) = 61.04, p<.001, r2 = .21	--
Step 2: SMC x Word Recall LR Interaction	F(5, 934) = 51.30, p<.001, r2 = .22	r2 = .008, p=.002
Word Recall RLS
Step 1: Age, Sex, SMC, Word Recall RLS	F(4, 935) = 60.79, p<.001, r2 = .20	--
Step 2: SMC x Word Recall RLS Interaction	F(5, 934) = 48.62, p<.001, r2 = .20	r2 = .000, p=.70
Word Recall LOT
Step 1: Age, Sex, SMC, Word Recall LOT	F(4, 935) = 60.86, p<.001, r2 = .20	--
Step 2: SMC x Word Recall LOT Interaction	F(5, 934) = 48.76, p<.001, r2 = .20	r2 = .000, p=.49

Note: SMC = Subjective Memory Complaint, RAVLT = Rey Auditory Verbal Learning Test, LR = Learning Ratio, RLS = Raw Learning Score, and LOT = Learning Over Trials, Word Recall = Word Recall subtest from the Alzheimer’s Disease Assessment Scale – Cognitive Subscale.

Figure 2.

Interaction effect of Subjective Memory Concern (SMC) status and A) RAVLT Learning Ratio and B) Word Recall Learning Ratio on the predicted bilateral hippocampal volume.

Table 5 displays the findings for hierarchical linear regressions predicting β-amyloid positivity from learning slope performance and SMC status. Across all learning slopes (RAVLT LR, RLS, LOT, and Word Recall LR, RLS, LOT), models including age, SMC status, and learning slope significantly predicted β-amyloid positivity (ps<.001, r²s=0.02-0.04). When considering the interaction effect between learning slopes and SMC status, no significant findings were observed across learning slopes, although the interaction between SMC status and RAVLT LR performance was close to significance (p=.07, r² change=.003). As can be seen in Figure 3, it appears that RAVLT LR exhibited a negative relationship between SMC status and β-amyloid, whereas the CN group displayed no relationship.

Table 5.

Hierarchical regression of learning slopes and subjective memory concern status predicting β-amyloid positivity (n = 950).

	Model F(df), p, r 2	Incremental r2 change, p
RAVLT LR
Step 1: Age, Sex, SMC, RAVLT LR	F(3, 946) = 9.79, p<.001, r2 = .03	--
Step 2: SMC x RAVLT LR Interaction	F(4, 945) = 8.19, p<.001, r2 = .03	r2 = .003, p=.07
RAVLT RLS
Step 1: Age, Sex, SMC, RAVLT RLS	F(3, 946) = 8.90, p<.001, r2 = .02	--
Step 2: SMC x RAVLT RLS Interaction	F(4, 945) = 6.84, p<.001, r2 = .02	r2 = .001, p=.33
RAVLT LOT
Step 1: Age, Sex, SMC, RAVLT LOT	F(4, 935) = 9.15, p<.001, r2 = .02	--
Step 2: SMC x RAVLT LOT Interaction	F(5, 934) = 6.87, p<.001, r2 = .02	r2 = .000, p=.81
Word Recall LR
Step 1: Age, Sex, SMC, Word Recall LR	F(3, 946) = 11.22, p<.001, r2 = .03	--
Step 2: SMC x Word Recall LR Interaction	F(4, 945) = 8.63, p<.001, r2 = .03	r2 = .001, p=.35
Word Recall RLS
Step 1: Age, Sex, SMC, Word Recall RLS	F(4, 935) = 9.21, p<.001, r2 = .02	--
Step 2: SMC x Word Recall RLS Interaction	F(5, 934) = 6.90, p<.001, r2 = .02	r2 = .000, p=.87
Word Recall LOT
Step 1: Age, Sex, SMC, Word Recall LOT	F(4, 935) = 8.93, p<.001, r2 = .02	--
Step 2: SMC x Word Recall LOT Interaction	F(5, 934) = 6.70, p<.001, r2 = .02	r2 = .000, p=.87

Note: SMC = Subjective Memory Concern, RAVLT = Rey Auditory Verbal Learning Test, LR = Learning Ratio, RLS = Raw Learning Score, and LOT = Learning Over Trials, Word Recall = Word Recall subtest from the Alzheimer’s Disease Assessment Scale – Cognitive Subscale.

Figure 3.

Interaction effect of Subjective Memory Concern (SMC) status and A) RAVLT Learning Ratio and B) Word Recall Learning Ratio on β-amyloid status.

DISCUSSION

Learning slopes derived from the RAVLT and Word Recall subtest of the ADAS-Cog were sensitive to differences in subjective memory concerns in this study, such that models that incorporated them could successfully predict SMC status. This effect was observed for all three learning slopes derived from the RAVLT, and for the LR metric derived from Word Recall. When examining learning slope means between SMC groups, CN participants consistently outperformed SMC participants. The results remained after accounting for baseline self-reported depression. To ensure confidence in our groups based on ADNI’s classification of SMC, we additionally found that the SMC group endorsed significantly higher symptoms of self- and informant-rated subjective cognitive concerns on separate subjective-concern inventories in ADNI (ECog and CCI), and that learning slopes were significantly and negatively associated with endorsements on these inventories. The examination of learning slopes in participants with SMC has thus far been limited to a report of raw learning score calculation in 60 participants with SMC undergoing memory treatment (Frankenmolen et al., 2018), consequently our work is the initial documentation of differences in learning slope in participants with SMC. As such, future research should endeavor to replicate these findings. Our results, however, are consistent with studies observing that participants with SMC display worse memory and cognitive functioning (Park et al., 2019; Sohrabi et al., 2019) and daily functioning (Ryu et al., 2016), and possess greater likelihoods of developing MCI and dementia within one year (Mitchell et al., 2014) than cognitively unimpaired participants without SMC. Consequently, while the association between SMC and objective cognitive impairments tends to vary across a range of factors (including depression status or clinical diagnosis [Buelow et al., 2014; Loessner et al., 2022; Sohrabi et al., 2019]), SMC seems to represent initial signs of cognitive change – to which learning slopes appear to be sensitive. This is consistent with the conceptualization that SMC reflects an early stage of the NIA-AA Research Framework. These results also coincide with Thomas and colleagues’ recent work (Thomas et al., 2020; Thomas et al., 2018) suggesting that memory process scores – in particular LOT – are sensitive to subtle cognitive decline, which is a similar but not identical condition along the early AD continuum. Our findings further correspond with previous research involving learning slopes and MCI, which is the next step in the AD continuum once objective cognitive impairments have reached the threshold of impairment. In particular, previous investigation has shown that learning slopes are sensitive to MCI, and that participants with MCI outperform those with dementia due to AD on metrics of learning slope (Gifford et al., 2015; Hammers, Gradwohl, et al., 2021; Hammers, Kostadinova, et al., 2022; Hammers, Suhrie, Dixon, Gradwohl, Duff, et al., 2021).

Examining between learning slope metrics (Table 3), LR displayed smaller AIC and BIC values than RLS or LOT for both RAVLT and Word Recall. These results are consistent with prior findings that LR tends to be the preferred learning slope metric, based on stronger magnitudes of the effect across a variety of neurodegenerative groups including AD (Hammers, Gradwohl, et al., 2021) and early-onset AD (Hammers et al., 2023). Conversely, learning slopes displayed mostly comparable variance relative to their respective traditional memory measures when predicting SMC status (RAVLT LR r² change =.02 vs. RAVLT Total Recall r² change =.02, Word Recall LR r² change=.02 vs. Word Recall Immediate Recall r² change =.03), though AIC and BIC values tended to suggested better goodness-of-fit for the traditional memory measures. These results generally suggest that the sensitivity of the LR metric towards SMC complements that of traditional memory measures, which mostly coincides with previous research indicated that learning slopes explained comparable variance in neurodegeneration as summary indexes of total learning (Gifford et al., 2015; Hall et al., In Press; Hammers et al., 2023).

It is of note that such comparisons to traditional learning or memory measures are not intended to suggest that learning slope metrics be used in their place. In the current results, the effect of total learning was not included in the prediction of SMC by learning slopes due to the high degree of multicollinearity between learning slopes and their associated traditional memory measures (rs=.60-.83) in this sample. When combined with the AIC and BIC comparisons, we are therefore unable to make the claim that there is evidence of LR possessing incremental advantages of predicting AD biomarkers in participants with SMC relative to traditional memory measures. Instead, we aim to highlight that by considering trial-by-trial acquisition of information for a particular participant or patient, learning slopes may enhance treatment recommendations and clinical decision-making – including in those patients with SMC.

Additionally, our results further previous investigation indicating that learning slopes are sensitive to AD biomarkers of hippocampal volume (Gifford et al., 2015; Hammers, Gradwohl, et al., 2021; Hammers, Suhrie, Dixon, Gradwohl, Archibald, et al., 2021) and β-amyloid positivity (Hammers, Kostadinova, et al., 2022; Hammers, Suhrie, Dixon, Gradwohl, Archibald, et al., 2021). Specifically, we observed that a model containing demographics, SMC status, and learning slope performance predicted bilateral hippocampal volumes for each of the six learning metrics evaluated (Table 4). Similar models predicted β-amyloid positivity for all six learning metrics (Table 5). This corresponds to learning slopes tapping into both episodic-memory-related and working memory/attention-related aspects of cognition (Gifford et al., 2015).

Although close inspection of effect sizes indicate that magnitudes of effect were small across learning slopes and AD biomarkers (r²=.20-.21 for hippocampal volume and r²=.02-.03 for β-amyloid positivity), these results are not necessarily surprising given that our sample was composed of participants without objective cognitive deficit. For example, in the four aforementioned studies in the literature examining slopes and either hippocampal volume or β-amyloid values, three considered the relationship across samples of cognitively normal, MCI, and dementia/AD. While Gifford et al. (2015) examined hippocampal volume prediction in cognitively intact participants, they observed β values of 0.00 (ps=.46 to .96) when predicting hippocampal volume across a range of learning slopes. This is because the range of AD biomarkers observed in cognitively normal samples are generally too restricted to result in the strength of associations typically seen across clinical samples. For example, the standard deviation surrounding the mean hippocampal volume for participants with normal cognition in a large-scale cohort like the ADNI (Weiner et al., 2017) is 44% smaller than the deviation around MCI or AD samples. Similarly, the standard deviation surrounding β-amyloid SUVR values is 76% smaller in cognitively normal samples than MCI or AD samples (Weiner et al., 2017). Although the magnitudes of effect are small, these significant results coalesce with meta-analytic findings that cognitive performance has “small but nontrivial associations” with increased amyloid burden in cognitively normal samples (Hedden et al., 2013), and suggest that learning slopes may be sensitive to AD biomarkers at the earliest stages of the disease. While initial research has suggested an association between learning slopes and tau pathology (Hammers, Kostadinova, et al., 2022), the current cognitively unimpaired sample had too few participants who were tau positive to render analysis possible. Given the pathological specificity of tau and its temporal proximity to cognitive change (Jack et al., 2013), future research with larger sample sizes should consider the role of tau pathology in learning in SMC.

Further, examination of interactions between factors suggests that learning slopes – LR in particular – predict bilateral hippocampal volumes differently as a function of SMC status. As seen in Figure 2, while either no relationship or a slightly negative relationship is observed between hippocampal volume and Word Recall/RAVLT LR in cognitively normal participants, strong positive relationships were evident between hippocampal volume and LR values in participants with SMC – such that worse LR was predictive of smaller hippocampal volume in SMC participants. Consequently, the presence of SMC appears to be driving the relationship observed above between learning slopes and hippocampal volume in our sample. These results are consistent with recent findings that smaller hippocampal volumes were associated with greater SMC in a large cohort of non-demented older adults (Dauphinot et al., 2020), and that participants with SMC possessed significant relationships between memory performance and hippocampal volume (that was not observed in healthy controls; Caillaud et al., 2019). Similarly, while no significant interaction existed between learning slopes and SMC status for β-amyloid positivity, visual inspection of Figure 3.A hints that greater β-amyloid positivity may be related to worse RAVLT LR performance in our SMC group – but not in the CN group. No interaction effect is observed between Word Recall LR and β-amyloid positivity in Figure 3.B, suggesting that decreases in LR for both SMC and CN groups were related to increased β-amyloid positivity. Together, the main effects between LR and β-amyloid positivity seen herein support previous work in cognitively unimpaired older adults that both β-amyloid positivity and neurodegeneration are associated with SMC – independent of objective memory performance (Amariglio et al., 2015; Pavisic et al., 2021).

Finally, it can be observed that Word Recall LR performed comparably to – or better than – RAVLT LR across analyses. This is evident despite some notable differences in their respective word lists. Specifically, the ADAS-Cog Word Recall word list includes 10 words presented over 3 trials (range 0-30 points), whereas the RAVLT LR possesses 15 words across 5 trials (range 0-75 points). This supports previous findings that trial length and number of trials of the original word list appear to have limited impact on learning slope (Hammers, Spencer, et al., 2022; Hammers, Suhrie, Dixon, Gradwohl, Duff, et al., 2021). Additionally, it adds to limited literature suggesting validation of not only a learning slope metric from the Word Recall subtest of the ADAS-Cog, but for validation of any use of ADAS-Cog beyond its summary total score. Previously, we observed that ADAS-Cog memory subtests and learning slopes could be used to predict status among the Research Framework for AD (“ATN model”; Hammers, Kostadinova, et al., 2022; Jack et al., 2018). This criterion validity of Word Recall LR in SMC is important for two reasons. First, because the ADAS-Cog is already traditionally administered in AD clinical trials, this learning metric could be easily incorporated into study designs to aid in the prediction of AD pathology, or potentially increase the sensitivity of trial outcomes. Second, due to considerations that SMC may reflect the earliest stages of detectible cognitive change due to AD, it has been proposed that investigation of disease-modifying treatments in AD and implementation of secondary prevention approaches should be considered in cognitively unimpaired participants with SMC (Caillaud et al., 2019). Together, this suggests that LR derived from Word Recall may be a useful tool in AD clinical trials moving forward, particularly if incorporating samples with SMC.

The following limitations are of particular relevance. First, as is seen in a number of ADNI studies, our sample was highly homogenous with respect to ethnicity and education (mostly Caucasian and highly educated). As this potentially restricts the generalizability of results, future research into learning slopes in more heterogenous samples is strongly encouraged. Second, ADNI uses stringent exclusion criteria typical of industry-sponsored clinical trials, making this cohort of individuals across the AD biomarker continuum potentially not be representative of the general population. Third, when applying actuarial criteria to the ADNI data in the current study, a combination of two different sets of normative comparisons were required (Shirk et al., 2011; Weintraub et al., 2018) because of modifications to ADNI protocols over time (ADNI2, 2011; ADNI3, 2016). It remains possible that the use of multiple norms for the same measures could have introduced some additional variance into the study. Fourth, these results are unique to learning slopes derived from either the ADAS-Cog Word Recall subtest or the RAVLT, and cannot be generalized to verbal memory tasks that are not related to list learning (such as Craft Story 21 Memory; Besser et al., 2018) or to visual memory tasks. Fifth, we did not examine all possible learning slope calculations, but only some commonly used slope metrics (RLS, LOT, LR). Consequently, these results cannot speak to relationships between SMC and/or AD biomarkers and all learning slopes. As such, the calculation of LR from Spencer and colleagues (2020) may overlook some patterns of performance across trial-based learning tasks, such as those seen in Ribeiro et al.’s (2007) work on non-linear fitting of learning rates. Future research to identify different performance patterns of learning – which have not yet been fully idealized – is recommended. Sixth, our classification of participants into the SMC group was based on either participant- or informant-report of subjective memory concerns. Therefore it is possible that some discrepancy may exist between sources of information, though research suggests that the relationship between AD biomarkers and SMC is consistent whether concerns were from patients or collaterals (Pavisic et al., 2021), and our corresponding findings with the ECog for both patients and informants reduces this concern to some degree. Seventh, while our study focused on subjective memory concerns, it is possible that patients may have subjective concerns for cognitive domains that are not amnestic in nature. Although our SMC group endorsed higher levels of both Memory and Total cognitive concerns on both the ECog and CCI, it is possible that our learning slope, SMC, and AD biomarker findings may not generalize to individuals with only non-amnestic cognitive concerns.

Finally, it is important to note that the use of the actuarial criteria for the diagnosis of MCI does not mean that CN participants cannot display performances on memory tests that were below normal limits. As described above (Bondi et al., 2014; Jak et al., 2009), the actuarial criteria used normative performances on two separate measures of memory, language, and executive functioning (each) to classify participants as CN versus MCI. In the current study, performance on the RAVLT or Word Recall from the ADAS-Cog was intentionally not used in the application of actuarial criteria – so as to avoid diagnostic circularity. As such, it is possible that CN participants in our study performed intactly on the cognitive measures used to diagnose, but possessed a relative weakness in either the RAVLT or Word Recall (as evidenced by some low performances observed in Table 2). However, we used the current methods to be consistent with the validation of the actuarial diagnostic process, given the known limitations of ADNI diagnoses.

Despite the aforementioned limitations, our results indicate that not only are learning slopes sensitive to SMC, but that the presence of SMC in cognitively unimpaired participants appears to impact the relationship between learning slopes and AD biomarkers. These findings advance our knowledge of SMC, which is an early condition along the NIA-AA Research Framework for the diagnosis of AD continuum, and suggest that LR derived from the Word Recall subtest of the ADAS-Cog may – in particular – be an important tool for the detection of AD pathology in both participants with SMC and in AD clinical trials.