Rhodiola rosea L. Improves Learning and Memory Function: Preclinical Evidence and Possible Mechanisms

Rhodiola rosea L. (R. rosea L.) is widely used to stimulate the nervous system, extenuate anxiety, enhance work performance, relieve fatigue, and prevent high altitude sickness. Previous studies reported that R. rosea L. improves learning and memory function in animal models. Here, we conducted a systematic review and meta-analysis for preclinical studies to assess the current evidence for R. rosea L. effect on learning and memory function. Ultimately, 36 studies involving 836 animals were identified by searching 6 databases from inception to May 2018. The primary outcome measures included the escape latency in Morris water maze (MWM) test on behalf of learning ability, the frequency and the length of time spent on the target quadrant in MWM test representing memory function, and the number of errors in step down test, dark avoidance test and Y maze test on behalf of memory function. The secondary outcome measures were mechanisms of R. rosea L. for learning and/or memory function. Compared with control, the pooled results of 28 studies showed significant effects of R. rosea L. for reducing the escape latency (P < 0.05); 23 studies for increasing the frequency and the length of time spent on the target quadrant (P < 0.05); and 6 studies for decreasing the number of errors (P < 0.01). The possible mechanisms of R. rosea L. are largely through antioxidant, cholinergic regulation, anti-apoptosis activities, anti-inflammatory, improving coronary blood flow, and cerebral metabolism. In conclusion, the findings suggested that R. rosea L. can improve learning and memory function.


INTRODUCTION
Lasting changes in behavior resulting from prior experience can be characterized as the result of learning, memory, and retrieval processes (Thompson, 1986). However, memory is vulnerable across the adult lifespan. A decrease in learning and memory functions is the most common complaint in normal aging process. In a large number of organic diseases, in which there is a physical change in the structure of an organ or part, such as amnesia, Alzheimer's disease (AD) and vascular dementia, the most prominent sign is memory impairment (Thompson, 1986). Currently, there is no valid treatment for cognition impairment in western medicine, although many potential agents exist through novel mechanisms (Parihar and Hemnani, 2004). Cholinesterase inhibitors

Study Selection
Two investigators independently screened the titles and/or abstracts based on the search strategy. Of the search results, we assessed the full-text articles for eligibility. Any uncertainty eligibility was resolved by discussion. Studies were eligible for our systematic review if they met: (1) Animal models were established for learning and memory injury; (2) Analyzed interventions were received R. rosea L. and/or its bioactive ingredients as monotherapy at any dose. Comparator interventions were isosteric non-functional liquid (normal saline) or no treatment; (3) the primary measured outcomes were indexes of learning and/or memory function tests, including Morris water maze (MWM), Y maze, step down test, dark avoidance test, active avoidance reaction and one step through test. The secondary outcome measures were mechanisms of R. rosea L. for learning and/or memory function. Pre-specified exclusion criteria were as follows: R. rosea L. was treated in conjunction with other compounds or R. rosea L.-based prescriptions, or without predetermined outcome index, or without in vivo model, or without control group, or duplicate publications. In the case of multiple publications from one study, we choose the articles with the largest sample or the earliest publication.

Data Extraction
The following details were extracted from each included study: (1) the first author's name, publication year; (2) individual data for each study, including animal species, number, sex, and weight; (3) type of animal model and anesthetic used in the model; (4) intervention characteristics, including timing for initial treatment, dosage and method of treatment, duration of treatment, and comparable treatment of control group;(5) main outcome measures on behavior tests and its corresponding pvalue. For each comparison, we extracted data of mean value and standard deviation from each treatment and control group of every study. If the data for meta-analysis were missing or only expressed graphically, we tried to contact authors for further information or calculated by ourselves if available. Otherwise we only performed qualitative analysis. The data of highest dose was selected when the treatment group included various doses of the target drug. The result of the peak time point was included when the data were expressed at different times.

Quality Assessment
Two authors independently assessed the methodological quality of the included articles according to the Collaborative Approach to Meta-Analysis and Review of Animal Data from Experimental Studies (CAMARADES) 10-item checklist (Sena et al., 2007): (1) peer-reviewed publication; (2) statements of temperature control; (3) randomization to treatment or control group; (4) blinded induction of model; (5) blinded assessment of outcome; (6) use of anesthetic without significant intrinsic neuroprotective activity; (7) appropriate animal model; (8) sample size calculation; (9) compliance with animal welfare regulations; and (10) declaration of potential conflict of interests. Each study was given an aggregate quality score based on one-point awarding for each item. Discrepancies were resolved by discussion or consultation with corresponding author.

Statistical Analysis
Meta-analyses and sub-analyses were performed using RevMan 5.3 software. Outcome measures were all considered as continuous data and given an estimate of the combined overall effect sizes utilizing standard mean difference (SMD) with the random effects model. SMD with its 95% confidence interval (CI) was used to assess the strength of efficacy of R. rosea L. and/or its bioactive ingredients for learning and memory function. Publication bias was assessed with a funnel plot. To clarify the impact of factors potentially modifying the outcome measures, we also conducted sensitivity analyses and subgroup analyses according to the following variables: animal species, anesthetic used, type of animal model and the treatment time. The I 2 statistic was used for assessment of heterogeneity among individual studies. Probability value P < 0.05 was considered significant.

Effectiveness
Twenty-eight studies reported the escape latency in MWM as the outcome measure of learning ability included in the analysis. We pooled the whole data to process and found a significant difference in favor of R. rosea L. treatment compared with control groups (P < 0.00001; SMD = −1.83, 95% CI [−2.03, −1.64]; Heterogeneity: χ 2 = 174.39, df = 28 (P < 0.00001); I 2 = 84%, Figure 2). Twenty-three studies reported the frequency and/or the length of time spent on the target quadrant as the indicator of memory ability. The pooled result showed that R. rosea L. significantly increased the frequency and the length of time spent on the target quadrant in MWM (P < 0.00001; SMD = 1.79, 95% CI [1.60, 1.98]; Heterogeneity: χ 2 = 131.87, df = 32 (P < 0.00001); I 2 = 76%, Figure 3). Seven studies reported memory outcome measure by the number of errors in step down test, dark avoidance test, the active avoidance test and Y maze. The pooled data showed that R. rosea L. resulted in a significant depression on the number of errors when comparing to that in control groups (P < 0.00001; SMD = −1.04, 95% CI [−1.35, −0.72]; Heterogeneity: χ 2 = 6.93, df = 8 (P = 0.54); I 2 = 0%, Figure 4).

Subgroup Analysis and Sensitivity Analysis
To explore potential confounding factors which affected the outcome measures, we stratified analysis of the escape latency based on variables including animal species, animal model, the duration of treatment, and the quality of study. In the subgroup analysis of these factors, the effect size of rat species was larger than mice (SMD = −2.09 vs. SMD = −1.08, Figure 10A). Animal model showed great discrepancy in the overall effect of outcome measure, which the model of hypobaric hypoxia with scale of 16.4% weight accounted for smaller effect size than any other model (SMD = −1.18 vs. SMD pooled = −1.96, Figure 10B). The longer period of R. rosea L. treatment also showed greater effect size than the shorter treatment with 2 weeks or less (SMD = −1.92 vs. SMD = −1.83, Figure 10C). Notably, the lower quality studies did not exhibit larger effect size than the higher ones (SMD = −1.65 vs. SMD = −2.55, Figure 10D). Sensitivity analyses showed that the results did not substantially alter after removing any one trial. However, when we only include studies using mice as animal models, meta-analysis of 5 studies (Liu et al., 2003;Wu et al., 2004;Wang et al., 2013;Ge et al., 2017) showed a small difference in favor of R. rosea L. treatment compared with control groups with lower heterogeneity (n = 61, SMD = −1.08, 95%CI [−1.47, −0.68], P < 0.00001; Heterogeneity: χ 2 = 6.22, df = 4 (P < 0.00001); I 2 =36%).

Summary of Evidence
In this meta-analysis, we assessed R. rosea L. treatment on learning and memory function based on 36 eligible studies. The results revealed that R. rosea L. could evidently reduce the escape latency, improve the frequency and the length of time spent in MWM and decrease the number of errors in step down test, dark avoidance test, and Y maze when comparing with control groups in animal models.

Limitations
Some limitations should be considered while interpreting this study. First, the methodological quality of the included studies was considerably variable and inferior. Nearly all of the included studies had an overall assessment as "high risk of bias, " so we could not exclude that our results may be biased. Second, calculation of sample size and blindness of model establishment and outcome measurement are pivotal in quality control of research, yet no studies provided these critical information in this systematic review. Third, it's not worthy that almost all the included studies declared random allocation to treatment and control groups, while the detailed procedure was not supported at all. Additionally, gender     difference was overlooked in the included study. Male/female mice models were used in the two studies Wang et al., 2013) for cognitive experiments. Although the mechanism is unclear, a male advantage for working memory and a female advantage for visual memory and social cognition in rodent models were highlighted (Leger and Neill, 2016). Moreover, funnel plots (Figure 11) showed potential publication bias existed in this research field, suggesting studies with null effect are missing. Studies achieved statistically significant outcomes have been shown to be three times more likely to be published than that with null outcomes (Dickersin et al., 1987). Publication bias is due to multiple factor such as researchers and journal editors prefer positive results rather than negative or inconclusive results (Wolfgang, 2007). Thus, the effect of R. rosea L. on learning and memory function cannot be excluded from overall over estimation of effect sizes and efficacy, which may weaken the validity of conclusions.

Interpretation of the Results
Considerably high heterogeneity was present in this metaanalysis, the summary positive results should be interpreted with caution. Given that there are many potential sources of heterogeneity in the outcome, several means are taken into consideration for the finding of the causes. Firstly, randomeffects models are used in our study. Heterogeneity is a key condition for the execution of meta-regression, but it can also cause confusion if confounding factors are not well-balanced. As small number of studies were included in this meta-analysis, we made the meta-regression with reservations and did subgroup analysis based on four potential confounding factors including animal species, animal model, the duration of treatment, and the quality of study. The results of subgroup analyses suggested that the first three factors were very likely to be the sources of heterogeneity in this research, while the poor quality of methodology still could not be exempted from the excuses for high heterogeneity. Sensitivity analyses have also been adopted to detect the effects of studies identified as being aberrant result, or being highly influential in the analysis (Haidich, 2010). While no studies identified as being aberrant result or being highly influential in the analysis from the results of sensitivity analyses in this review.

Implication for Further Studies
While mice models are increasing used for cognitive experiments involving learning and memory process that were originally designed for rat species, the stability of spatial cognitive representation in rats changes more slightly over time than in mice (Hok et al., 2016). In the subgroup analyses, rat species also showed greater effect size in depression of escape latency than that of mice. Thus, rat species were considered as suitable cognitive experiments involving learning and memory process. In addition, the impact of gender on cognitive function deserves attention. In the present study, male rats models and male/female mice models were used in the included studies of our review for working memory process, while no significant difference existed in the pooled result of meta-analysis in escape latency of MWM test after discarding two studies with male/female mice Wang et al., 2013). However, a male advantage for working memory and a female advantage for visual memory and social cognition in rodent models were highlighted in recent systematic review (Leger and Neill, 2016). Thus, using a single sex animal model is considered more reasonable for study learning and memory function in future experiments.
Two dementia models of AD and vascular dementia (VaD) are most commonly approached for learning and memory  research (Kalaria et al., 2008). However, there are several model methods for inducing these two dementia types and their differences of effectiveness and robustness are not investigated. For this systematic review, intra-peritoneal injection with scopolamine, combination with aluminum trichloride, Dgalactose and scopolamine, intracerebroventricular injection with streptozotocin, and hippocampal injection with Aβ 1−40 were the most approaches for AD models in the included studies. Different time scales of artery occlusion and different arteries selected for blood blocking were adopted for VaD models. In the subgroup analyses, six animal models including AD, VaD, hypoxia, sleep deprivation, epilepsy, and diabetes mellitus models were conducted for cognitive impairment, of which AD models accounted for 38.7% weight and VaD models accounted for 17.7% weight. These two most weight of models showed no significant difference in effect size on escape latency of MWM test, which can indirectly reflect the effectiveness and robustness of the two dementia models for cognitive impairment.
A lower-quality study trends toward better outcomes, leading to the global estimated effect overstated (García-Bonilla et al., 2012). In the present study, many domains had flaws in aspects of randomization, allocation concealment, and blinding and sample size calculation, which are the core standards of study design (Moher et al., 2015). Thus, we recommended that the experimental research of R. rosea L. for learning and memory function need be promoted by means of incorporating the ARRIVE guidelines (Kilkenny et al., 2012).
Long-term treatment for dementia progression with Gingko biloba showed great effect on prevention of cognitive decline (Dodge et al., 2008). In parallel, treatment with R. rosea L. more than 2 weeks showed greater effect size in the escape latency of MWM test than that of <2 weeks' treatment in the subgroup analyses, suggesting that long-term treatment with R. rosea L. has a greater benefit for cognitive function. In view of the number of studies in subgroup analyses was relatively small and may lack of statistical power to detect smaller effect sizes. Therefore, we recommend that future studies involving this problem are conducted strictly complying with standards of research methodology and report their adequate information clearly.

CONCLUSION
We have provided a first-ever comprehensive preclinical systematic review of R. rosea L. for cognitive behavior in