Data-Driven Analyses of Longitudinal Hippocampal Imaging Trajectories: Discrimination and Biomarker Prediction of Change Classes

INTRODUCTION

Hippocampal atrophy is a well-documented anatomical process that typically occurs during brain aging [1–4]. However, aged individuals may vary in several indicators of hippocampal atrophy, including level (e.g., overall volume loss), slope (e.g., rate of volume loss), and associated clinical outcomes (e.g., memory impairment, Alzheimer’s disease (AD)) [1, 5–7]. In a distribution of cognitively normal (i.e., unimpaired or asymptomatic) older adults, hippocampal volume trajectories characterized by relatively lower levels and steeper decline may be suggestive of elevated risk for subsequent clinical transitions to mild cognitive impairment (MCI) or AD [8–10]. Given its heterogeneity in level and change, further studies are required to ascertain and disentangle important features that characterize hippocampal atrophy in cognitively normal aging. Among the considerations are accumulating evidence of hippocampal hemispheric differences that are reflected in volume trajectories and various clinical outcomes [11–13]. For example, left and right hippocampal trajectories have been found to be differentially moderated by sex and APOE (McFall et al., unpublished data). Hemispheric differences in hippocampal subfields have also been observed between clinical cohorts (i.e., normal controls, subjective cognitive decline, MCI, and AD) [14]. We investigated this issue by deploying a sequence of two data-driven analytic approaches (i.e., latent class growth analysis, random forest classification) in parallel for the left (LHC) and right (RHC) hippocampi: 1) objectively discriminating classes within a distribution of individualized volume longitudinal trajectories, and 2) identifying key biomarkers and risk factors that discriminated between the observed classes.

Previous hippocampal atrophy research has been conducted with both cross-sectional (comparing age or clinical groups at one time point) and longitudinal (following groups over two or more time points) designs [3, 7, 9, 15, 16]. Although useful for determining average group differences or mean-level change in multiple domains of asymptomatic brain and cognitive aging, these variable-oriented approaches (i.e., focused on relationships between variables in assumed homogeneous populations) are not typically aimed at scrutinizing the well-established individual heterogeneity in either the level or slope of trajectories [3, 17–19] as compared to person-oriented approaches (i.e., focused on similarities and patterns among individuals in an assumed heterogeneous population) [20]. Recently, the growing interest in examining heterogeneity in brain aging and dementia [21, 22] has led to a corresponding effort to adapt data-driven technologies to the 1) examination of individualized trajectories of cognitive changes in older adults and 2) determination of possible underlying classes of trajectory patterns [19, 21, 23]. These latent classes, which are determined via application of algorithms based on performance intercept (level) and slope (rate of change) parameters [20], may later be clarified by identifying predictors most associated with reduced or exacerbated risk for cognitive decline or clinical impairment [21].

A growing body of neurocognitive aging and dementia research has demonstrated the viability of applying data-driven technologies to model heterogeneity in both cross-sectional and longitudinal (trajectory) distributions, including the identification of detectable asymptomatic classes and the determination of differential biomarker predictors [19, 21, 24]. One such longitudinal example in an AD sample identified atrophy subtypes associated with differing degrees of memory performance [25]. In asymptomatic individuals, three cross-sectional biomarker profile subtypes were extracted from a combination of magnetic resonance imaging (MRI) data and cerebrospinal fluid (CSF) biomarkers [26]. One of these subtypes, similar in biomarker profile to a comparative AD group, was associated with accelerated cognitive decline and lower baseline scores on cognitive tests [26]. Although few studies have explored longitudinal data-driven subtypes [21], separate cross-sectional studies of cognitively unimpaired older adults have previously reported distinct imaging subtypes [27–32]. As both cognitively unimpaired aging and AD are characterized by progressive hippocampal atrophy, the possible presence of detectable longitudinal subtypes of hippocampal trajectories in cognitively normal older adults and their potential associations with AD-related risk factors merit further investigation.

Research on early detection of AD risk in asymptomatic older adults has identified a large number of modifiable and non-modifiable factors (e.g., APOE genetic risk, education, metabolic health, sex) which are associated with increased risk of (or protection from) accelerated cognitive decline, MCI, and AD [33–35]. Similarly, previous studies of normal aging and hippocampal atrophy in normal aging and clinical groups have identified predictors from multiple domains. For example, both traditional CSF AD-related biomarkers, such as baseline p-tau_181p and Aβ_1–42 [36, 37], and disparate lifestyle risk factors as smoking [38] and complex mental activity [39], have been associated with hippocampal atrophy. In addition, three CSF biomarkers [37] have been previously used in a multiple linear regression model to predict longitudinal hippocampal atrophy. Although some recent biomarker reports have featured data-driven technologies applied to large numbers of predictors of AD outcomes [40], longitudinal studies of hippocampal atrophy in cognitively unimpaired older adults have not included a large number of biomarkers or biomarker domains. Previous reports have emphasized the need to include biomarkers from multiple modalities in prediction models over the use of a single biomarker or domain in order to achieve increased prediction accuracy [41, 42].

We aimed to address a knowledge gap regarding hippocampal volume trajectories in cognitively asymptomatic aging. Specifically, the gap refers to the extent to which the heterogeneity of trajectory distributions can be clarified by the detection of underlying longitudinal latent classes and the determination of leading risk factor and biomarker predictors. Because hippocampal hemispheric atrophy differences have been reported both cross-sectionally [13, 43] and longitudinally [44–46], we implemented this aim by testing two main research goals, both of which included parallel analyses of LHC and RHC. For the first research goal (RG1), we analyzed distributions of hippocampal volume trajectories (up to six time points, maximum of 7.2 years) for predominantly cognitively normal (asymptomatic) participants from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). We used latent class growth analyses (LCGA) to detect discriminable classes of trajectories. LCGA is a data-driven longitudinal quantitative modeling technology that applies an algorithm of level and slope to identify statistically separable trajectory classes. Our study focused on a brain aging phase not yet characterized by clinical impairment. Despite normal cognitive function, some individuals may exhibit relatively lower and declining hippocampal volume likely associated with increased risk of future cognitive decline or AD. Notably, membership in higher volume trajectory classes may indicate reduced risk for (or protection from) age-typical morphological shrinkage, membership in lower volume trajectory classes may indicate elevated risk for impending pathological changes. For our second research goal (RG2), we compiled a large, multi-modal set of 38 AD-related biomarkers and risk factors (e.g., CSF Aβ_1–42, body mass index, hypertension, sex) from the ADNI database. Whereas most studies have investigated these factors independently or in relatively small clusters, we examine them simultaneously in the context of a competitive quantitative model. We used random forest analyses (RFA), a machine-learning technology for evaluating the relative importance of multiple biomarker and risk factors predictors to the discrimination of higher and lower classes of LHC and RHC atrophy trajectories.

METHODS

RESULTS

RG1: LHC and RHC trajectory classes

Left hippocampal volume trajectories

Model fit statistics for all analyses are presented by number of classes in Table 3. All tested models had acceptable entropy values (i.e., > 0.8). The two-, three-, and five-class models were selected as possible candidate models as they had lower AIC, BIC and SABIC values than the baseline model and sufficient participants in each class. We selected the three-class model as the final model following the inspection of a scree plot (see Supplementary Figure 1) and in the context of past findings in the related domain of memory aging trajectory analyses [19]. The three-class model is portrayed in Fig. 1c, with parameter means (level and slope) reported in Table 4. Discriminated and ranked by a combination of both level and slope, from highest to lowest volume in the trajectory distribution, the three classes can be characterized as follows. Class 1 (n = 100; the group at the top of the distribution) was characterized by the highest combination of level and slope, followed by Class 2 (n = 173), the group in the middle of the distribution, and Class 3 (n = 78), the group at the bottom of the distribution. Informally, the classes appear to differ more in level than in slope (with Class 2 and 3 having the steeper slopes), but both parameters contributed to the latent class solution. Specifically, the LCGA algorithm identifies distinguishable trajectory classes based on simultaneous consideration of level and slope, both of which are essential parameters in model identification. It is important to note that the resulting trajectory classes are statistically differentiated even though they may not appear visually as dramatically distinct at their edges. This between-class distinction is clearly indicated by the entropy values (revealing good class separation) and the level and slope parameters (and 95% confidence intervals) for each class (see Table 4).

Table 3

Latent class growth analyses model fit statistics and class proportions for left and right hippocampal volume

Volumetric variable	Number of classes	Class proportions	AIC	BIC	SABIC	Entropy
Left hippocampus	1	–	403.50	442.12	410.39	–
	2	0.49/0.51	–909.04	–851.13	–898.71	0.90
	3*	0.49/0.29/0.22	–1907.10	–1829.88	–1893.33	0.92
	4	Did not replicate	–	–	–	–
	5	0.10/0.26/0.22/0.13/0.30	–2707.13	–2591.31	–2686.48	0.89
Right hippocampus	1	–	399.19	437.80	506.08	–
	2	0.46/0.54	–885.82	–827.91	–875.49	0.90
	3*	0.25/0.27/0.48	–1997.35	–1920.14	–1983.58	0.93
	4	0.12/0.34/0.23/0.31	–2450.80	–2354.28	–2433.59	0.92
	5	0.12/0.09/0.36/0.22/0.21	–2765.27	–2649.45	–2744.62	0.90

AIC, Akaike information criteria; BIC, Bayesian information criteria; SABIC, Sample-size adjusted BIC. * Identified as best model fit based on low AIC, BIC, SABIC and no class proportion less than 10%.

Table 4

Final latent class growth analyses models statistics and parameters

Volumetric variable	Class	n (%)	Level (intercept) [95% CI]	Slope [95% CI]
Left hippocampus	1	100 (28.5)	2.50 [2.50–2.51]	–0.02 [–0.025—0.021]
	2	173 (49.3)	2.14 [2.13–2.14]	–0.03 [–0.028—0.024]
	3	78 (22.2)	1.79 [1.78–1.80]	–0.03 [–0.030—0.022]
Right hippocampus	1	96 (27.4)	2.53 [2.53–2.54]	–0.02 [–0.025—0.021]
	2	167 (47.6)	2.21 [2.20–2.21]	–0.03 [–0.028—0.023]
	3	88 (25.1)	1.83 [1.83–1.84]	–0.03 [–0.027—0.023]

Class 1 refers to the higher group; Class 2 refers to the middle group; Class 3 refers to the lower group.

Fig. 1

Distribution of left (1a) and right (1d) hippocampus volume data. Individual trajectories of left (1b) and right (1e) hippocampal volume. Three classes were identified within left (1c) and right (1f) hippocampal volume trajectories: Class 1 (Blue; highest, least atrophied), Class 2 (Green; middle), and Class 3 (Red; lowest, most atrophied). Hippocampal volume was corrected for head size using (hippocampal volume / intra cranial volume) x 10³.

Right hippocampal volume trajectories

Model fit statistics for all analyses are presented by number of classes in Table 3. Similar to the LHC models, all tested models had acceptable entropy values (> 0.8). The four-class model was removed from consideration as the loglikelihood failed to replicate, indicating that no global solution was reached. The five-class model was removed from consideration due to insufficient participants in one class (9%). The two- and three-class models were selected as possible candidate models as they had lower AIC, BIC, and SABIC values than the baseline model and sufficient participants in each class. As with LHC trajectories, we selected the three-class model as the final model based on past findings and inspection of the scree plot of relative fit indices for the inflection point (see Supplementary Figure 2). Thus, we identified three unique classes of RHC volume trajectories within the overall sample (Fig. 1f). Parameter means (level and slope) are reported in Table 4. Discriminated and ranked by a combination of level and slope, from highest to lowest volume in the trajectory distribution, the classes can be characterized as follows. Class 1 (n = 96; the group at the top of the distribution) was characterized by the highest combination of level and decline, followed by Class 2 (n = 167), the group in the middle of the distribution, and Class 3 (n = 88), the group at the bottom of the distribution. Comparable to the LHC trajectory class distribution, the classes appear to differ in level more than slope; however, both parameters contributed to the latent class solution. Informally, the level (but not slope) of each RHC class appears to be consistently higher than that of the corresponding LHC class.

RG2: Important predictors of LHC and RHC class membership

We performed RFA to identify biomarkers and risk factors that best discriminated between the highest (Class 1) and lowest (Class 3) trajectory classes within LHC and RHC volume separately.

Left hippocampal volume trajectory classes

The higher and lower LHC volume trajectory classes were discriminated by seven biomarkers and risk factors from four modalities: biospecimen (plasma Aβ_1–40, plasma tau, plasma Aβ_1–42), demographic (sex, education), co-morbidities (geriatric depression scale [GDS] score), and lifestyle (body mass index; C = 0.80; Fig. 2a). As informed by post-hoc correlational analyses, we found that individuals belonging to the lower LHC volume trajectory class were more likely to have lower levels of plasma Aβ_1–40, Aβ_1–42, and tau, greater number of years of education, higher GDS scores (indicating more depressive symptoms), a lower BMI, and be male (see Table 5 for biomarker/risk factor frequencies and means per class).

Fig. 2

Variable importance (permutation accuracy) in the discrimination of the (2a) lowest (n = 78) versus highest (n = 100) classes of left hippocampal volume trajectories (C = 0.80, ntree = 1000, mtry = 4), and (2b) lowest (n = 88) versus highest (n = 96) classes of right hippocampal volume trajectories (C = 0.78, ntree = 1000, mtry = 4). GDS, Geriatric Depression Scale score; BMI, body mass index; APOE, Apolipoprotein E genotype; MH, medical history; ADAS-Cog, Alzheimer’s Disease Assessment Scale-Cognitive Subscale; CSF Aβ_1–42, cerebrospinal fluid amyloid β_1–42; CSF t-tau, cerebrospinal fluid total tau; CSF p-tau, cerebrospinal fluid phosphorylated tau; MMSE, Mini-Mental State Examination score.

Table 5

Biomarker and risk factor means and frequencies for LHC and RHC trajectory classes

Significant biomarker	Lowest LHC trajectory class	Highest LHC trajectory class	Lowest RHC trajectory class	Highest RHC trajectory class
N	78	100	88	96
Plasma Aβ1–40	139.72 (56.78)	171.46 (47.03)	142.23 (47.19)	168.31 (45.30)
Sex (%, female)	33.33	64.0	31.82	69.80
Plasma t-tau	2.41 (0.94)	2.65 (1.05)	2.50 (1.42)	2.55 (1.07)
Plasma Aβ1–42	34.71 (10.58)	41.00 (14.62)	34.35 (10.13)	42.04 (14.52)
Education, y (SD)	17.15 (2.42)	15.73 (2.56)	17.17 (2.43)	15.33 (2.73)
GDS	0.91 (1.27)	0.52 (0.88)	0.81 (1.19)	0.67 (1.01)
BMI	26.06 (4.47)	27.35 (4.69)	26.11 (4.43)	27.36 (5.07)
Follow-up cognitive status documentation
# of person-waves (observations)	398	496	442	473
% of person-waves that are non-AD	98.2	99.8	98.6	100
% of person-waves that are non-AD and non-MCI	93.0	98.6	92.5	98.7

Right hippocampal volume trajectory classes

The higher and lower RHC volume trajectory classes were discriminated by three biomarkers and risk factors from the following two modalities: demographic (sex, education) and biospecimen (plasma Aβ_1–42; C = 0.78; Fig. 2b). As informed by post-hoc correlational analyses, we found that individuals belonging to the lower RHC trajectory class were more likely to be male, have lower levels of plasma Aβ_1–42, as well as have greater number of years of education (see Table 5 for biomarker frequencies and means per class).

DISCUSSION

This study applied data-driven technologies to longitudinal imaging data to 1) extract computationally separable classes based on individual level and slope from LHC and RHC trajectory distributions and 2) subsequently identify key AD-related biomarkers and risk factors that discriminate between the higher and lower trajectory classes. To our knowledge, no previous study has used these technologies to 1) identify trajectory classes based on separate LHC and RHC volume change in a sample of predominantly cognitively normal older adults and 2) assemble and test a large pool of putative biomarker and risk factor predictors of trajectory class.

Overall, the class structures (number and membership) and constituent trajectory characteristics (levels and slopes) for the two hemispheres were similar. One exception is that RHC volumes appeared consistently higher (in level) for each corresponding class. This RHC advantage is consistent with previous research indicating that RHC volumes are generally more preserved at corresponding ages than LHC volumes in cognitively normal older adults [43, 46, 83]. Our results provide a new and discriminating indicator of this advantage; namely, the advantage can be observed at all corresponding classes (higher, middle, and lower) of aging change. For both hemispheres, the slope means across classes were relatively similar; however, the two lowest classes (middle, lowest) exhibited steeper slopes than the highest class. This pattern was expected as the current sample consisted of uniformly cognitively normal older adults at baseline and who remained clinically non-impaired over 96% of the analyzed longitudinal observations. Notably, even in the more limited heterogeneity of a cognitively unimpaired older adult sample (as compared to a more clinically diverse sample), our analytic approach detected discriminable classes of HC volumetric change. In addition, although there was some overlap between the participants classified into the LHC and RHC classes, there were a substantial number of individuals (n = 93) who were uniquely classified (e.g., were in the lowest LHC but not the lowest RHC) in the two hemispheric analyses. These findings provide further evidence for the consideration of LHC and RHC differences in future research.

As increasing hippocampal atrophy is associated with incipient clinical progression [8, 9, 84], two potential implications of our data-driven latent class approach could be considered. First, these classes of hippocampal trajectories could be provisionally considered as “secondary phenotypes” of brain aging in that they 1) differ in objective and salient brain aging trajectory characteristics and 2) may be associated with differential outcomes or clinical phenotypes such as cognitive impairment or AD. A post-hoc informal check of the current data revealed that cognitive performance over time decreased in a stepwise manner across hippocampal trajectory classes (see the Supplementary Material for ADAS-Cognition and ADNI Memory Composite scores by wave). In addition, higher scores on the CDR were somewhat more prevalent in the lowest classes and none of the participants with a CDR of 1 were classified in the highest trajectory classes. Similarly, a recent study identifying four spatiotemporal trajectory subtypes of tau deposition found that longitudinal MMSE outcomes differed between subtypes [85]. The interpretation was that data-driven groups based on other AD-related biomarkers (tau) have also identified differences in cognitive trajectories [85]. Taken together, the present and complementary findings chart an important direction for future research, in which studies with comprehensive clinical outcome information could provide insights into AD or impairment risk based on long-term pre-clinical trajectory class membership. Second, members of higher trajectory classes may have lower exposure to AD risk factors. We investigated these implications in the next research goal by testing associations with AD biomarkers and risk factors.

Accordingly, we tested predictor importance for a roster of 38 multi-modal AD risk factors and biomarkers. The machine learning technology (RFA) evaluated the relative importance of all of the predictors in a quantitatively competitive context. The leading predictors of extreme classes (higher versus lower) were thus identified for their prediction importance with both independent and interactional contributions considered. The present prediction models do not establish mechanisms of association, but instead identify the risk factors that emerge in data-driven analyses from a large panel of potential predictors and thereby point to promising future directions of both validation and mechanistic research. The full roster of predictors was presented earlier and listed (by modality) in Table 2. Three aspects of the results are discussed: 1) the subset of predictors that were observed for both LHC and RHC, 2) any predictors that were selectively associated with either hemisphere, and 3) notable predictors (e.g., factors that have been associated in candidate biomarker studies) that did not emerge in the present analyses. In all cases, we refer to any available candidate biomarker and risk factor literature to establish the context. Three important predictors from two modalities were robust across the hemispheres: demographic (sex, education) and biospecimen (plasma Aβ_1–42). Four additional predictors were observed selectively in the LHC analyses. We characterize the three common predictors briefly and then discuss the unique predictors for LHC.

Regarding predictors in common for LHC and RHC classes, the sex factor indicated that being male was associated with membership in the lower trajectory classes. For hippocampal atrophy in cognitively unimpaired aging, a common result is that, for given ages, males experience more overall atrophy than females [86]. Our results conducted separately on LHC and RHC extend this pattern to both hemispheres. As an illustration, for both LHC and RHC we noted that membership of the upper (less atrophied) class was predominantly female (64–70%) whereas the lower class membership was predominantly male (66.7–68.2%). Notably, our current multimodal approach highlights the importance of sex relative to other established AD biomarkers and risk factors in predicting differential hippocampal atrophy. This female advantage is concordant with 1) findings in the cognitively asymptomatic aging literature, whereby cognitively normal females often perform at higher the levels than males, and 2) our post-hoc check regarding cognitive trajectories for this sample (see Supplementary Material). Specifically, mean memory scores for the lowest HC trajectory classes (predominantly male) were lower than for the highest trajectory classes (predominantly female), which is consistent with the growing evidence of a male disadvantage in asymptomatic memory aging [24, 87, 88]. However, it should be noted that this female advantage may be reversed in persons living with AD or even preclinical AD. For example, studies have found that females with AD exhibit more rapid hippocampal atrophy [89] and similar associations have been reported for females with AD-related neuropathology [44]. In contrast, we found that in predominantly cognitively unimpaired individuals, men made up a higher proportion of the hippocampal trajectory class characterized by the lowest level and steepest decline (i.e., most atrophy). Thus, future research can aim to resolve whether there is 1) a selectively accelerated rate of hippocampal volume loss for preclinical and clinical (where AD-related neuropathology, such as low CSF Aβ₄₂ levels, would be evident) females or 2) some other factor accounts for the contrasting observations.

More years of education was associated with the lowest (most atrophied) classes of both LHC and RHC volume trajectories. In cognitively unimpaired older adults, non-significant cross-sectional associations between hippocampal size (volume and thickness) and education have been reported [90, 91]. In contrast, education has been previously identified as a potential protective factor in the AD epidemiological literature [92]. Longitudinal findings regarding associations with cognitive reserve (including education) have also been mixed [93, 94]. These inconsistencies may originate from a number of study-related differences, including: 1) design (cross-sectional versus longitudinal), 2) measurement (years of schooling versus attainment), 3) cohort (education differing across generations), 4) study sample (cognitively normal versus clinical; higher versus lower education), 5) analytic approaches (most often single variable versus multi-variable prediction models), 6) study role (correlate, covariate, and even AD protective factor), and 7) outcome (cognitive differences/changes, brain differences/changes). In the current ADNI sample, the majority of participants were relatively highly educated (M years of total schooling at baseline = 16.3). Previous findings regarding the moderation of hippocampal volume by education [95] indicate that these effects are diminished among those with higher education attainment. A relevant previous result [96] led us to explore whether the commonly used proportional approach to correcting for head size [97–99] could lead to potential overcorrections in volume estimates for highly educated samples. Specifically, the common approach corrects the numerator (hippocampal volume) by the denominator (intracranial volume). In a post-hoc check we observed a positive correlation between intracranial volume and education [96]. We suggest (a) careful monitoring of education effects in cognitively normal brain aging, (b) further specific attention to intracranial HC volume corrections when education levels are high, and (c) increasing attention to education effects in research on other brain regions and related biomarkers (e.g., hippocampal to cortex atrophy ratio [94]).

Lower levels of plasma Aβ_1–42 were associated with the lower trajectory classes for both LHC and RHC. Although a conventional biomarker of AD, Aβ_1–42 has been found to be more strongly related to overall neurodegeneration (versus AD specifically) as increased levels in the brain and decreased levels in CSF also occur in other neurodegenerative diseases [34]. Evidence for brain atrophy associations with plasma levels of Aβ_1–42 have been mixed. For example, higher plasma Aβ_1–42 levels and lower volumes of hippocampal subfields have been linked in older adults with, but not those without, subjective complaints [100]. In a separate study using a large sample of cognitively normal older adults, decreased levels of plasma Aβ_1–42 were associated with smaller hippocampal volumes and increased risk of dementia [101]. Similarly, plasma levels of Aβ_1–42 were found to be lower in amnesic MCI individuals as compared to cognitively normal older adults [102]. Our results contribute to the existing and emerging evidence that 1) lower Aβ_1–42 levels are a detectable biomarker of emerging neurodegeneration (hippocampal trajectory classes) in initially cognitively normal individuals and 2) less invasive biomarker collection procedures (e.g., plasma) provide reliable indicators of this early trend toward neurodegeneration [34, 103].

Four additional predictors discriminated LHC trajectory classes only. From the biospecimen modality, plasma Aβ_1–40 and plasma tau predicted class membership uniquely for the LHC. Specifically, lower levels of both plasma Aβ_1–40 and plasma t-tau were associated with membership to the lowest LHC trajectory class. Our findings support and extend previous reports of lower levels of plasma Aβ_1–40 in preclinical AD and AD-related neurodegeneration [101, 102]. Specifically, our results indicate that lower baseline levels of plasma Aβ_1–40 predict trajectories associated with more left (but not right) hippocampal atrophy prior to detectable disease stages. For plasma t-tau, increased levels have been associated with lower gray matter volumes in Aβ+ (but not Aβ-) older adults [104] as well as higher risk of incident dementia [105]. However, our results suggest that lower plasma t-tau may be differentially associated with “secondary phenotypes” of clustered individuals representing different patterns of longitudinal atrophy in cognitively normal adults. A possible explanation is the potential effect of age on plasma t-tau levels. In a recent study, older adults (compared to middle-aged adults) were found to have higher levels of plasma t-tau after controlling for sex and APOE. Although not directly testable in the present data, the average age of the lowest class LHC class (M_W1 = 73.9, M_W2 = 74.3, M_W3 = 74.8, M_W4 = 75.7, M_W5 = 77.0, M_W6 = 78.6) was somewhat lower than that of the highest LHC class (M_W1 = 75.1, M_W2 = 75.6, M_W3 = 75.9, M_W4 = 76.7, M_W5 = 78.2, M_W6 = 79.5) at each time point. It is possible that the reported age-related effects extend to a higher age range and to subtler age differences, representing an important area of future investigation.

Depressive symptoms (at a non-clinical level) were a selective predictor of LHC trajectory classes, with higher mean GDS score associated with the lowest trajectory class. This result is concordant with previous literature in which depression has been linked with increased AD risk [33]. Similarly, depressive symptoms have been associated with increased limbic and prefrontal atrophy over a four-year follow-up in cognitively normal older adults [107]. The left hippocampus (but not the right hippocampus) has also been found to be reduced in major depression disorder in adults [108]. In our sample, only 2% of individuals were considered mildly depressed at baseline and no individuals had GDS scores indicating moderate or severe depression. The present findings suggest that the association between mild depressive symptomology and prefrontal/limbic atrophy also extends to the left hippocampus. Although the mechanism of this relationship remains largely unknown, it is possible that such mood or affect symptomology is associated with the subtle changes in cognition as a function of emerging hippocampal and cortical atrophy [109]. Another perspective is that hippocampal atrophy may be directly affecting networks that are associated with mood and impact depressive symptomology through numerous mechanisms such as estrogen depletion and deregulation of certain neural circuits [110].

A lower body mass index (BMI) was associated with the lower LHC, but not RHC, trajectory class. BMI associations with brain and cognitive aging are complex [111–113]. A previous studying using BMI as a predictor of HC volumetric change reported a negative association between hippocampal volume (across hemispheres, but with stronger effects for the LHC) and BMI [114]. Participants of that study were, on average, a decade younger than those of the current study. Our findings indicate that a protective effect of higher BMI persists in an older cohort, and further support that this effect occurs more strongly in the LHC. Potential protective effects of increased BMI in older age (versus midlife or young-old cohort) have been reported in the context of AD risk [115, 116] and cognitive decline [117] and may act similarly for risk reduction for hippocampal atrophy. Notably, it appears that higher BMI might be an important AD risk factor in midlife, but this association reverses towards protection or risk-reduction in later life and older age, perhaps due to weight changes occurring in preclinical AD phases [118, 119].

We tested 38 biomarkers and risk factors as potential predictors of trajectory class membership. Our analytic approach considered all predictors simultaneously in a computationally competitive context. In addition to the seven predictors of trajectory classes, we note that there were 31 AD-related predictors that did not successfully emerge in either (LHC or RHC) of the analyses. Within the biospecimen modality, plasma measures of Aβ and tau outperformed CSF Aβ and tau to discriminate between hippocampal trajectories. Although CSF measures of Aβ have been consistently reported as sensitive biomarkers of MCI and AD, recent developments have identified less invasive and lower cost alternatives such as blood-based biomarkers [103]. Potentially, these peripheral biomarkers are more useful in predicting specific pathological changes and broader neurodegeneration, such as hippocampal atrophy. Alternatively, it is possible that the present plasma markers are better suited as predictors of non-clinical aging outcomes (i.e., hippocampal classes representing a dynamic distribution of cognitively normal longitudinal trajectories) as compared to related findings for CSF markers and associations with AD diagnosis and clinical progression patterns. For the genetic modality, although APOE genetic risk is the most important genetic risk factor for sporadic AD [120], it did not appear as one of the important or leading predictors of the lowest HC atrophy class (although it was among the lesser contributing predictors). This may point to an attenuated importance of single genetic factors within an interactive network of wide-ranging AD risk factors. The inclusion of a polygenic AD-related risk score may have revealed more predictive utility in the context of other risk-related AD predictors and should be investigated in future research [22]. Within the vascular/metabolic modality, no factors reached sufficient variable importance to be considered important predictors despite past findings suggesting possible associations [17, 121]. For the demographic modality, chronological age was not found to be an important predictor of the lowest hippocampal trajectory class membership. Instead, our findings indicate that, when available, certain aging-related mechanistic predictors may be more important than age per se for predicting adverse brain aging outcomes in predominantly cognitively normal samples. This provides additional support to the growing evidence that markers of biological age (versus chronological age) are important to consider in predictions of exacerbated decline in non-demented aging [122–125]. Given the current analytic approach and the use of a conditional variable importance measure, we identified the most prominent predictors of hippocampal trajectory classes in the context of other previously identified and often closely related AD-related biomarkers and risk factors.

There were several limitations to the present study. First, previous reports have acknowledged some limited generalizability of the ADNI cohort due to convenience sampling and possible biases in recruited participants (e.g., familial history of AD) [126]. However, these potentially at-risk individuals are key targets of clinical trials and prevention efforts. As our study aimed to identify biomarkers and risk factors associated with morphometric change in cognitively normal older adults, we have identified biomarker associations in individuals that are likely to be targeted for these purposes. Second, although variables included in the current study had few missing data (0–3.9%), there was a notable exception for biomarkers in the biospecimen modality. For the biospecimen biomarkers, missing data ranged from 35 to 51.3%. Missing data were imputed using the ‘missForest’ package in R which utilizes a random forest to iteratively predict missing values. The present imputation procedure and RFA models allowed for the inclusion of many predictors from multiple modalities despite some with higher rates of missing data. We consider this a notable strength of our approach, as previous studies predicting AD risk have often employed fewer biomarker or risk factor predictors, possibly due to analytical restrictions (e.g., multiple comparison issues) [127–129]. Replicating and validating these findings using additional biomarker data would be an important future step. Third, because of data limitations we were unable to investigate whether preclinical trajectory class membership would predict clinical diagnostic outcomes such as MCI or AD. As shown in Table 5, 96.3% of the analyzed longitudinal observations were with participants who were free of MCI or AD and over 99% included persons who were non-AD. In total, there were very few participants who transitioned to AD (n = 8) or MCI (n = 32, with 5 reverting back to CN) within the six waves under study—and together they contributed data for only 3.7% of the analyzed longitudinal observations (AD = 0.56%). By design, the present sample was selected initially to be cognitively asymptomatic (all were cognitively normal at baseline) and remained predominantly so throughout the study. The very small number of observations that could be characterized as impaired was appropriate for our objectives and expected in our design. No separate machine learning prediction analysis of this small cluster is possible due to severely imbalanced groups. However, a post-hoc check revealed that, in general, most of the individuals transitioning to impairment status were members of the lower trajectory classes. Accordingly, we suggest future work aimed at testing whether lower HC trajectory class membership is a reliable precursor condition for impairment and AD diagnosis. Fourth, the correlational analyses to clarify predictor directionality were focused more on describing associations with predictor variables than interpreting potential underlying mechanisms. Specific mechanisms should be further explored in future studies. Fifth, due to the ADNI MRI methods and protocols, almost all participants from ADNI1 were scanned using 1.5T scanners and all participants from ADNI2 were scanned using 3T scanners. However, we found no significant associations between scanner strength and hippocampal trajectory classes. This indicates that scanner strength was properly corrected for at the modelling stage, as has been done in previous studies [44]. Sixth, other (non-AD specific) pathologies and risk factors unavailable in this study may have contributed to the observed hippocampal volume and atrophy trajectories.