Structural Neuroanatomy of Semantic Retrograde Memory in Older Adults

Abstract

Studies of brain lesions or volumes indicate that the integrity of medial and lateral temporal lobe structures are important for news event memory accuracy, but the relationship between cortical thickness and news event memory accuracy has not yet been investigated in older adults. In a mixed sample of 70 older adults with variable cognitive abilities, but without dementia, we investigated the relationship between cortical volume, hippocampal volume, and cortical thickness with news event recognition memory accuracy across the entire adult lifespan using the Retrograde Memory News Events Test (RM-NET). Partial Least Squares analysis was used to identify brain regions where news event memory accuracy scores significantly correlated with cortical volume, hippocampal volume, and cortical thickness. We found that mean news event memory accuracy significantly correlated with volume/thickness for a network of regions that included the hippocampus, medial/lateral temporal lobe, medial/lateral parietal lobe, and specific areas within the medial/lateral prefrontal cortex. Poorer performance was associated with a thinner cortex (and smaller volumes). Almost all regions in this network exhibited decreasing brain-behavior correlations as the age of memory increased, thus retrieval of remote memories was less reliant on the network. We also found regions in this network that were not identified by the RM-NET post-test (a measure of episodic anterograde memory for the RM-NET content) nor traditional neuropsychological tests. The regions identified as uniquely contributing to news event memory overlap with regions known to exhibit increasing AD pathology and cortical thinning when pathology begins to spread outside of the medial temporal lobe.

Introduction

Older adults who develop Alzheimer’s disease (AD) are thought to pass through a transitional stage between normal aging and AD, known as Mild Cognitive Impairment (McIntosh & Lobaugh, 2004; Petersen et al., 1999). The hallmark feature of MCI is a mild impairment in cognition, usually in episodic anterograde memory (EM)(Petersen et al., 1999), which reflects difficulty remembering information together with linked information about the learning context (e.g., time, place; Tulving, 1983). This observation is consistent with the neuropathological changes in the medial temporal lobe (MTL) that appear early in disease progression(Braak & Braak, 1991; Thal & Braak, Postmortale Diagnosestellung bei Morbus Alzheimer. Stadiengliederungen der kennzeichnenden Hirnveranderungen./2005). Yet, there is long-standing evidence that semantic abilities are also impacted early in the disease(Barbeau et al., 2012; Gardini et al., 2013; Hodges et al., 2006; Salmon & Bondi, 2009), which is consistent with observations that, early in the disease process, neuropathological changes spread from MTL to lateral temporal lobe cortex (LTL) (Edmonds et al., 2016; McEvoy et al., 2009)and beyond, where semantic memory is thought to be stored.

Unlike EM, semantic memories have lost the linked information about the learning context, and the learned information remains as knowledge or facts. Traditional neuropsychological assessments of semantic (retrograde) memory assess knowledge about semantic identities, semantic concepts, and language and they can reveal dramatic impairment in AD and other dementias. Unfortunately, they are relatively insensitive to the mild semantic memory deficits in MCI, such that individuals with MCI exhibit only mild difficulty on these tests and frequently the impairment is not statistically reliable(Venneri et al., 2016). By contrast, amnestic MCI and mixed MCI groups (i.e., amnestic and non-amnestic MCI) are consistently impaired on novel semantic retrograde memory measures, such as memory for news facts or famous personalities(Asp et al., 2024; Barbeau et al., 2012; Flicker et al., 1987; Irish et al., 2010; Leyhe et al., 2010; Leyhe et al., 2009; Murphy et al., 2008; Seidenberg et al., 2009; Smith, 2014; Thomann et al., 2012). Thus, memory for this type of semantic information appears to be more vulnerable to disruption than memory for the long-established and enduring information queried by traditional semantic memory tests.

Within the context of novel semantic memory tests, those that query news events or personalities that were in the public eye for only a limited time (transient events/personalities) are more sensitive to deficits in MCI and AD than measures that query events that experienced long-lasting attention (enduring events/personalities)(Benoit et al., 2017; Langlois et al., 2016; Seidenberg et al., 2009). In addition, measures of news events or personalities that came to prominence in the recent past are more sensitive to the cognitive deficits in MCI than measures of more remote events(Asp et al., 2024; Benoit et al., 2017; Langlois et al., 2016; Smith, 2014). These findings are consistent with the idea that novel tests of semantic retrograde memory hold significant promise for identifying the early, subtle cognitive and neural changes that lead to AD (Jack et al., 2010), particularly if the tests query transient events and recent memories.

The semantic retrograde memory network (see Figure 1) overlaps substantially with brain regions that undergo the most cortical thinning when one transitions from normal aging to early AD(Braak & Braak, 1991; Dante Mantini and Wim, 2013; Dede & Smith, 2016; Eskildsen et al., 2013). For MCI, the brain regions that appear to support memory for news event memory or famous personalities are not only in the MTL, but also in the LTL, medial/lateral parietal lobe, and in specific regions of medial/lateral prefrontal cortex(Barbeau et al., 2012; Gardini et al., 2013; Serra et al., 2022; Smith, 2014). Yet, these studies used either voxel-based morphometry (i.e., cortical volume, hippocampal volume, and cortical thickness were not measured) or examined a small number of participants with MCI (N=11). Therefore, it is important to examine brain-behavior relationships in a mixed sample where there is more variability in both cognitive difficulties and neuroanatomical decline (i.e., individuals with normal cognition [NC] or MCI). Moreover, because recent memories are more vulnerable to disruption than remote memories, it is important to also test whether these brain-behavior relationships change as a function of memory age.

Figure 1. Visualization of the semantic retrograde memory network adapted from Dede & Smith (2016).

Brain regions involved in this network are labeled on the lateral (left) and medial (right) surfaces of the brain. The hippocampus cannot be seen on the brain surface, so it is shown near the parahippocampal gyrus on the medial surface of the brain.

Although novel retrograde memory measures have a long history of revealing impairment after brain injury or disease, they show reliable impairment in MCI groups, and they appear to depend on structures that are first affected in AD, these sensitive measures are not used clinically to identify those at risk for dementia (e.g., MCI). Instead, tests of episodic anterograde memory are most frequently used, even though both EM and the MTL structures that support it decline in normal aging. To reliably assess AD cognitive risk in older adults, and determine which patients would benefit from new therapies, sensitive measures are needed that detect cognitive changes associated with abnormal aging and that depend on structures that are largely unaffected by normal aging.

Therefore, we examined the relationship between structural neuroanatomy (cortical volume, hippocampal volume, and cortical thickness) and retrograde memory in a mixed sample with variable cognitive difficulties (N=70 with NC or MCI) and using the Retrograde Memory News Events Test (RM-NET), which queries memories across the entire adult lifespan (and avoids enduring memories). Behavioral Partial Least Squares (B-PLS) analysis was used to examine correlations between news event memory accuracy scores and measures of neuroanatomy. B-PLS was also used to examine whether these brain-behavior correlations changed as a function of memory age. To identify regions unique to semantic retrograde memory, we compared these findings with findings from B-PLS analyses using performance on traditional neuropsychological tests as well as a novel measure of anterograde episodic memory (RM-NET post-test). We predicted that the RM-NET would (1) identify regions in the retrograde memory network (see Figure 1), that (2) many of these regions would exhibit brain-behavior correlations that changed with memory age, and that (3) be associated with unique regions not be identified by traditional measures of cognition, including the RM-NET post-test.

Method

Participants

A mixed sample of seventy older adults with variable cognitive difficulties but without dementia (34 with NC and 36 with MCI) completed the study (see Table 1). Although differences between the NC and MCI participants were not the focus of the current study, we did include some behavioral comparisons and limited comparisons of the brain-behavior associations (described below) across these subgroups. Accordingly, we briefly describe how these subgroups were identified below and include a more detailed description in the Supplemental Information. Participants were classified as having NC or MCI (72% amnestic, 28% non-amnestic) using the conceptual framework from Jak et al. (2009) and Bondi et al. (2014). MCI was identified if performance on traditional cognitive assessments (see below and Supplemental Table 1) was impaired (> 1SD below demographically appropriate norms) on (1) at least 2 tests within a single cognitive domain, OR (2) at least 1 test across three or more cognitive domains (see Cognitive Assessment, below).

Table 1.

Demographic Information, Participant Characteristics, and Cognition (N=70)

	Mean or #	SD or %	Min	Max
Demographic Information
Age	73.2	6.4	65	91
Education	15.8	2.3	12	20
Gender (%Men)	46	65.7%	-	-
Ethnicity (%Hispanic)	8	11.4%	-	-
Race
American Indian (%)	4	5.7%	-	-
Asian (%)	2	2.9%	-	-
Black (%)	5	7.1%	-	-
Pacific Islander (%)	1	1.4%	-	-
White (%)	63	90.0%	-	-
Comorbidities
Medical Health Burden	3.4	1.9	0	8
Mental Health Burden	1.5	1.3	0	4
RM-NET
Lifespan Accuracy	63.7	13.0	35.2	91.3
Post-Test Accuracy	78.7	14.3	44.8	96.3
Autobiographical Memories	2.6	3.1	0	13
RM-NET Test Interval (days)	564.9	452.3	0	1297
News Habits Score	8.2	1.9	2	11
Cognitive Composite Scores
Episodic Memory	0.1	0.9	−2.3	2.0
Semantic Memory/Language	0.2	0.7	−1.3	1.8
Executive Functions	0.0	0.6	−1.5	1.5
Attention/Processing Speed	0.0	0.5	−1.4	1.3
Visuospatial Function	0.2	0.4	−1.2	0.9
Global Cognition
Mini-Mental State Exam	28.3	1.5	25	30

Note. More than one race could be selected, and no participants indicated “Other” or “Unknown”. Three participants indicated “Unknown” ethnicity. Medical and Mental Health burden reflects the total number of comorbidities.

All participants provided written informed consent and were recruited and enrolled without regard to ethnicity or race. Participants were recruited from the Veterans Affairs San Diego Medical Center, local outpatient VA clinics, the University of California San Diego Alzheimer’s Disease Research Center (ADRC), and the San Diego community.

Inclusion Criteria.

65 years old or older; fluent in English; lived in the United States for most of adulthood; right-handed (due to the same sample participating in a functional MRI investigation of news event memory retrieval, which will be reported elsewhere); able to successfully complete a RM-NET practice test.

Exclusion Criteria.

Individuals were excluded if they had any major psychiatric or neurological conditions, including dementia, that could affect cognition, brain structure, or brain function: diagnosis or self-report of dementia; impaired activities of daily living (≤40, low functioning) according to the “health and safety” or “managing money” subscales of the Independent Living Scales Test(Loeb, 1996); impaired score on MMSE (< 25); impaired reading ability (< 2SD below norms) according to the Wide Range Achievement Test Blue Word Reading Test; uncontrolled high blood pressure; traumatic brain injury or head injury with loss of consciousness >30 min; stroke or transient ischemic attack; chronic disorders of the lung or heart; seizures or other neurological conditions such as Parkinson’s disease; type I diabetes or uncontrolled type II diabetes; general anesthesia in the previous 4 months; chemotherapy or full-body radiation for cancer treatment; diagnosed with schizophrenia, bipolar disorder, or psychotic disorders; untreated major depression or exhibiting moderate to severe depression symptoms (score ≥ 7) according to the Geriatric Depression Scale(Sheikh & Yesavage, 1986); contraindication for MRI; or currently enrolled in an alcohol/drug treatment program. See the Definitions of Impairments section in Supplemental Information for more information.

Excluded Participants.

Ninety-six potentially eligible participants were enrolled. Twenty-six participants were excluded for the following reasons after providing informed consent: moderate/severe depression symptoms (N=3), uncontrolled high blood pressure (N=3), incompatible with MRI/abnormality found on MRI (N=7), too slow on news events practice test (N=1), and attrition (N=12). Note that there was a relatively high attrition rate, likely influenced by the COVID-19 pandemic which interrupted enrollment and testing. Many of these individuals could not or would not return to finish the study following the pandemic (other reasons included moving out of the area, stroke, and unable to re-contact). No individuals were excluded for exhibiting impaired activities of daily living, impaired MMSE score, or impaired reading ability. The remaining 70 participants (24 women) completed the study.

Materials

Cognitive Assessments

Traditional Neuropsychological Tests.

All participants were given a comprehensive neuropsychological battery with tests that measured five cognitive domains (Episodic Memory, Semantic Memory/Language, Executive Functions, Attention/Processing Speed, and Visuospatial Functions). Cognitive ability in each of the 5 domains was estimated using 4-7 tests for a total of 26 tests (see Supplemental Table 1).

Retrograde Memory News Events Test (RM-NET).

The recognition portion of the RM-NET (Cawley-Bennett et al., 2022) was used to obtain estimates of news event memory across the entire adult lifespan (1948-2017), separated into sixteen 3–5-year time periods (231 questions in total, see Table 2). The number of questions needed to cover the adult lifespan for each participant varied according to the participant’s age at the time of testing. On average, 210 ± 10 questions were needed to cover the adult lifespan (i.e., events that occurred after each participant was 15 years or older until 2017). Each question was presented in multiple-choice format. An example question was “Which Asian country impeached its first female president?”. Participants were instructed to read each question and select their answer from among 4 options (e.g., A. Indonesia; B. South Korea; C. India, and D. Nepal) within 12.8 seconds, followed by a confidence judgment within an additional 3.2 seconds using a button-box (see Cawley-Bennett et al. 2022 for more details about RM-NET administration).

Table 2.

RM-NET Time Periods and Accuracy Scores

Memory Age	Years	Duration (years)	Number of Items	Number of Participants	% Correct Mean (SD)
1-3	2017-2015	3	20	70	62.5 (18.9)
4-6	2014-2012	3	20	70	64.9 (17.5)
7-9	2011-2009	3	20	70	58.9 (16.4)
10-12	2008-2006	3	20	70	63.0 (15.4)
13-15	2005-2003	3	20	70	65.1 (16.2)
16-20	2002-1998	5	20	70	60.9 (14.8)
21-25	1997-1993	5	20	70	66.0 (16.1)
26-30	1992-1988	5	20	70	63.6 (15.1)
31-35	1987-1983	5	10	70	57.3 (20.6)
36-40	1982-1978	5	8	70	66.3 (18.1)
41-45	1977-1973	5	8	70	64.6 (24.4)
46-50	1972-1968	5	10	70	67.0 (21.0)
51-55	1967-1963	5	8	70	60.0 (19.3)
56-60	1962-1958	5	9	46	73.7 (19.1)
61-65	1957-1953	5	10	24	71.1 (21.7)
66-70	1952-1948	5	8	11	37.9 (21.0)

Note. Memory age refers to the approximate age of the news event memory relative to the year of testing, which occurred between 2018-2021. The number of questions administered to each participant covered the adult lifespan, included questions from 2017 and back to the year when each participant was 15 years old.

RM-NET post-test.

About 20 minutes after completing the first 160 items of the RM-NET, participants began the 160-item post-test to obtain additional information about each news event (i.e., news events from 2017 to 1988). Specifically, participants were asked questions about each news event.

1. Subsequent memory accuracy: participants were asked to identify the specific topic they had been asked about earlier (e.g., Which topic were you asked a question about?: 1. The Asian country that impeached its first female president; 2. The company tied to the impeachment of the first female president of an Asian country; or 3. The reason an Asian country’s first female president was impeached). This surprise recognition memory test was used to obtain a measure of anterograde memory (delayed recognition memory, EM) for the RM-NET content. Assessing memory for information learned incidentally during experimental testing is a standard way of obtaining a measure of EM (Dede et al., 2016; Smith & Squire, 2009).

2. Participants indicated whether or not they had a specific autobiographical memory associated with the news event using the definition of an EM from the Autobiographical Memory Interview (Kopelman et al., 1989) (i.e., a score of 3 on this measure). For example, if they could report specific details about the time and place when they learned about the event and not if they simply remembered hearing it on the radio or seeing it on television. This component provided information about the prevalence of news events that were accompanied by episodic memories.

Procedure

All procedures were approved by the Institutional Review Board at the Veterans Affairs San Diego Healthcare System. Data collection took place across four visits.

The recognition memory portion of the RM-NET was administered between March 2018 and September 2021. For more details about the RM-NET development and full administration procedures, see Cawley-Bennett et al.,(Cawley-Bennett et al., 2022). E-Prime software (Psychology Software Tools, Inc.) was used to administer the RM-NET.

Participants were assessed with a comprehensive neuropsychological battery across Visits 1-2, depending on the amount of testing required to complete the battery. On Visit 3, participants answered RM-NET questions about events that occurred between 2017 and 1988 inside the MRI scanner (160 questions) while functional MRI scans were collected (manuscript in preparation). Structural and functional neuroimages were obtained. After scanning, participants completed a news habits questionnaire and a surprise recognition memory post-test that asked for additional information for each news event. Specifically, for each question we obtained: 1) subsequent memory accuracy – incidental encoding for the content of each news event question and 2) autobiographical memory – whether the news event was associated with a personal and specific autobiographical memory. See Cawley-Bennett et al. (2022) for additional information about the content of the post-test. On Visit 4, participants completed another structural neuroimaging session and then answered RM-NET questions about events that occurred between 1987 and 1948 (up to 71 test items, depending on the age of the participant) using the same setup as was used in the scanner. See Supplemental Information for a complete description of visits and more details about the administration of the RM-NET and the post-test.

Measures of Brain Structure

To obtain measures of cortical volume, hippocampal volume, and cortical thickness, high-resolution, T1-weighted, sagittal structural MRI scans (magnetization-prepared rapid gradient echo [MPRAGE], 208 slices, 1 mm³ voxels, TE = 0.003, TI = 900 ms, Flip angle = 8°, Field of view = 25.6 mm) were obtained using a 32 channel NOVA coil on a General Electric 3 Tesla scanner at the UCSD Center for Functional Magnetic Resonance Imaging on visits 3 and 4. PROMO (PROspective MOtion correction;(White et al., 2009)) was used to adaptively compensate for motion during structural scanning, resulting in no loss of T1-weighted data due to motion.

Computing Variables of Interest

RM-NET.

Time Periods.

For each of the 16 time periods, mean news event memory accuracy was computed (2017–1948, 1 to 70 years before testing, see Table 2). For trials where participants failed to provide a response for a news event question within the allotted time (12.8 seconds), the trial was counted as incorrect. For the three most remote time periods, data were only available from the oldest participants (65% or less of the sample), therefore these time periods were excluded from analyses that examined performance across time periods.

Lifespan Scores.

We computed total accuracy scores across the entire adult lifespan by averaging the mean scores from each time period that was available for each participant (2017-2015; 2014-2012; and so on until reaching the last time period when the participant was 15 years of age). In this way, the Lifespan scores reflected equal weighting of each time period, regardless of how many time periods were needed to cover each participant’s adult lifespan.

Cognitive Domain Composite Scores:

Performance on individual tests from the neuropsychological test battery was converted into z scores based on published norms. Composite scores (z scores) were computed for each participant and domain by averaging the individual z scores for the tests in that domain resulting in 5 composite scores representing the 5 cognitive domains measured. A full description of these methods is provided in(Cawley-Bennett et al., 2022).

Comorbidities

Medical Health Burden and Mental Health Burden.

Due to the wide variety of comorbidities and medications reported by participants, we computed a comorbidities burden measure following the method developed by(Charlson et al., 1987). Individual comorbidities (including medications) were summed across organ systems (e.g., vascular, neurological) to create a continuous measure (medical and mental health burdens examined separately). See (Cawley-Bennett et al., 2022) for a full description of this computation.

Measures of Brain Structure

T1-weighted scans from visits 3 and 4 were averaged prior to the calculation of cortical thickness and volumes using the FreeSurfer version 6.0.0(Dale et al., 1999; Fischl et al., 2002; Fischl et al., 1999; Fischl et al., 2004). Cortical thickness and volume were obtained for each region in the Brainnetome (BN) Atlas(Fan et al., 2016). The BN atlas was chosen based on its fine-grained structural parcellation of the cortex, which provides 105 cortical regions for each hemisphere. Left and right hippocampal volumes were calculated using Freesurfer’s anatomical segmentation (aseg). Manual corrections of FreeSurfer parcellations were carried out to fix errors in the distinction between brain and pia-skull and the distinction between gray matter and white matter. These corrections were carried out blind to group membership by JB and ST. Final data quality checks were carried out by JB and CNS. No participants were excluded due to poor data quality.

Data Analysis

Primary analyses investigated (1) which brain regions were significantly associated with RM-NET Lifespan accuracy scores, and (2) which regions exhibited changes in the strength of the relationship between the brain and RM-NET Time Period accuracy scores as a function of memory age. Means and standard deviations are reported unless explicitly noted otherwise.

All analyses were carried out across the full sample and did not examine NC and MCI groups separately. However, because we were interested in investigating the RM-NET in the context of AD, we examined whether performance on the RM-NET and whether the brain-behavior correlations from the primary and secondary analyses were significantly different across these groups. For completeness, we also reported anatomical differences between these groups in the supplemental information.

Behavioral data analysis examining the NC and MCI groups was carried out using Analysis of Covariance and Repeated Analysis of Covariance on SPSS version 28. Differences between correlations between NC and MCI groups were tested using Systat software version 13. All tests were two-tailed and a probability value less than 0.05 was considered statistically significant. Inverse correlations between cognition and brain measures were not examined.

Identification of Covariates

Pearson correlations were carried out to identify whether demographic (self-reported age, education, gender, and race/ethnicity) and participant characteristics (news habits, comorbidities) were associated with variables of interest (see Supplemental Table 2). Covariates with significant (p < 0.05) associations were identified for inclusion in group-level analyses. The RM-NET test interval reflects the relative number of days that elapsed between when the most recent news event occurred and when the participant was tested. This covariate was included to account for possible forgetting of news events across the data collection period. Due to small sample sizes for individual race and ethnicity categories, race/ethnicity data were combined to create two groups depending on whether participants were both non-Hispanic and white or whether they were not.

Group Level Analysis of Brain-Behavior Associations: Behavior Partial Least Squares

Primary Analyses:

The primary analysis identified the structural neuroanatomy that supported news event memory. Associations between cognitive and brain variables were tested with Multivariate Behavioral Partial Least Squares (B-PLS) analysis (McIntosh & Lobaugh, 2004) using MATLAB (R2023a). Prior to carrying out B-PLS analysis, behavioral and brain measures were corrected for the influence of significant covariates and analyses were carried out on the residuals.

Singular value decomposition (SVD) of the brain-behavior correlation matrices was implemented to yield latent variables (LVs). Each LV contains a singular value, which indicates the total amount of variance in the correlation matrices that is accounted for by an LV. The significance of LVs was assessed through permutation tests on the singular value (10,000 permutations). Bootstrapping was conducted to assess brain regions from significant LVs (10,000 iterations). Both hemispheres (105 cortical regions and 1 hippocampus per hemisphere) were included in one matrix. The bootstrapping ratios (BSRs) for all 212 brain regions were tested for significance. These regions reflected the 105 BN cortical regions and the Freesurfer hippocampus region (only for volumetric analyses) from each hemisphere. Since this is a normal bootstrapping distribution, the BSRs are equivalent to z scores(Efron & Tibshirani, 1986).

We carried out the same procedure for the RM-NET Time Period B-PLS, except we used the RM-NET accuracy score residuals from the 13 most recent time periods. The 13 time periods yielded 13 LVs. The correlation coefficient of each time period from significant LV(s) was extracted to identify if the pattern of correlations identified by an LV changed significantly with respect to the age of the news event memory. Bootstrapping and permutations were the same as described previously.

To characterize the pattern of change across time periods, we used Curve Estimation regression analysis in SPSS and tested for specific patterns thought to reflect the memory consolidation (linear, cubic, and power functions). The predictor was the age of the memory for each time period (e.g., time period 1-3 = 1, time period 4-6 = 4, and so on) and the dependent variable was the correlation coefficients for each time period. The best function was identified as one that both maximized the variance accounted for (R²) and minimized the number of coefficients in the model. For example, if both linear and cubic functions explained a similar amount of variance, the linear function was selected.

An overlap analysis was carried out to identify which brain regions were uniquely identified by the Lifespan analysis or the Time Periods analysis and which regions were common to both analyses. A p value of 0.025 was used for tests of significance to correct for testing of the two primary analyses: RM-NET Lifespan accuracy scores and RM-NET Time Period accuracy scores, for an overall p value of 0.05.

Secondary Analyses:

To identify which brain regions were uniquely identified by the RM-NET (i.e., identified by either the Lifespan accuracy or Time Period analyses) and which regions were common to analyses of other cognitive tests, two secondary analyses were carried out. These analyses examined more traditional types of cognitive cognition used to identify risk for AD. First, using the same procedures as for the RM-NET B-PLS, we identified brain regions that correlated with performance on traditional neuropsychological tests (i.e., each cognitive composite z score). Second, we identified brain regions that correlated with performance on the RM-NET post-test, which reflects anterograde episodic memory. Finally, we created an overlap map to visualize the brain regions that were unique to the RM-NET, unique to more traditional cognitive measures, or common to both analyses. Similar to the primary analyses, a p value of 0.025 was used for tests of significance for each domain of cognition.

Availability of Study Data and Preregistration:

T1-weighted images supporting this work are available by contacting the corresponding author. This study was not preregistered.

Results

Covariates.

Given the known effects of age and brain size on brain structure, age and estimated intracranial volume (from Freesurfer) were included as covariates for brain volumes(Wang et al., 2024). Age alone was included as a covariate for cortical thickness measures because thickness measures are not affected by brain size. Correlations of demographic and participant characteristics with cognitive variables of interest appear in Supplemental Table 2. Age and the RM-NET test interval were significantly correlated with the RM-NET Lifespan and post-test accuracy scores. For the 5 cognitive composite scores, the significant covariates were the following, designated by parentheses: Episodic Memory (education), Semantic Memory/Language (gender); Executive Functions (mental health comorbidity burden), Attention (none), and Visuospatial Function (Race/Ethnicity). Descriptive information about the NC and MCI groups appears in Supplemental Table 1. Due to disruptions in data collection from the COVID-19 pandemic, the NC and MCI groups were not matched on gender or education. Therefore, these demographic characteristics were included as covariates in analyses comparing these groups and we did not to examine the effect of gender in the primary analyses because it would have been difficult to disentangle effects based on gender from effects based on NC-MCI groups.

Performance on RM-NET and RM-NET Post-test.

The mean Lifespan accuracy score was 63.7 ± 13.0% correct (Table 1, Figure 2A) and the mean RM-NET post-test accuracy scores was 78.7 ± 14.3% correct (see Table 1). There was overlap between the NC and MCI groups in the Lifespan and post-test accuracy scores, suggesting there is a behavioral continuum across the groups. RM-NET Time Period accuracy scores are shown in Table 2. For the 160 items queried by the RM-NET post-test which spanned 1-30 years prior to testing there were very few items that were accompanied by autobiographical memories (2.6 ± 3.1 items; 1.6%).

Figure 2. RM-NET accuracy scores.

A. Swarm and violin plot of RM-NET Lifespan accuracy scores, color coded separately for participants with normal cognition (NC; black) or mild cognitive impairment (MCI; red [grey]). B. The MCI group was impaired on the RM-NET relative to the NC group. C. RM-NET time period accuracy scores for NC (black) and MCI (red [grey] groups). The extent of retrograde amnesia in the MCI group extended back to memories that were 45 years old. The groups performed similarly for older memories. Memory age refers to the approximate age of the memories queried. Error bars indicate SEM.

Though there was overlap in the RM-NET Lifespan scores across the NC and MCI groups (Figure 2A), we examined if there were significant group differences in these scores. The MCI group’s Lifespan accuracy score was significantly worse than the score in the NC group (p < 0.001; Figure 2B). To identify if impairment in the MCI group was related to the age of the news event memory, we conducted a repeated measure Time Period by Group ANCOVA. There was a significant two-way cubic interaction of Group (MCI vs. NC) by Time Period (F_(1,37) = 4.445, p = 0. 039), which is consistent with the idea that MCI produced temporally-limited retrograde amnesia (see Figure 2C). Follow-up tests on individual RM-NET time periods were carried out to estimate the extent of retrograde memory impairment in the MCI group relative to the NC group. The MCI group exhibited significant impairment that extended from 1-45 years prior to testing (p values < 0.026; Figure 2C). Accuracy scores for older events were not significantly different between the MCI and NC groups (p values > 0.680). See Asp et al. (Asp et al., 2024)for further analysis of the RM-NET in the NC and MCI groups. The RM-NET post-test also revealed group differences (p < 0.001) between the NC (86.7 ± 2.1% correct) and MCI (71.2 ± 2.0% correct) groups.

PLS Behavior-Brain Associations: RM-NET Lifespan Accuracy Score

Volume.

One significant latent variable (LV) was identified (r = 0.56, p < 0.001 , variance accounted for = 100%; see Figure 3A and Table 3) that reflected a predominantly left-sided network of regions primarily in the medial and lateral temporal lobe (including hippocampus), lateral parietal lobe, and lateral prefrontal cortex. The behavior-brain correlation for these regions was r = 0.560 and the correlation was not significantly stronger (p = 0.176) in the MCI group (r = 0.635) versus the NC group (r = 0.390).

Figure 3. Behavioral-PLS results for the RM-NET.

Brain regions where the RM-NET significantly predicted cortical and hippocampal volumes and cortical thickness in older adults (N = 70). A. Brain regions identified by the Lifespan LV (light blue), the Time Period LV (dark blue), or both LVs (blue) for volumes. B. Brain regions identified by the Lifespan LV (yellow), the Time Period LV (red), or both LVs (orange) for thickness. C. Left. Scatter plot showing the significant correlation between behavioral (RM-NET Lifespan Score) and brain scores (Network Brain Volume Score) for regions identified by the Lifespan Volume LV (light blue and blue regions in panel A). Individual data points reflect whether participants had normal cognition (NC, black) or mild cognitive impairment (MCI, red). C. Right. Scatter plot showing the correlation coefficient for each Time Period (brain-behavior correlations) as a function of memory age for regions identified by the Time Period Volume LV (blue and dark blue regions shown in A). Correlation coefficients followed a cubic function. D. Left. Same as C (left), but for regions identified by Lifespan Thickness LV (yellow and orange regions in panel B). D. Right. Same as C (right), but for regions identified by the Time Period analyses in panel B (orange and red regions). Correlation coefficients decreased according to a linear function. Brain regions are displayed using the Brainnetome Atlas, with the hippocampus shown on the medial surfaces of the brain on panel A. LVs and BSRs thresholded at p < 0.025, BSRs > 2.24.

Table 3.

Associations between regional cortical and hippocampal volumes and the RM-NET

Brain Region	Brainnetome Region
	Lifespan	Time periods
Frontal Ctx.
Sup. Frontal G.	L A9l, L A10m*	Lifespan regions
Mid. Frontal G.
Inf. Frontal G.	L A44d*, L A44op*, L A44v**, L IFS, L A45c*, L A45r	Lifespan regions
Orbital G.	L A12/47l*, R A12/47o, R A14m, B A11m (L*, R**)	Lifespan regions
Precentral G.	L A4hf*, L A4tl*, L A6cvl*	Lifespan regions
Paracentral Lob.
Insular Ctx.
Insular G.	L G**, L vId/vIg, L dId**	Lifespan regions
Limbic Ctx.
Ant. Cingulate Ctx.
Post. Cingulate Ctx.		L A23c, R A23d
Lateral Temporal Ctx.
Sup. Temporal G.	B A38m (L**, R**), B A38l (L*, R**)	Lifespan regions
Mid. Temporal G.	L aSTS, B A21c (L*, R*), B A21r (L, R*)	Lifespan regions + R A37dl
Inf. Temporal G.	L A20r, R A37vl, B A20il (L, R), B A37elv (L, R*)	Lifespan regions
Fusiform G.	L A20rv**, L A37mv**, B A37lv (L*, R)	Lifespan regions + R A20rv
Post. Sup. Temporal S.
Medial Temporal Ctx.
Parahippocampal Ctx.	L TI *	Lifespan regions
Entorhinal Ctx.	L A28/34 *	Lifespan regions
Perirhinal Ctx.	B A35/36r (L*, R**)	Lifespan regions
Parietal Ctx.
Sup. Parietal Lob.	R A5l**	L A7pc
Inf. Parietal Lob.	L A39c, L A40rv, R A39rv*	Lifespan regions + R A39rd
Postcentral G.	L A2	Lifespan regions
Occipital Ctx.
Med. Vent. Occipital Ctx.	L cCunG*, R cLinG*	Lifespan regions
Lat. Occipital Ctx.	L msOccG, R iOccG, B V5/MT+ (L, R)	Lifespan regions
Subcortical
Hippocampus	N/A, L**, R*	N/A, Lifespan regions

Note. Lifespan: regions unique to the Lifespan analysis and that were not identified by the Time Period analysis are underlined. Time periods: regions identified by the Time Period analysis were very similar to the regions identified by the Lifespan analysis. Therefore, common regions are represented by the ‘Lifespan regions’ label and only the regions uniquely identified by the Time Period analysis are listed. Regions that were not identified by an analysis were intentionally left blank. Bolded regions are unique to the RM-NET (Lifespan and time period analyses combined) relative to the traditional cognitive measures (see Supplemental Table 3). Unbolded regions overlapped with regions identified by analyses of traditional cognitive measures. All regions listed below are p < 0.025.

^*p < 0.01,

^**p < 0.001.

Lob. = lobe, Ctx. = cortex, G. = gyrus, S. = sulcus, Ant. = anterior, Post. = posterior, Sup. = superior, Mid. = middle, Inf. = inferior, Lat. = lateral, L = left, R = right, B = bilateral. For the abbreviation of the Brainnetome regions, see (Fan et al., 2016). The hippocampal volume was obtained using FreeSurfer, not the Brainnetome atlas, therefore, N/A indicates there is no Brainnetome label.

Cortical Thickness.

One significant LV was identified (r = 0.54, p < 0.0001, variance accounted for = 100%) (see Figure 3B and Table 4). Unlike the findings for cortical volume, the identified network was largely bilateral, consisting of medial and lateral temporal and prefrontal cortex, and lateral parietal cortex. The behavior-brain thickness correlation for these regions was r = 0.420 and the correlation was significantly stronger (p = 0.019) in the MCI group (r = 0.557) than in the NC group (r = 0.04).

Table 4.

Associations between regional cortical thickness and the RM-NET

Brain Region	Brainnetome Region
	Lifespan accuracy scores	Time periods
Frontal Ctx.
Sup. Frontal G.	R A9m*, R A10m	Lifespan regions
Mid. Frontal G.	L A9/46v, R A6vl	Lifespan regions
Inf. Frontal G.	B A45r (L*, R*)	Lifespan regions
Orbital G.	R A14m**, R A12/47o**, R A11m**, B A11l (L*, R*), B A13 (L, R*)	Lifespan regions + L A12/47l
Insular Ctx.
Insular G.	R vIa*, R dIa*, R dId**, B G (L, R*), B vId/vIg (L**, R**)	Lifespan regions
Limbic Ctx.
Ant. Cingulate Ctx.	L A32p*, L A32sg*	Lifespan regions
Lateral Temporal Ctx.
Sup. Temporal G.	L TE1.0/TE1.2*, B A38m (L**, R**), B A38l (L**, R**)	Lifespan regions + R TE1.0/1.2
Mid. Temporal G.	L A37dl*, R aSTS**, B A21r (L**, R**)	Lifespan regions
Inf. Temporal G.	L A20r*, R A20iv, B A20il (L, R*)	Lifespan regions
Fusiform G.	L A37mv, L A37lv, B A20rv (L, R)	Lifespan regions
Medial Temporal Ctx.
Parahippocampal Ctx.	R TL, R TI	Lifespan regions
Entorhinal Ctx.	L A28/34*	Lifespan regions
Perirhinal Ctx.	L A35/36c*, R A35/36r**	Lifespan regions
Parietal Ctx.
Inf. Parietal Lob.	B A40rv (L*, R)	Lifespan regions
Postcentral G.	L A1/2/3tonIa	Lifespan regions

Note. Lifespan: regions unique to the Lifespan analysis and that were not identified by the Time Period analysis are underlined. Time periods: regions identified by the Time Period analysis were very similar to the regions identified by the Lifespan analysis. Therefore, common regions are represented by the ‘Lifespan regions’ label and only the regions uniquely identified by the Time Period analysis are listed. Regions that were not identified by an analysis were intentionally left blank. Bolded regions are unique to the RM-NET (Lifespan and time period analyses combined) relative to the traditional cognitive measures (see Supplemental Table 3). Unbolded regions overlapped with regions identified by analyses of traditional cognitive measures. All regions listed below are p < 0.025.

^*p < 0.01,

^**p < 0.001.

Lob. = lobe, Ctx. = cortex, G. = gyrus, S. = sulcus, Ant. = anterior, Post. = posterior, Sup. = superior, Mid. = middle, Inf. = inferior, Lat. = lateral, L = left, R = right, B = bilateral. For the abbreviation of the Brainnetome regions, see (Fan et al., 2016).

Behavior-Brain Associations: RM-NET Time Periods

Volume:

One significant LV was identified, showing a decrease in brain-behavior correlations as the age of the memory increased (p < 0.001, 78% variance accounted; see blue and dark blue regions in Figure 3A and Table 3). The network of brain regions identified by the Time Period analysis was highly overlapping with the regions identified by the Lifespan analysis (see blue regions in Figure 3A). Additional regions were identified in medial and lateral parietal cortex (see dark blue regions in Figure 3A). The changes in correlation coefficients followed a cubic function (adjusted R² = 0.699, p < 0.001, see Figure 3C, right), indicating that correlation coefficients remained relatively unchanged for memories aged 1-45 years then abruptly decreased for older memories.

Cortical Thickness:

One significant LV was identified (p = 0.0139, 75.9% variance accounted; see orange and red regions in Figure 3B and Table 4), showing a decrease in brain-behavior correlations as the age of the memory increased Like the findings for volume, the network of regions was highly overlapping with the regions identified by the Lifespan analysis (see orange regions in Figure 3B). Additional regions were identified in right superior temporal gyrus and left inferior-lateral prefrontal cortex (see red regions in Figure 3B, Table 4). Unlike the findings for volumes, the changes in correlations followed a linear function (r = −0.939, p < 0.001; see Figure 3D, right) indicating that correlations coefficients consistently declined as memory age increased.

Secondary Analyses

To identify the brain regions uniquely identified by the RM-NET (i.e., Figure 3A for volume [Lifespan and Time Period regions] and 3B for thickness [Lifespan and Time Period regions]) and not by other cognitive tests, we carried out similar analyses but instead used traditional cognitive tests to identify brain regions. We first identified regions where cortical and hippocampal volume or cortical thickness correlated with each of the cognitive composite scores. For these analyses, significant LVs were identified only for episodic memory (r = 0.42, p = 0.0054, 100% variance accounted) and attention/processing speed (r = 0.39. p = 0.0129, 100% variance accounted). For cortical and hippocampal volumes, significant LVs were also identified only for episodic memory (r = 0.49, p = 0.0247, 100% variance accounted) and attention/processing speed (r = 0.50, p = 0.0131, 100% variance accounted). These behavior-brain correlation coefficients were not significantly different for the MCI and NC groups (p values ≥ 0.250). Brain regions identified either by of these domains were combined into one network that reflected regions that support performance on traditional neuropsychological tests in this cohort (see Supplemental Table 3). Note that RM-NET performance significantly correlates with only measures of episodic memory and attention and processing speed (Cawley-Bennett et al., 2022), suggesting that this traditional neuropsychological test network is suitable for identification of unique regions in the RM-NET network.

Next, we identified which regions in the RM-NET network (regions identified by either the Lifespan or Time Period analyses) overlapped with regions in the traditional test network. For cortical and hippocampal volumes, the RM-NET identified unique regions bilaterally in the medial and lateral temporal lobes, medial and lateral parietal cortex, and in left prefrontal cortex (Figure 4A, dark blue). For cortical thickness, the RM-NET identified unique regions bilaterally in the posterior MTL, lateral temporal and parietal lobe, and the left prefrontal cortex (Figure 4B, dark red). See bolded regions in Tables 3 and 4 for a complete list of regions unique to the RM-NET for volume and cortical thickness, respectively.

Figure 4. Behavioral-PLS results for the RM-NET and Traditional Neuropsychological Tests.

Brain regions where the RM-NET (as in Figure 3) or traditional tests significantly predicted cortical and hippocampal volumes (A) and thickness (B) across all participants (N = 70). The RM-NET identified unique regions that were not identified by traditional neuropsychological tests (dark blue, dark red). The traditional tests reflect the domains of episodic memory and attention. Brain regions are displayed using the Brainnetome Atlas, with the hippocampus shown on the medial surfaces of the brain on panel A. LVs and BSRs thresholded at p < 0.025, BSRs > 2.24.

We also identified brain regions that were similar or unique to retrograde versus anterograde memory. Specifically, we carried out a similar analysis as described above, but compared the networks associated with retrograde memory (i.e., the RM-NET networks in Figure 3A and 3B) with the networks associated with the RM-NET post-test scores, which reflect episodic anterograde memory. Analysis of RM-NET post-test scores revealed one significant LV (r = 0.55, p = 0.003, 100% variance accounted for) for cortical and hippocampal volumes (see Figure 5A and Table 5), and one significant LV (r = 0.41, p = 0.0055, 100% variance accounted for) for cortical thickness (see Figure 5B and Table 5). Cortical and hippocampal volume behavior-brain correlations were not significantly different (p = 0.472) between MCI (r = 0.510) and NC (r = 0.365) groups, whereas the correlations were significantly different (p = 0.038) between MCI (r = 0.405) and NC (r = −0.088) for cortical thickness. Next, we identified regions unique to retrograde memory. For volume, unique regions were located bilaterally in medial and lateral temporal cortex, parietal cortex, and in left prefrontal cortex. (Figure 5A, dark blue regions). For thickness, unique regions were located primarily in lateral temporal and parietal cortex, right parahippocampal cortex, and medial prefrontal cortex (Figure 5B, red regions).

Figure 5. Behavioral-PLS results for the RM-NET and the Post-Test.

Brain regions where the RM-NET Lifespan or Time Period accuracy scores (as in Figure 3) or the RM-NET Post-Test significantly predicted cortical and hippocampal volumes (A) or cortical thickness (B) across all participants (N = 70). Brain regions are displayed using the Brainnetome Atlas, with the hippocampus shown on the medial surfaces of the brain on panel A. LVs and BSRs thresholded at p < 0.025, BSRs > 2.24.

Table 5.

Associations between volume and thickness and the RM-NET post-test

Brain Region	Brainnetome Region
	Cortical volume	Cortical thickness
Frontal Ctx.
Sup. Frontal G.	L A6dl, L A8m, L A9l, L A10m	L A8m**, L A8dl*, L A10m*, R A9m, B A9l (L*, R)
Mid Frontal G.	B A9/46v (L*, R*), R A8vl	L A9/46d*, L A9/46v*, L A46*, L A10l, B IFJ (L*, R**), B A6vl (L*, R)
Inf. Frontal G.	L A45r, L A44v*	L A45c*, L A44op, L A44v, B A45r (L**, R)
Orbital G.	R A11m**, R A13*, R A14m*	L A12/47l*, R A11m*, R A13*, R A14m*, B A12/47o (L, R**), B A11l (L*, R**)
Precentral G.	L A6cvl, R A4t*	L A4hf, B A6cvl (L**, R*)
Paracentral Lob.
Insular Ctx.
Insular G.	L G**	L G, L vId/vIg**, R dId*
Limbic Ctx.
Ant. Cingulate Ctx.		L A32p*, L A32sg**
Post. Cingulate Ctx.	L A23v*, R A23c*	R A23c *
Lateral Temporal Ctx.
Sup. Temporal G.	R A38l**, B A38m (L*, R**)	L A22r, L TE1.0/1.2*, R A38l**, B A38m (L, R*)
Mid. Temporal G.	L A21r*, L A21c*	L A37dl, B A21r (L**, R*), B aSTS (L*, R**)
Inf. Temporal G.	L A20r, L A20cl*	L A20r*, L A20il
Fusiform G.	L A20rv, L A37mv*, L A37lv*	L A20rv, L A37lv*
Post. Sup. Temporal S.		R rpSTS*, R cpSTS
Medial Temporal Ctx.
Parahippocampal Ctx.	R TH *	R TI
Entorhinal Ctx.		L A28/34**
Perirhinal Ctx.	B A35/36r (L**, R**)	L A35/36c, B A35/36r (L*, R**)
Parietal Ctx.
Sup. Parietal Lob.	R A5l*
Inf. Parietal Lob.	L A39c	L A40rd, L A40rv*
Precuneus	R A7m *	L A31 *
Postcentral G.		L A1/2/3tonIa*
Occipital Ctx.
Med. Vent. Occipital Ctx.	L cCunG**, L rCunG, L vmPOS*, R cLinG**
Lat. Occipital Ctx.	R mOccG**, R lsOccG*, R msOccG	R V5/MT+
Subcortical
Hippocampus	N/A, L**, R**

Note. The bolded Brainnetome regions are those unique to the RM-NET post-test versus the RM-NET, i.e. Table 3 for cortical and hippocampal volume and Table 4 for cortical thickness. All regions listed are below p < 0.025.

^*p < 0.01,

^**p < 0.001.

Lob. = lobe, Ctx. = cortex, G. = gyrus, S. = sulcus, Ant. = anterior, Post. = posterior, Sup. = superior, Mid. = middle, Inf. = inferior, Lat. = lateral, L = left, R = right, B = bilateral. For the abbreviation of the Brainnetome regions, see (Fan et al., 2016). The hippocampal volume was obtained using FreeSurfer, not the Brainnetome atlas, therefore, N/A indicates there is no Brainnetome label.

Discussion

We identified brain regions where brain volume or cortical thickness was significantly associated with news event memory in older adults. The regions identified were primarily in the medial and lateral temporal and parietal lobes bilaterally, as well as left prefrontal cortex (Figure 3A-B and Tables 3–4). Most of these regions exhibited decreasing associations with news event memory as the age of the memory increased (Figure 3C and 3D, right). We also identified which of these regions were unique to news event memory and which regions overlapped with those associated with performance on traditional neuropsychological tests (Figure 4A-B). Finally, we identified regions unique to retrograde memory (RM-NET Lifespan and Time Period scores) versus anterograde memory (RM-NET post-test scores, Figure 5), by comparing results from the two measures of memory that can be computed from the RM-NET.

Structural Neuroanatomy of Retrograde Memory for News Events

Frontal lobe damage and MTL/LTL damage have been associated with retrograde amnesia in patients with focal brain damage from various etiologies (Bayley et al., 2006). In concordance with these findings, we found that news event memory was correlated with the structure of MTL, LTL, and prefrontal cortex, particularly for cortical thickness. An earlier study in MCI (N=11) found significant correlations between news event memory and thickness/volume in bilateral hippocampus, left parahippocampal cortex, left posterior cingulate cortex, and bilateral paracentral lobules, with strong trends in lateral temporal cortex bilaterally (Smith, 2014). In addition to these findings for cortical thickness/volume, three studies examined grey matter intensity in MCI using voxel-based morphometry and news event memory (Barbeau et al., 2012; Serra et al., 2022) or famous personalities (Gardini et al., 2013). Like the current findings, these studies also identified significant brain-behavior correlations in the MTL, LTL, and prefrontal cortex. Finally, studies of brain activity associated with news events/famous personalities have found these regions are related to memory retrieval (Douville et al., 2005; Haist et al., 2001; Smith & Squire, 2009; Woodard et al., 2007).

The concordant findings from the above studies of the functional and structural neuroanatomy of semantic retrograde memory suggest a common network for retrograde memory retrieval. First, the MTL is well known for its role in memory retrieval, especially for retrieval of recent (but not remote) semantic memory, consistent with our findings that the hippocampus and parahippocampal cortices exhibited decreases in brain-behavior correlations with memory age. Second, in contrast to the MTL, the LTL is thought to store the content of semantic memory because patients with LTL volume reduction exhibit dense (ungraded) retrograde amnesia (Bright et al., 2006; Squire & Wixted, 2015). This idea is consistent with our finding that left middle temporal cortex volume (i.e., rostral BA 21 and anterior STS) was one of the few regions that did not exhibit decreases in brain volume-behavior associations with memory age (Table 3, Figure 3A). Finally, regions in the prefrontal frontal cortex are thought to guide memory retrieval/search, as well as support verbal working memory, a type of cognition that is strongly associated with performance on the RM-NET (Cawley-Bennett et al., 2022). Our robust findings of left prefrontal cortex volume correlations with RM-NET accuracy scores (Table 3, Figure 3A), together with the findings that many of the regions uniquely associated with the Lifespan score were in medial or lateral prefrontal cortex (light blue and yellow regions in Figure 3A and 3B, respectively), suggest that prefrontal cortex is additionally needed to retrieve news event memory, regardless of memory age.

Structural Neuroanatomy of Retrograde Memory and Memory Age

When comparing the network of regions associated with RM-NET Lifespan and Time Period accuracy scores, most regions were common to both analyses, suggesting that the bulk of the retrograde memory network exhibited decreases in brain-behavior correlations with increasing memory age. A striking difference between the findings for the volume and thickness outcomes was the way the brain-behavior correlations changed over time. For volume, the correlations were relatively stable for memories aged 1-45 years and then decreased precipitously for more remote memories (Figure 3C, right). By contrast, the correlations for thickness followed a gradual, linear decrease across all the time periods (Figure 3D, right), steadily weakening with memory age. For both volume and thickness, the strength of the behavior-brain relationships started near 0.4 for the most recent memories and ended near zero for the most remote memories. What differed between these measures was the trajectory taken as memories aged from 1 – 55 years. Thus, measures of cortical thickness would be better candidates for examination of the effects of memory age on brain structure. Additionally, our findings suggest that measures of brain volume are robust to changes in memory age for almost all but the oldest populations.

It was surprising that the hippocampus was included in a network that exhibited decades-long positive correlations with memory retrieval, given that its role in semantic memory retrieval is thought to be limited to about 10 years (Bayley et al., 2006; Manns et al., 2003), in patients with acute lesions limited to the hippocampus. It is possible that a small subset of regions, including the hippocampus may exhibit a different relationship between volume and memory age, but there were not enough of them to yield a significant LV in our whole-brain analysis approach. It may also be the case that when there is damage to both the hippocampus and parahippocampal gyrus (see Supplemental Table 4 and Supplemental Figure 1 for damage in the MCI group vs. the NC group) it is sufficient to cause decades-long retrograde amnesia. This idea is supported by findings from patients with large acute lesions of the hippocampus and parahippocampal gyrus due to viral encephalitis (but with minimal damage outside the MTL) who were tested with a precursor to the RM-NET (Bayley et al., 2006). Similar to patients with MCI (Asp et al., 2024; Benoit et al., 2017; Langlois et al., 2016; Smith, 2014), patients with large acute MTL lesions exhibit retrograde amnesia affecting news event memories formed decades prior to injury and also exhibit intact memories for events formed during young adulthood. This idea of the importance of combined hippocampus and parahippocampal gyrus injury is also supported by findings from patients with AD, who exhibit retrograde amnesia affecting all time periods (including young adulthood) (Beatty et al., 1988; Benoit et al., 2017; De Simone et al., 2020; Langlois et al., 2016; Leyhe et al., 2010; Sagar et al., 1988). This neurodegenerative group exhibits less damage inside the MTL but more damage outside the MTL than the acute encephalitic patients, suggesting that it is the regions outside the MTL that support retrieval of the most remote memories. Beyond these earlier studies, our findings reveal the intricate network of regions (including the MTL) that support retrieval of memories decades after they are formed, with damage to the hippocampus and parahippocampal gyrus together being sufficient to cause severe, but temporally-limited retrograde amnesia. Examination of the brain-behavior temporal dynamics in MTL subregions, including hippocampal subregions, will elucidate the individual contributions of MTL regions to performance on the RM-NET and will be reported in a separate publication.

Structural Neuroanatomy of News Event Memory versus Other Cognitive Tests

It could be the case that any cognitive test, particularly episodic memory tests, may depend on the network identified by the RM-NET. If so, then performance on RM-NET would not identify unique regions that support memory retrieval. Previous studies found that the performance on traditional neuropsychological assessments was associated with measures cortical volume or thickness in the MTL, LTL, lateral parietal lobe, posterior cingulate cortex, and regions in the medial/lateral prefrontal cortex (Ahn et al., 2011; Amaefule et al., 2021; Cheng et al., 2018; Csukly et al., 2016; Hartikainen et al., 2012; Julkunen et al., 2009; Kang et al., 2019). Although the RM-NET identified regions in these areas as well (see Figure 4, light blue and light red), the RM-NET also identified additional unique regions, primarily in the MTL, LTL, and medial/lateral parietal lobe bilaterally, as well as left medial prefrontal cortex and lateral prefrontal cortex bilaterally (see Figure 4, dark blue and dark red). Of all the RM-NET regions, 69.2% were unique to the RM-NET for volume network and 41.5% were unique to the RM-NET for the thickness network. One previous study examined correlations between grey matter intensity (VBM) and performance on a novel semantic test (naming of famous personalities) versus traditional semantic memory/language tasks (semantic fluency and naming(Gardini et al., 2013). Their test of famous personalities identified 13 regions, six (46.2%) of which were unique to the famous personalities task (lingual gyrus, precuneus, cuneus, inferior parietal lobule, posterior cingulate cortex, and superior frontal gyrus). Therefore, the RM-NET, like their famous personalities test, represents a single test where performance depends on regions associated with existing cognitive tests as well as numerous unique regions.

Structural Neuroanatomy of Anterograde versus Retrograde Memory

Given that MTL-dependent episodic anterograde memory tests are used to identify MCI and that retrograde memory is related to the extent of both MTL and LTL damage(Bayley et al., 2006; Bright et al., 2006; Kopelman, 1991; Kopelman et al., 1999; Kroll et al., 1997), we directly compared the networks that supported anterograde versus retrograde memory. The RM-NET anterograde memory analyses (using the post-test) identified regions in the MTL and in the LTL, medial/lateral parietal lobe (thickness only), medial (thickness only)/lateral prefrontal cortex, and hippocampus, as did the RM-NET retrograde memory analyses [see Figure 5, blue (A) and orange (B) regions]. Notably, the retrograde memory analyses also identified unique regions within these large areas [see Figure 5, dark blue (A) and red (B) regions]. Like our findings comparing the RM-NET to traditional neuropsychological tests (see above), there was a high percentage of unique retrograde memory regions, when comparing anterograde and retrograde memory networks, with more overlap for volume than for cortex. Specifically, retrograde memory regions (see Figure 3A,B) overlapped 59.1% and 28.6% for volumes and thickness networks, respectively (see Figure 5A,B).

Only one prior study has examined the overlap between retrograde memory and anterograde memory performance in a small MCI group, and only in a number of a priori ROIs(Smith, 2014). That study found that the hippocampus correlated with both retrograde and anterograde memory scores, while the LTL and posterior cingulate were correlated only with retrograde memory scores. Similar to this study, we found that the hippocampus was correlated with both anterograde and retrograde memory. Although the earlier study examined LTL cortex as one ROI, we confirmed that several brain regions within the LTL exhibited this pattern. Thus, the RM-NET identified unique regions in MTL and LTL cortex, when compared to a novel measure of episodic anterograde memory as well as to traditional measures of cognition. These findings regarding unique regions in LTL for retrograde memory are important because these regions are largely unaffected by normal aging. With a single measure, such as the RM-NET Lifespan score, one can identify the integrity of both the MTL and LTL (particularly the inferior LTL), which together signal high risk for progression to dementia.

News Event Memory in the Context of AD

The retrograde memory network, identified by the RM-NET, appears to overlap with the network of regions that are susceptible to early Alzheimer’s disease. Specifically, medial and lateral temporal and parietal lobes undergo the most cortical thinning when one transitions from normal aging to early AD(Eskildsen et al., 2013), medial and inferior temporal cortex, medial parietal cortex, and medial prefrontal cortex are susceptible to tau deposition in the early stages (stages 3-4) of the disease (Braak & Braak, 1991) when it starts to spread outside of the MTL. Finally, these regions are highly overlapping with those that exhibit abnormal functional activity (i.e., default mode network study (Dante Mantini and Wim, 2013). As a result of these similarities, novel measures of semantic retrograde memory, such as the RM-NET, may be particularly sensitive predictors of cognitive decline and conversion to AD, because they tap into additional structures in inferior temporal lobe and medial parietal lobe that are largely unaffected by normal aging, but are significantly affected in Alzheimer’s disease.

Limitations

There were a few limitations in this study that are important to note. First, there were modest differences in education level and age between the NC and MCI groups, requiring the inclusion of these factors as nuisance covariates. Second, we were unable to identify significant LVs associated with traditional semantic memory/language tests. This limited direct comparisons between the RM-NET and more commonly used tests of semantic memory. Third, this sample lacked racial and ethnic diversity. Finally, the older time periods of the RM-NET (30-55 years) had about half the number of questions relative to more recent time periods (see Table 2). This could have resulted in less variability in the behavioral scores, leading to lower effect sizes for brain-behavior correlations, and potentially driving the significant changes in effect sizes identified by the Time Period PLS analyses. We do not believe this was the case because measures of error (SEM or SD) for these time periods were similar to the more recent time periods (also see Figure 2C), and the time-dependent effects detected by the PLS analyses do not appear to be driven by abrupt changes at the divide between the more recent (1-30 years) and more remote (31-55 years) time periods (see Figures 3C,D). It should be noted that the RM-NET’s time periods with the fewest questions have about the same number of questions (or more questions) than other measures of retrograde memory for equivalent (or larger) time periods.

Summary

In summary, our findings indicate that news event memory is impaired in MCI and depends on a widespread network of brain region that overlap with brain regions that undergo the most thinning when an individual transitions from normal aging to AD. Furthermore, the RM-NET identified unique regions when compared to traditional measures of cognition and of episodic anterograde memory, in particular. This suggests that novel tests of semantic memory, such as the RM-NET, will be useful clinical tools in detecting early risk for AD because they rely on the regions known to decline in early AD, including susceptible regions not captured by traditional tests.

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.