Non-occupational physical activity during pregnancy and the risk of preterm birth: a meta-analysis of observational and interventional studies

A meta-analysis was conducted to evaluate the association between non-occupational physical activity (PA) during pregnancy and the risk of preterm birth (PTB). By searching PubMed and EMBASE from inception to August 20, 2016, 25 observational studies (18 cohorts and 7 case-controls) and 12 interventional studies were identified. Comparing the highest to the lowest category of leisure-time PA during pregnancy, the pooled relative risk (RR) of PTB was 0.83 [95% confidence interval (CI) = 0.74–0.93] for cohort studies and 0.60 (95% CI = 0.43–0.84) for case-control studies. No overall significant association was found between domestic or commuting PA and the risk of PTB. In addition, PA intervention did not indicate significant beneficial effect on the risk of PTB. Evidence from the observational studies suggested that leisure-time, but not domestic or commuting, PA during pregnancy was inversely associated with the risk of PTB. The findings were not supported by small-scale and short-term interventional studies. Further research with objective measurement on leisure-time PA is warranted.


Quality assessment. All identified observational studies received quality assessment based on the
Newcastle-Ottawa quality assessment scale 16 , which evaluates observational studies from three aspects: the selection of study population (4 criteria with 4 stars), the comparability of study population (1 criterion with 2 stars), and the assessment of exposure (3 criteria with 3 stars) for a cohort study or the ascertainment of outcome (3 criteria with 3 stars) for a case-control study. Each star was assigned 1 point with a total of 9 points. The quality of study was considered high if the sum score was ≥ 8 points, and moderate if the sum score ranged from 5 to 7 points.
Data extraction. Two co-authors (J. W. and P. X.) independently reviewed each included study and extracted the relevant information. Discrepancies were resolved by group discussion with a third co-author (K. H.). The following information was extracted: last name of the first author, the year when the paper was published, the country where the study was conducted, study design, study period, the number of participants/cases, participant age, exposure and its assessment method, outcome ascertainment, measures of the associations of interest [i.e., RR, HR, or OR with the corresponding 95% CIs], and the covariates adjusted in the final model.

Data synthesis and analysis.
Since PTB is a relatively rare disease, we ignored the distinction among the association measures (i.e., RR, HR, and OR) and undertook a random-effects meta-analysis to estimate the pooled relative risk (RR) and 95% CIs comparing the highest to the lowest category of PA level. If a study did not present multivariable-adjusted models, the unadjusted data was used. When no effect estimate was given, a crude estimate was calculated directly from a 2 by 2 table based on available information. If the estimates were reported for different trimesters respectively, they were combined first with a random-effects meta-analysis model.
Heterogeneity among studies was examined by using Cochran's Q test and quantified by using the I 2 statistic. To reduce the likelihood of drawing a false negative conclusion (type II error), a P value of ≤ 0.10 is considered as statistically significant for Cochran's Q test. Very low, low, moderate, and high degree of heterogeneity were defined as ≤ 25%, 26-50%, 51-75% and > 75%, respectively. Publication bias was assessed by Egger's regression asymmetry test. The Duval and Tweedie nonparametric "trim and fill" method was used to estimate the pooled association of interest if publication bias was suggested 17 . Sensitivity analyses were performed to evaluate the robustness of our findings: 1) to remove one study from the pooled analysis each time; and 2) to replace random-effects with fixed-effects models.
All analyses were performed with STATA software (Version 14, STATA Corporation LP, College Station, TX). A two-sided P value of ≤ 0.05 was considered statistically significant if not otherwise specified.

Results
Study selection process. Figure 1 shows the detailed selection process. We retrieved 214 relevant studies from PubMed and EMBASE. Of them, 189 studies were excluded for one of the following reasons: (1) not an original study; (2) not in English; (3) occupational population, such as nurses, military personnel, physicians; (4) no information on the association of interest and such information cannot be derived from available data; (5) no regular measurement in PA; or (6) duplicated publication. In addition, we identified 12 studies through Google Scholar and the relevant reference lists. Therefore, 37 eligible studies (7 case control studies, 18 cohort studies, and 12 interventional studies) were included in the meta-analysis. All the included 25 observational studies (7 case-control studies and 18 cohort studies) were rated as either high 18 or moderate 9,19-41 quality (see Supplemental Tables S1 and S2). Tables 1 and 2 summarize the characteristics of the included observational studies. Information on leisure-time PA and the risk of PTB were provided in 13 cohort studies 9,20,[25][26][27][28][29][30][32][33][34][35][36] (167,087 participants and 9,096 cases), and 4 case-control studies 21 28,32 . Data on domestic PA and the risk of PTB were presented by 7 cohort studies 9,[22][23][24][25]31,34 (11,009 participants and 747 cases), and 3 case-control studies 18,21,40 (391 cases and 651 controls). Of these studies, 3 were conducted in North America 9,21,34 , 2 in Europe 22,23 , 2 in the Asia-Pacific region 24,25 , 2 in Africa 18,31 , and 1 in South America 40 . Information on commuting PA and the risk of PTB was available in 5 cohort studies 9,19,23,24,35 (5,489 participants and 592 cases), and 1 case-control study 38 (2,230 cases and 3,907 controls). Of these studies, 3 were conducted in Europe 19,23,38 , 2 in North America 9,35 , and 1 in the Asia-Pacific region 24 . Leisure-time physical activity and the risk of preterm birth. Thirteen cohort and 4 case-control studies have data on leisure-time PA and the risk of PTB. Comparing the highest to the lowest category of leisure-time PA, the pooled RR of PTB was 0.83 (95% CI = 0.74-0.93) for cohort studies and 0.60 (95% CI = 0.43-0.84) for case-control studies (Fig. 2). No significant heterogeneity (I 2 = 18.5%, P = 0.26) was observed in cohort studies, but a moderate heterogeneity was found in case-control studies (I 2 = 54.3%, P = 0.09). Since publication bias was  Sensitivity analysis indicated that no single study appreciably changed the results, and the pooled associations persisted when a fixed-effects model was used instead of a random-effects one. Notable, the pregnancy period (first, second, third trimester, or mixed) and format (intensity, duration or frequency) of PA assessed were different among these studies. However, the pooled RR was similar [0.83 (95% CI = 0.78-0.88)] when combining data from 11 cohort studies in which leisure-time PA was assessed using frequency (i.e., yes or no, minutes per week, hours per week, times per week and times per month). In addition, the pooled estimate was essentially unchanged [0.80 (0.69-0.94)] when we combined data from 4 cohort studies in which leisure-time PA was measured in the first two trimesters.

Study characteristics.
Domestic physical activity and the risk of preterm birth. Seven cohort and 3 case-control studies reported results on domestic PA during pregnancy and the risk of PTB. No significant association was revealed. The pooled RR was 0.86 (95% CI = 0.65-1.14) for cohort studies and 0.64 (95% CI = 0.39-1.07) for case-control studies (Fig. 3). Neither significant heterogeneity (I 2 = 29.1% and P = 0.21 for cohort; I 2 = 17.4% and P = 0.30 for case-control) nor publication bias (Egger's test: P = 0.56 for cohort and P = 0.74 for case-control) was evident. The pooled results generally remained when using a fixed-effects model. However, the pooled association became statistically significant [0.78 (95% CI = 0.60-0.997)] after omitting Misra 9 et al. among cohort studies. Of note, the domestic PA in that study was defined as lifting heavy objects at home. This inverse association was slightly strengthened when further excluding another study 22 that also included lifting objects at home as the domestic PA [0.74 (95% CI = 0.55-0.998)].
Commuting physical activity and the risk of preterm birth. Five cohort studies and 1 case-control study presented data on commuting PA during pregnancy and the risk of PTB. No significant association was found among cohort studies comparing the highest to the lowest level of commuting PA (the pooled RR = 1.08; 95% CI = 0.67-1.75). Also, publication bias was not evident (Egger's test: P = 0.79). The observed null association was not appreciably altered by any single study and the pooled results persisted when the random-effects model was replaced with a fixed-effects model in the sensitivity analyses.
Physical activity intervention and the risk of preterm birth. Twelve interventional studies presented data on PA intervention and preterm birth, and found no significant association (the pooled RR = 1.15; 95% CI = 0.82-1.61). Neither significant heterogeneity (I 2 = 0.0% and P = 0.95) nor publication bias (Egger's test: P = 0.78) was documented. Sensitivity analysis indicated that no single study appreciably changed the results, and the pooled associations persisted when a fixed-effects model was used. When two studies specifically on sedentary women and one study on overweight women were excluded, the results were materially unchanged (the pooled RR = 1.14; 95% CI = 0.81-1.62) (Fig. 4).

Discussion
In the meta-analysis of observational studies, we found a significant inverse association of leisure-time PA during pregnancy with the risk of PTB. Domestic PA was inversely associated with the risk of PTB only if studies defining domestic PA as lifting heavy objects at home were excluded. No significant association was observed between commuting PA and the risk of PTB. However, findings from the observational studies were not supported by interventional studies, which indicate null association.  risk of PTB, which significantly increased the statistical power to detect potential associations. Specifically, we assessed the association separately for each domain of non-occupational PA. Also, all included observational studies were assessed as moderate or high quality using a standardized protocol, so that the likelihood was reduced that the pooled results were substantially biased. Nevertheless, findings from the observational studies should be interpreted in caution because of the following considerations: first, misclassification of PA levels is a concern since PA was assessed with an interview-based questionnaire during pregnancy in the primary studies, which might be subject to recall bias. However, the misclassification is likely to be non-differential and may attenuate the observed associations. To provide more accurate information on PA, objective measurements such as an accelerometer should be used. Second, although the meta-analysis was mainly based on fully adjusted models in the primary studies, the possibility that results were biased by residual confounding or unknown factors could not be completely excluded given the nature of observational study. For example, only a few primary studies considered occupational activity and socioeconomic status in the analysis. This might be an inherent limitation that might affect our findings in the meta-analysis. Third, moderate heterogeneity was observed in a couple of pooled analyses. The sources of heterogeneity include variations in study population, study region, sample size, exposure assessed at different stage of pregnancy, and adjustment for different covariates. We used a random-effects model in concordance with the heterogeneity. Fourth, publication bias due to unpublished data or publications in other languages could not be ruled out. Nevertheless, we used the Duval and Tweedie's "trim and fill" method to adjust for publication bias. Thus, our findings should not be substantially biased. Fifth, the primary studies did not provide sufficient information to enable us to investigate some important effect modifications such as the age of the women at pregnancy. By design, intervention studies or clinical trials have certain advantages over the observational studies. However, a few limitations should be acknowledged when interpreting the pooled results from the interventional studies in this meta-analysis. First, the sample size and the number of cases of PTB are relatively small, which indicates the statistical power may not be sufficient. Second, most of the included studies were not designed specifically for studying PTB, i.e., the primary outcome was not PTB (a rare disease) but other outcomes, such as the newborn's body size 43,44,46,48 , maternal aerobic capacity change 42,45 or weight gain 49,51 during pregnancy, and pregnancy-induced hypertension 53 , which from the other angle explained the low power for the analysis. Third, the most important limitation of the interventional studies was the practical difficulty of maintaining a high compliance in the exercise group due to logistical and family constraints; similarly, the control group may be aware the benefit of exercise and consequently continue or increase their PA, which may explain the null association.

Strengths and limitations.
Comparison with other reviews. Several reviews [54][55][56][57][58][59][60][61][62] of observational studies investigated the associations of non-occupational PA during pregnancy with the risk of PTB. While most of them concentrated on leisure-time PA, only three discussed different domains of non-occupational PA during pregnancy in relation to the risk of PTB [55][56][57] . Of these studies, one systematic review 62 of literature up to 2014 qualitatively assessed the association of leisure-time PA during pregnancy with the risk of PTB and supported the assertion that healthy pregnant women can engage in low, moderate, and even some vigorous levels of leisure-time PA without risk for preterm birth. Another review 60 quantitatively combined data from only 4 cohort studies, but found null association between leisure-time PA during pregnancy and the risk of PTB, which may be due to insufficient statistical power.
Several other reviews of interventional studies have discussed the effect of PA during pregnancy on the risk of PTB. For example, a Cochrane review published in 2010 63 , which combined data from 3 studies with a total of 6 PTB cases concluded that the data are insufficient to draw any conclusion. In addition, a meta-analysis of interventional studies 64 , which used maternal weight as the primary outcome, found that PA had a trend of reducing the risk of PTB, though the pooled result from 5 trials (450 participants with 20 cases) was statistically non-significant. One recent systematic review and meta-analysis of 9 interventional studies 65 , including one abstract and one published in other language, concluded that aerobic exercise was not associated with an increased risk of PTB. Similarly, a meta-analysis 62 of 17 trials found no significant difference in gestational age at delivery between the PA group and the control group.
Although our results are generally consistent with the previous findings, we think the present meta-analysis provides more robust results and additional information to the literature by combining evidence from both observational and interventional studies and focusing on the different domains of non-occupational PA.
Potential mechanisms. It is generally recognized that pregnant women can get tremendous benefit from regular PA. First, maintaining PA during pregnancy will help pregnant women maintain a general condition of health via improving their lipid profiles and lowering their blood pressures 3,4 . Second, regular PA during pregnancy will help women relieve symptoms during pregnancy (e.g., nausea and vomiting) 66-68 via hormonal and metabolic adaptations associated with improved cardiovascular functioning and alterations in catecholamine release and response 69 . Third, it can help pregnant women reduce the risk of developing chronic diseases such as gestational diabetes mellitus and preeclampsia 5,6 via improved insulin sensitivity, decreased concentrations of proinflammatory cytokines in peripheral circulation, reduced oxidative stress, and improved plasma lipid and lipoprotein concentrations.
There are several explanations for the potential beneficial effect of leisure-time PA on the risk of PTB. First, leisure-time PA may be less strenuous than the other two domains of non-occupational PA. Second, compared with domestic and commuting PA, women who engage in leisure-time PA may represent a select group who are more relaxed, with less stress, since gestational depression is an established risk factor of PTB 70 .
In addition, two cohort studies 9,22 reported results on domestic PA during pregnancy and risk of PTB, in which the domestic PA was defined as lifting objects at home. Non-heavy domestic PA may provide a similar benefit as leisure-time PA after omitting these two studies. Weight lifting may raise the blood pressure and does little or nothing to benefit the heart and cardiovascular system in general, which may explain the change in the result. However, a potential effect of lifting objects at home on risk of PTB cannot be firmly established because the available data were derived from a limited number of studies.
Implications for clinical practice and future research directions. Based on the best currently available evidence, the results of this meta-analysis show a beneficial effect of leisure-time PA during pregnancy in reducing the risk of PTB. The optimal dose of PA is still unknown, but the present results recommend that appropriate leisure-time PA during pregnancy has the potential to reduce the risk of PTB. This study also indicates that non-heavy domestic PA (e.g., care giving) might benefit the pregnant women with respect to PTB. Future studies, especially well-controlled experimental/interventional studies with sufficient power, are encouraged to better understand the dose-response relationship of leisure-time PA during pregnancy and the risk of PTB.
Our systematic review lends support to the hypothesis that leisure-time physical activity during pregnancy may protect against the incidence of preterm birth. Further studies are needed to identify the most appropriate levels of intensity, duration and frequency of leisure-time PA during pregnancy. Future studies should consider the four domains of PA and potential moderators (e.g., age, race), as well as utilize tools that reliably measure exposure variables. Such studies would provide useful guidelines for pregnant women and clinicians.
In conclusion, evidence from the observational studies suggests that leisure-time PA but not commuting PA during pregnancy was inversely associated with the risk of PTB. Domestic PA may provide a similar benefit, with the exception of lifting heavy objects. Results from the observational studies are not supported by the interventional studies that indicate null associations. Future studies are needed to determine the optimal intensity and frequency of leisure-time PA during pregnancy with respect to the risk of PTB and to elucidate the potential mechanisms.