Deficits in hippocampal-dependent memory across different rodent models of early life stress: systematic review and meta-analysis

Exposure to early life stress (ELS) causes abnormal hippocampal development and functional deficits in rodents and humans, but no meta-analysis has been used yet to quantify the effects of different rodent models of ELS on hippocampal-dependent memory. We searched PubMed and Web of Science for publications that assessed the effects of handling, maternal separation (MS), and limited bedding and nesting (LBN) on performance in the Morris water maze (MWM), novel object recognition (NOR), and contextual fear conditioning (CFC). Forty-five studies met inclusion criteria (n = 451–763 rodents per test) and were used to calculate standardized mean differences (Hedge’s g) and to assess heterogeneity, publication bias, and the moderating effects of sex and species (rats vs. mice). We found significantly lower heterogeneity in LBN compared to handling and MS with no consistent effects of sex or species across the three paradigms. LBN and MS caused similar cognitive deficits in tasks that rely heavily on the dorsal hippocampus, such as MWM and NOR, and were significantly different compared to the improved performance seen in rodents exposed to handling. In the CFC task, which relies more on the ventral hippocampus, all three paradigms showed reduced freezing with moderate effect sizes that were not statistically different. These findings demonstrate the utility of using meta-analysis to quantify outcomes in a large number of inconsistent preclinical studies and highlight the need to further investigate the possibility that handling causes different alterations in the dorsal hippocampus but similar outcomes in the ventral hippocampus when compared to MS and LBN.


Introduction
Early life stress (ELS) impairs the development of several interconnected brain regions including the hippocampus, prefrontal cortex, striatum, and amygdala 1,2 . Of these, reduced hippocampal volume and abnormal hippocampal function are some of the most reproducible findings in individuals exposed to ELS 1,3,4 . This is not surprising given that the hippocampus undergoes substantial developmental changes during childhood that are highly sensitive to stress [5][6][7] . The hippocampus also forms extensive connections with other brain regions involved in declarative memory, spatial learning, threat detection, stress reactivity, and mood regulation 8,9 . These features make the hippocampus a potentially important hub in mediating several of the cognitive, emotional, and physiological abnormalities seen in individuals exposed to ELS 1,3,4 .
Elucidating the mechanisms by which ELS affects hippocampal development is a challenging task in humans because of the inherent complexity and heterogeneity of the adversities, genetic variability, and numerous additional variables that are difficult to control in clinical settings 10 . In addition, causally linking structural and functional changes in the hippocampus with alterations in anxiety, stress reactivity, or cognition are practically impossible to do in humans. Nevertheless, the conserved nature by which the hippocampus develops in mammalian species 5,11 and the observation that rodent models of ELS cause significant impairment in its development and long-term function 10 suggest that work in rodents may clarify important details about the role that abnormal hippocampus development plays in the longterm consequences of ELS. Indeed, an elegant body of work has shown that chronically elevated levels of the neuropeptide corticotrophin-releasing hormone (CRH) mediate the synaptic and hippocampal-dependent memory deficits in rodents exposed to a rodent model of ELS known as limited bedding and nesting (LBN) 12 .
Several key questions regarding the effects of ELS on hippocampal function in rodents remain unresolved. For example, it is currently unclear whether some types of ELS lead to more severe hippocampal-dependent deficits, whether males and females are equally sensitive, and if some tests are more sensitive for detecting hippocampaldependent cognitive deficits caused by ELS. Even within the same ELS paradigm, different research groups report variable outcomes and the literature is replete with examples of inconsistent findings 10,13,14 . These inconsistencies appear to be related to many factors such as different genetic backgrounds, poor standardization of procedures across labs, and other stochastic variables 10,[14][15][16] . To address these issues, we conducted a systematic review and subsequent meta-analysis examining the effects of three types of ELS: handling (also known as brief maternal separation), maternal separation (MS), and LBN on hippocampal function in rodents (mice and rats). Hippocampal-dependent memory was assessed using 3 behavioral tests: (1) the Morris water maze (MWM), (2) Novel Object Recognition (NOR), and (3) Contextual Fear Conditioning (CFC). We chose these tests because they all require normal hippocampal function 8,17,18 and are commonly tested in rodents exposed to ELS 13,14 . Moreover, the MWM and the NOR rely more on the dorsal hippocampus 8,[19][20][21] , whereas freezing behavior in the CFC is thought to have a strong ventral hippocampus component 8,[22][23][24][25] . Nevertheless, performance in these tests, especially steps that involve consolidation and retrieval, also require the prefrontal cortex and other cortical regions that need to be considered when interpreting behavioral outcomes in these tests 17,18,[26][27][28] .
Although two systematic reviews have previously examined the effects of different types of ELS on behavioral outcomes in rodents, including spatial learning 13,14 they were qualitative in nature and did not include a meta-analysis. Meta-analysis is routinely used in clinical settings but has rarely been used to address conflicting results in preclinical studies in rodents 29 . In fact, we are aware of only two meta-analyses examining the effects of ELS on behavioral outcome in rodents, one assessing pain sensitivity 30 and the other examining anxiety-likebehavior 31 . No meta-analysis to date has assessed the effects of ELS on spatial learning and none of the available meta-analysis included the LBN which is one of the most commonly used ELS paradigms in rodents 32 .
The primary goals of this meta-analysis were to compare the effect sizes and heterogeneity of the three different ELS paradigms in the MWM, NOR, and CFC. We also examined the effects of several moderators such as sex, species (i.e., mice vs. rats), separation index (i.e., the number of days pups were separation X length of the separation), and temperature of isolation on cognitive outcomes. Lastly, we assessed for possible publication bias and its impact on behavioral outcomes in these tests.

Search strategy
Two reviewers (MR, DW) searched the electronic databases of PubMed and Web of Science on September 30th, 2019 for relevant studies using the following searches: (early life stress OR ELS OR postnatal stress OR maternal separation OR neonatal stress OR limited bedding and nesting OR LBN OR brief maternal separation OR BMS OR handling) AND (mice OR mouse OR mus musculus OR rats OR rat) AND (Morris water maze OR MWM OR novel object recognition OR NOR OR novel object location OR NOL OR contextual fear conditioning OR CFC). Studies were limited to English language articles. Additional citations came from previous systematic reviews 13,14,32,33 . Studies obtained from the search, the titles, and abstracts were examined by the two reviewers (MR, DW) to determine preliminary inclusion. Discrepancies were addressed by the reviewers through discussion and through conversation with senior reviewers (AK, MHB). Despite a concerted effort to identify all relevant studies, it is possible that we have missed some studies due to the enormity of the literature and the different names used to describe these paradigms. In addition, we did not register this study with PROSPERO prior to extracting the data and therefore are not able to do this at this point. However, we are confident that no previous meta-analyses have examined this question before.

Study selection and data extraction
After preliminary inclusion, studies were carefully read to determine if inclusion criteria were met. See Supplemental Information for additional details regarding selection criteria and data extraction are available in the supplementary information.

Data analysis
Statistical Analyses were conducted using the Comprehensive Meta-Analysis Version 3.0 software (Biostat, 2016). Outcomes of interest were measures of hippocampal-dependent Memory in the MWM, NOR, and CFC. Details regarding specific behavioral measurements are available in the supplementary information. Hedge's g was used as the pooled measurement of effect size as it is preferred over Cohen's d for small samples, which are common in animal studies 34 . Both fixed and random effects were calculated and are presented in the results section. Random effect analysis was used as the primary analysis to account for the large heterogeneity between studies. Fixed-effect analysis was conducted as a sensitivity analysis because random-effects models may be overly conservative in situations where a relatively small number of studies is available. Publication bias was assessed using funnel plots and the Egger's test, and heterogeneity was measured utilizing the I 2 statistics and Chi-square test for heterogeneity 35,36 . Details regarding the analyses for moderating effects of sex, species, separation index and temperature are available in the supplemental information. Figure S1 depicts the selection strategy for included studies. In total, 1435 articles were identified for consideration, of which 45 studies were eligible for inclusion. Reasons for exclusion of studies are identified in the Supplemental Information Fig. S1. Table 1 describes the characteristics of our included studies and additional details are available in the supplemental information Tables S1 and S2.
In contrast to the reduced latency seen in handling, exposure to MS was associated with a significant increase in the latency to find the platform (random effect: Fig. 1, middle). There was significant heterogeneity between studies (I 2 = 62%, Q = 37, df = 14, p = 0.0007) with no evidence of publication bias based on the funnel plot (Supplemental information Fig. S3B) or the Egger's test (p = 0.14). Stratified subgroup analysis demonstrated no significant effect of sex (Q = 1.9, df = 1, p = 0.17), and all MS studies were done in rats, preventing us from assessing the effect of species on latency to find the platform. Moderator analyses found no significant effect with longer latency to find the platform associated with incubation at a higher temperature during the MS procedure.

MWM: probe trial
The forest plot summary for the effects of handling (2 studies, n = 60 rodents), MS (15 studies, n = 343 rodents), and LBN (7 studies, n = 162 rodents) on performance in the probe trial of the MWM are shown in Fig. 2. Rodents exposed to MS and LBN spent significantly less time swimming in the correct quadrant whereas handling was associated with a non-significantly increased time spent in the correct quadrant compared to controls. Tests  Since only two handling studies were included (Table  S2), the outcomes for the probe trial should be considered exploratory and did not include analyses of publication bias or on the effects of sex and species. Nevertheless, handling increased time swimming in the correct quadrant (Fig. 2, top), an outcome that was not significant using random effect (Hedge's g = 0.72 ± 0.49, 95% CI = −0.25-(+1.69), z = 1.5, p = 0.14, k = 2) but significant using a fixed-effect analysis (Hedge's g = 0.58 ± 0.27, 95% CI = 0.06-1.10, z = 2.2, p = 0.029, k = 2). The heterogeneity between the studies was large but did not reach statistical significance given the small number of studies (I 2 = 65%, Q = 3, df = 1, p = 0.091).
Exposure  Fig. 3 middle). The heterogeneity between the studies was large and statistically significant (I 2 = 72%, Q = 50, df = 14, p < 0.0005). Funnel plot asymmetry suggested a possible publication bias (Supplemental information Fig. S4A), a finding that was confirmed using the Egger's test (p = 0.009). Nevertheless, the effect of MS remained significant after adjusting for funnel plot asymmetry using Duval and Tweedie's Trim-and-Fill method, Hedge's g = −0.30, 95% CI = −05-(−0.09). There was no significant effect of sex (Q = 0.24 df = 1, p = 0.62) and all MS studies were conducted in rats (Table S2) . However, as with the latency to find the platform there was a significant effect of separation temp on outcomes in the probe trial with worse performance in rodents exposed to higher temp during the MS procedure (β = −0.33 ± 0.12, 95% CI = −0.56-(−0.10), Z = −2.80, p = 0.05, k = 8).
Similar to MS, rodents exposed to LBN showed significantly reduced time swimming in the correct quadrant ( Fig. 2 bottom). This outcome had a large effect size that was highly significant (random effect: Hedge's g = −0.

Novel object recognition (NOR)
A forest plot summarizing the effects of handling (3 studies, n = 67 rodents), MS (8 studies, n = 197 rodents), and LBN (11 studies, n = 232 rodents) on performance in the NOR test is shown in Fig. 3. Tests for subgroup differences found no significant differences in effect size using random effect (χ 2 = 2.45, df = 2, p = 0.29), but significant differences using fixed-effect analysis (χ 2 = 7.94, df = 2, p = 0.019). However, the relationship between the three paradigms was similar to outcomes in the latency to find platform and probe trial (Fig. 3).
In contrast to the improved cognition in the MWM and the lack of impact in the NOR, exposure to handling reduced contextual freezing suggesting impaired hippocampal-dependent memory. The effect size was moderate and did not reach statistical significance using random-effect analysis, most likely due to the small number of available studies (Hedge's g = −0.48 ± 0.33, 95% CI = −1.12-(+0.17), z = −1.44, p = 0.15, k = 4). The effect of handling on CFC was significant when the fixed-effect analysis was performed (Hedge's g = −0.41 ± 0.21, 95% CI = −0.83-(−0.001), z = −1.97, p = 0.049, k = 4, Fig. 4 top). There was a large heterogeneity that did not reach statistical significance (I 2 = 57%, Q = 6.92, df = 3, p = 0.075) and too few studies were available to assess publication bias or the effects of sex and species.
MS also reduced contextual freezing, an effect that was not significant for random effect (Hedges g = 0.05 ± 0.32, 95% CI = −0.58-(+0.67), z = 0.15, p = 0.88, k = 12), but significant for fixed-effect analysis (Hedges g = −0.28 ± 0.13, 95% CI = −0.53-(−0.03), z = −2.24, p = 0.025, k = 12, Fig. 4, middle). As with other studies involving MS, there was large and highly significant heterogeneity between studies (I 2 = 83%, Q = 64.8, df = 11, p < 0.0005). There was also substantial evidence for publication bias based on funnel plot asymmetry and the Egger's test (p = 0.006) that were highly influenced by 1 outlier study (46) (Supplemental information Fig. S6). We were unable to assess the effect of species because all studies were conducted in rats (Table S2)  Rodents exposed to LBN also showed reduced contextual freezing, with a moderate effect size that did not reach significance most likely due to the small number of  Fig. 4 bottom). Heterogeneity between studies was small and did not reach statistical significance (I 2 = 12%, Q = 2.27, df = 2, p = 0.32) and too few studies were available to assess publication bias or the effects of sex or species (Table S2).

Discussion
This is the first meta-analysis that examines the effects of different rodent models of ELS on spatial learning providing several new insights to some key questions in the field. For example, we show that MS and LBN cause similar cognitive deficits in the MWM latency to find the platform, the probe trial, and the NOR test (Summarized in Fig. S2). The effect sizes of MS and LBN in these tasks were mostly moderate (Hedge's g = −0.3-(−0.6), except for large effect size in the probe trial for LBN (Hedge's g = −0.82). In contrast, exposure to handing showed improved performance in the latency to find platform and probe trial of the MWM, with no clear difference compared to control condition in the NOR (Fig. S2). These results demonstrate both quantitative and qualitative differences between handling and MS and LBN in the MWM and NOR and are consistent with previous work showing that handling causes different outcomes in stress reactivity 33,37 and sensitivity to pain compared to MS 30 . Sustained elevation of CRH in the hippocampus has been shown to play a central role in inducing synaptic abnormalities and spatial-learning deficits in rodents exposed to LBN 12,38 , but its role in modifying hippocampal function in MS and handling has not been studied extensively. Nevertheless, Wang et al. found increased CRH levels in the hippocampus in rats exposed to MS, and administrating the CRH receptor antagonist (CP-154526) improved performance in the MWM and NOR in rats exposed to MS 39 . To the best of our knowledge, no group has yet shown that handling causes a reduction in CRH levels in the hippocampus, but work by Fenoglio et al. found that transient administration of CRH receptor 1 antagonist to control pups from P10-17, enhanced performance in the MWM and NOR to levels seen in handled animals. A similar procedure did not affect cognitive performance in handled animals, consistent with the notion that CRH levels are low in this group 40 . Additional studies are therefore warranted to compare CRH levels in the dorsal and ventral hippocampus (see below), across these three models of ELS.
One of the most intriguing findings of our analysis is that exposure to handling causes similar deficits in contextual freezing compared to LBN and MS (Fig. S2). This result needs to be interpreted with caution given the relatively small number of studies contributing to this outcome and that the impacts of handling and MS were only significant using fixed-effect analysis. Nevertheless, it demonstrates that handling shares some similarities with other ELS paradigms and raises the intriguing possibility that all three ELS paradigms cause similar deficits in the ventral hippocampus. This assertion is consistent with data showing that the ventral hippocampus plays an important role in contextual freezing 8,[22][23][24][25] and provides a parsimonious explanation for these seemingly conflicting results. Additional support comes from studies showing that the offspring of low licking and grooming (LG) dams show reduced longterm potentiation (LTP) and poor spatial learning in the dorsal hippocampus, but increased LTP in the ventral hippocampus 41 and enhanced contextual freezing compared to offspring of high-LG 42 . Important differences in the effects of prenatal stress on the dorsal versus the ventral hippocampus have also been reported 43 . However, given the role that other brain regions, such as the prefrontal cortex, play in these tasks 17,18,[26][27][28] , it is also possible that changes in these other areas and not the hippocampus are responsible for these behavioral outcomes.
The heterogeneity in LBN studies was significantly lower compared to both MS and handling studies across 3 of the 4 cognitive tests. Large heterogeneity was also found in a recent meta-analysis examining the effects of MS on anxiety-like behavior 31 . This is not surprising given the lack of standardization in the separation procedure associated with the MS and handling paradigms (e.g., number of days, length of separation, temp, single vs. whole litter separation). In addition, regular human contact in the MS and handling procedures may also increase variability due to the effects of the sex of the researcher and his/her experience scruffing and transferring animals 31 . Indeed, moderator analyses focusing on MS studies found that longer separation procedures (i.e., greater separation index) were associated with more severe cognitive outcomes in the NOR, but not MWM. Surprisingly, we found that incubating pups at higher temp during the separation procedure was associated with worse cognitive outcomes in the MWM (both latencies to find platform and probe trial) and were not able to examine this in the NOR because of the small number of studies available. This was an unexpected outcome because of the known sensitivity of rodent pups to hypothermia 44 and will require additional replication.
A previous meta-analysis has found that maternal separation increases anxiety-like behavior in rats, but not mice 31 . This was not the case for hippocampal-dependent tasks where most of the outcomes were similar in rats and mice. The only exception was a performance in the latency to find a platform where rats exposed to LBN were significantly more affected than mice exposed to LBN (Fig. 1). A systematic review without meta-analysis conducted by Loi et al. reported that males had more cognitive deficits in non-stressful learning paradigms compared to female rodents 13 . This was confirmed for NOR in rodents exposed to MS, but overall there was no clear sex effect on hippocampal-dependent function. These different outcomes might be due to the inclusion of non-hippocampal learning tests in the Loi et al. study and the more quantitative approach used in this study.

Conclusions
This work provides several new insights on the impact of different rodent models of ELS on spatial learning. First, LBN and MS, cause similar deficits in tasks such as the MWM and NOR that rely heavily on the dorsal hippocampus. In contrast, handling improved performance in the MWM and had no significant effect in the NOR. Second, all ELS paradigms, reduced contextual freezing, suggesting similar abnormalities in the ventral hippocampus and/or other brain regions. Third, heterogeneity was significantly lower in LBN compared to handling and MS. Fourth, ELS causes similar cognitive deficits in male and female rodents with no differences in the sensitivity between mice and rats.