Information about fetal movements and stillbirth trends: Analysis of time series data

To investigate the implementation of the Count the Kicks campaign in Iowa to increase maternal awareness of fetal movements and its association with stillbirth rates.


| I N TRODUC TION
Stillbirth, defined as the death of a baby before birth after 20 weeks of gestation, occurs in approximately 23 000 births annually In the USA. 1 Applying the World Health Organization definition for international comparison (≥28 weeks of gestation), the stillbirth rate in the USA ranks 23rd out of 49 high-income countries; concerningly, the annual rate of reduction (ARR) between 2000 and 2019 was only 0.5%. 2 There is also wide variation in both stillbirth rates and ARR between states. As stillbirths place significant psychological, social and economic burden on families, health services and wider society, efforts are urgently needed to reduce stillbirths. 3,4 Maternal perception of reduced fetal movements is related to stillbirth. 5 Women who track fetal movements less frequently are more likely to experience stillbirth, as are women who are less frequently asked to do so by healthcare providers. 6 Data regarding the effect of maternal awareness of fetal movements on the stillbirth rate are inconclusive. Four large cluster randomised studies have failed to demonstrate a statistically significant reduction in perinatal mortality, [7][8][9][10] whereas the results of a meta-analysis were consistent with a small reduction, but this finding was not conclusive (RR 0.92, 95% CI 0.85-1.00). 11 Furthermore, there have been few studies of the impact of maternal awareness of fetal movements on stillbirth rates in the USA. In Iowa, the charity Healthy Birth Day promotes maternal awareness of fetal movements through the Count the Kicks (CtK) campaign, with the goal of reducing stillbirth; the campaign includes the provision and promotion of information and a free mobile phone app that can be used to monitor fetal movements. This study aimed to investigate the implementation of this campaign in Iowa and to compare stillbirth rates there with the rates in three neighbouring states.

| M ET HODS
Prior to conducting this study, we posted a short protocol (https://osf.io/ktve9/). As all outcome data used were in the public domain, ethical approval was not required. We did not use a core outcome set for this study as presently there is no agreed core outcome set for studies of fetal movement. As this study did not include direct involvement with participants, we did not undertake any specific participant involvement activities in this project.
We assembled two versions of a time series data set covering the period 2005-2018 (inclusive), containing the numbers of stillbirths, live births and aggregated data related to known risk factors for stillbirth, for each of Iowa, Illinois, Minnesota and Missouri. The two versions considered data aggregated to 3-and 6-month time periods; we had anticipated a trade-off between granularity and precision. We undertook all analyses on both data sets to check the robustness of the results. Illinois, Minnesota and Missouri were selected as comparators because they neighbour Iowa. The inclusion of comparator states was intended to control for the possibility that any changes in stillbirth rates in Iowa could arise from general improvements in the standard of care over the study period or from changes in the risk factor profiles of women with pregnancies. The corresponding assumption is that these improvements would be present and comparable in the neighbouring states.

| Data sources and data preparation
The numbers of live births and stillbirths to US residents by month between 2005 and 2018 in the states of Iowa, Illinois, Minnesota and Missouri were obtained from the Centers for Disease Control and Prevention Wonder (CDC-W) system. 12,13 Full details of the data extraction processes are contained in the appendix 'Extraction of Centers for Disease Control and Prevention (CDC) Wonder Data' of the study protocol (https://osf.io/ktve9/).
Individual-level data were not available but it was possible to access aggregated data on known risk factors for the time periods in question from the CDC-W Natality database. 12 The following pre-specified risk factor variables were downloaded by state, year and month: mean maternal age, percentage of live births that were first live births (birth order), percentage of women who smoked cigarettes, percentage of women with maternal diabetes, percentage of women with chronic hypertension (HT), percentage of women with pregnancy-associated HT and percentage of births delivered by caesarean section. As body mass index (BMI) was not available before 2016, we used the Behavioural Risk Factor Surveillance System (BRFSS) Prevalence and Trends Data to get an estimate for the percentage of women who were obese (BMI ≥ 30 kg/m 2 ) amongst US adult women in each year. 14 This database was also used to obtain information on the annual percentage of US adult women who were current smokers, as CDC-W had gaps in coverage (for this variable only) for the three neighbouring states (unavailable for 2007-2009 inclusive for Illinois and Missouri; suppressed for 2007-2011 inclusive for Minnesota). As data on deprivation was not available on CDC-W, we used US census data on poverty (percentage below 100% of poverty using weighted person count estimates) at the state and year level from the Current Population Survey (CPS). 15 We also obtained data corresponding to the implementation and uptake of CtK activities from Healthy Birth Day. Their CtK campaign included the distribution of educational materials and eventually the release of a 'kick-counting' app, which can be used to monitor fetal movements during pregnancy. Thus, implementation included annual counts of users of the kick-counting app in each of the four states, and data describing the delivery of educational materials to counties within Iowa. We used the app data to aid interpretation (as wide uptake of the app in the comparator states would have prevented us from attributing any differences in stillbirth rates to the app) and used the educational materials data for analyses at the county level (described below).

| Division into time tranches
Evaluating the association between the CtK campaign and stillbirths is complicated by the fact that the charity was active throughout the study period. We considered state-level stillbirth trends over the whole period and also in relation to key implementation phases in the campaign. The study period was divided into time tranches relating to these implementation phases. These were selected on the basis of discussion with Healthy Birth Day prior to assembling or analysing the data sets, and were prospectively specified in the outline protocol (https://osf.io/ktve9/). We divided the study period into the inclusive tranches 2005-2007 (prior to the launch of CtK), 2008-2013 (covering early CtK activity, prior to the launch of the kick-counting app), 2014-2016 (following the launch of the app), and 2017-2018 (following promotional appearances on Good Morning America and a wide variety of national media outlets, resulting in a large increase in downloads of the app).

| County-level analyses
In order to consider the potential impact of educational materials supplied by Healthy Birth Day to counties in Iowa, we examined the annual stillbirth rates in three large counties, Polk, Linn and Black Hawk. County-level stillbirth data were obtained from Iowa Department of Public Health. 16 We selected large counties for this analysis because stillbirth counts below 10, which would occur in smaller counties, would not be disclosable.

| Statistical methods
We tabulated the annual number of app users in each state, as a measure of uptake. All state-level analyses were conducted on both the 3-monthly aggregated data and the 6-monthly aggregated data. County-level analyses were based on annual data, as increasing granularity beyond this level would result in an excessive portion of suppressed data. Only the CDC-W smoking variable had missing values. Multiple imputation was used for the missing CDC-W smoking values using each of the extracted risk factors, including BRFSS obesity and smoking variables, the CPS poverty values, time and the percentage of stillbirths out of the total births as regression covariates based on 100 imputations.
We plotted each risk factor against time by state, to consider whether changes in the risk factors could account for changes in stillbirth rates, and whether differences in these trends between states could account for inter-state differences in stillbirth trends. In particular, we examined whether any sharp changes in risk factor trends were apparent, which could account for any sharp changes in stillbirth trends. In the absence of sharp changes over time or distinct differences between states, it becomes more plausible that any such changes or differences may be attributable to the intervention. This type of reasoning plays an important role in time series analyses, as it can be difficult to distinguish intervention effects from confounder effects by way of statistical adjustment; as both intervention and confounding factors change over time, they are collinear.
All subsequent analyses involved fitting logistic regression models to the aggregated time series data, with stillbirths as a proportion of all births (live plus stillbirths) as the outcome variable. We performed analyses of each state individually, before fitting models to directly compare states. 17 Our simplest models were unadjusted state-specific models, with time as the only covariate. These estimate the trend in stillbirth rates over the study period. We attempted to fit a further version adjusted for the nine risk factors (mean maternal age, birth order, percentage of smokers, percentage with maternal diabetes, percentage with chronic HT, percentage with pregnancy-associated HT, percentage with caesarean delivery, percentage with obesity and percentage in poverty). Two problems became apparent with this model. First, some of the adjustment variables could plausibly be acting as mediators rather than confounding factors. Second, the fully-adjusted models displayed separation and did not converge to a solution. 18 To explore the latter issue, models were fitted with each confounding factor individually. It became clear that mean age was causing the problem with the fitting. Centering mean age (by subtracting the state-specific mean age for the whole study period from each mean age value) solved the convergence problem.
Five covariates were subsequently selected to include in the adjusted models as confounding factors (excluding potential mediators): percentage of smokers, percentage with chronic hypertension, centred age, percentage of obesity and percentage in poverty. This was based on clinical expertise, informed by the epidemiological literature. 19,20 We fitted unadjusted and adjusted versions of the model for each state with the addition of a time-tranche interaction (with four tranches corresponding to the implementation phases). The impact of each of the implementation steps would not be instantaneous, and so we were primarily interested in the rate of change in stillbirth for each of the studied tranches; these models evaluate whether slopes differed between tranches. The unadjusted interactions were tested by likelihood ratio test. The interactions for the adjusted models were tested by Wald test.
Analyses were then conducted to compare the states directly. First, we compared the stillbirth trends over the study period with a model including time and state, and their interaction. Next, we considered whether the tranche effects on the slopes differed between the four states (a three-way interaction between state, tranche and time). In an alternative version of this analysis, we coded Iowa as the 'intervention' state and the other three as 'control' states. We considered whether the tranche effects on the slopes differed in Iowa compared with our control states. State-specific intercepts had to be omitted from this model (a departure from our protocol) as they were not identifiable when included along with a binary term for intervention/control status.
In the analysis of three counties within Iowa, we had access to annual stillbirth data only. We looked at the year when each of these counties initially ordered information materials from Healthy Birth Day, and created a binary variable denoting whether each year fell before or on/after that year. Healthy Birth Day confirmed that materials were delivered promptly following each order. We then compared stillbirth trends before and on/after this year by fitting a logistic regression for grouped data for each county, with year, the binary indicator (pre-or post-order of information materials) and the interaction between these terms. This interaction was then tested by likelihood ratio test to consider whether the slope changed after materials were ordered. No sample size calculation was performed, as this was not under our control in this study. Table 1 shows the numbers of births and users of the app. App use was substantially greater in Iowa compared with the comparator states, both in absolute number and relative to the annual number of births. A national TV appearance promoting the app appears to have translated into a substantial increase in users in all states in the period 2017-2018, although in the comparator states the numbers of users remained low relative to the number of births. Figures 1 and 2 show time trends for risk factors for stillbirth in Iowa and the three comparator states. Trends were generally similar in all states, and did not display sudden changes over the study period. The majority of the risk factors became more prevalent over the study period, with the percentage of smokers being an exception. Table 2 shows the annual stillbirth trends over the study period in each state and, for comparison, in all US states combined. Figure 3 shows stillbirth trends over the study period in each state in 3-month intervals. We present the 3-month analyses here and note any differences with the 6-month analysis (presented in Figure S1). Of the four states, only Iowa demonstrated a clear decrease in stillbirths, which was statistically significant (OR 0.99, 95% CI 0.99-1.00, per 3-months; OR 0.96, 95% CI 0.96-1.00 when rescaled to reflect reduction per year; p < 0.001; Table S1). This would equate to an illustrative reduction from 5.8 stillbirths per 1000 births to 4.7 stillbirths per 1000 births over a 5-year period. There was no evidence of any trends in Illinois and Missouri, whereas there was evidence of an increase in stillbirths over the study period in Minnesota (Table S1).

| Trends in stillbirths
The red vertical lines in Figure 3 divide the study period into designated time tranches. In Iowa, there was a decline in stillbirths in the period 2008-2013, commencing with the launch and covering the early years of the CtK campaign. This was followed by an estimated increase over the period 2014-2016, following the launch of the app, and an estimated decrease over 2017-2018, following the national TV appearance and increase in app users. The differences in slopes were not significant, however (test of interaction: p = 0.060, without adjustment; p = 0.142, adjusting for confounding factors; Tables S2 and S3).
In post-hoc analyses comparing the last two periods in Iowa there was evidence of a change in slope following the TV appearance (test of interaction: p = 0.044, unadjusted for confounding factors). However, after adjusting for confounding factors there was no evidence of a change in slope (p = 0.812; Table S4). Stillbirth trends in the time tranches were qualitatively similar in Missouri and Iowa (Figure 3) (tests of interaction: p = 0.063 and p = 0.038 for unadjusted 3-and 6-month analyses, Tables S2 and S5; p = 0.319 and p = 0.164 for adjusted 3-and 6-month analyses, Tables S3  and S6).
Models comparing the states directly were fitted. The decrease in Iowa was statistically significantly greater than in each of the comparator states (test of interaction between state and time: p < 0.001, unadjusted and adjusted for both the 3-and 6-month data, Tables S7 and S8). We considered whether the tranche effects on the slopes differed in Iowa compared with the control states, which could be taken as evidence of an impact associated with specific stages of the implementation of the CtK campaign in Iowa. For the 3-month data there was no evidence of this in either unadjusted (p = 0.389) or adjusted models (p = 0.221) (Table S9). Similarly, there was no evidence using the 6-month data (Table S10). This conclusion did not change when we considered state as a binary variable (Iowa vs others) (Tables S11 and S12).  3.4 | Did the provision of educational materials at county level impact stillbirths in Iowa? Figure 4 shows the annual stillbirth rates in these counties prior to and following the first orders of educational materials. There was no evidence of a change in slope following the distribution of materials in any of the counties studied (p = 0.948).

| DISCUS SION
This study evaluated stillbirth trends in Iowa between 2005 and 2018 in comparison with neighbouring states and in relation to the activities of a state-wide campaign to reduce stillbirths by encouraging maternal awareness of fetal movements. A particular interest of the study was to consider the uptake and potential effectiveness of these interventions in a real-world context. We found that the uptake of a kickcounting app was modest relative to the number of births, although a national television campaign appears to have been successful in significantly increasing the numbers of users. This increase in use coincided with a decline in the stillbirth rate in Iowa, although the available data do not allow us to causally connect these events. We were not able to clearly associate any change in stillbirth trends with any particular stage in the campaign. Much of the overall decline in stillbirths occurred prior to the launch of the app. The decrease in stillbirths in Iowa contrasts with neighbouring states and national trends in the USA. Adjusting for potential confounding is difficult in this context, as confounding factors are also subject to time trends, resulting in collinearity in statistical models. Accordingly, estimates became imprecise with adjustment. Examining key potential confounding factors showed that, with the exception of smoking, they were increasing in frequency, and also that trends were similar in Iowa and the comparator states. This suggests that changes in confounding factors are not responsible for the observed decreases in stillbirth in Iowa. Furthermore, there have been no nationally implemented initiatives or guidelines relating to the awareness or management of fetal movements in the USA. Through discussions with stakeholder organisations we are aware of the Iowa Barriers to Prenatal Care project (active since 1991) and the Iowa Stillbirth Surveillance Project (last active publication in 2015); however, as these largely pre-date the study period, the changes in stillbirth rate since 2012 are unlikely to be related to these initiatives. Nonetheless, it is likely that changes in clinical practice have occurred over the study period, which would impact stillbirths. The inclusion of comparator states was intended to control for these non-specific changes, under the assumption that they would be comparable in neighbouring states. An important caveat is that all of our analyses were based on aggregated rather than individuallevel data, conferring a risk of ecological fallacy (incorrectly inferring that relationships observed at the aggregate level apply at the level of individuals). Moreover, with the data available we were unable to realise supplementary analyses around potential iatrogenic effects (for example, relating to early term births) or by gestational age.
Data regarding maternal awareness of fetal movements from 633 women indicate that women whose pregnancy ended in stillbirth were less likely to report having been aware of fetal activity and less likely to have been told to monitor fetal activity by healthcare professionals. 6 To address this, parent-led organisations such as Healthy Birth Day have developed written materials and a mobile app; an evaluation of 809 app users found that it increased maternal awareness of fetal movements and more frequently prompted assessment by professionals for maternal concerns. 21 It is well recognised that women obtain information about the progress of their pregnancy and information about signs and symptoms of pregnancy complications from a variety of sources, particularly the internet. 22,23 Thus, evaluating the impact of initiatives such as CtK is important.
Quantifying the impact of kick-counting information is challenging because, as observed in cluster randomised controlled trials, 7 only a small proportion of pregnant women downloaded the app. However, this does not reflect the proportion of women who may have accessed flyers, posters and paper leaflets. It is notable that over the period studied, the Iowa Barriers to Prenatal Care project reported that care providers increased the provision of information about fetal movements from 72% to 84% of respondents. 24 The data presented here suggest that the promotion of fetal movement awareness in the media is associated with a fourfold increase in the number of users. Thus, these engagement strategies could be usefully employed to disseminate health promotion advice and activities to pregnant women.
A meta-analysis of the impact of interventions relating to maternal awareness was inconclusive, suggesting that there was a reduction in perinatal mortality (RR 0.92, 95% CI 0.85-1.00) but no effect on stillbirth (RR 0.94, 95% CI 0.71-1.25). 11 However, this systematic review and meta-analysis grouped together interventions that raised awareness of fetal movements and provided intervention for reduced fetal movement; these two related interventions may have different effects. The prior Cochrane review that specifically assessed counting fetal movements found insufficient evidence to determine whether this approach reduced perinatal mortality in randomised controlled trials (RCTs); this Cochrane review identified relatively little data relating to stillbirth, and so was unable to determine whether counting fetal movements was effective for reducing this outcome. 25 Critically, both RCTs reporting on stillbirth were conducted in high-income countries (HICs), such that the outcome rates were relatively low.

| Conclusion
In 2015 the USA was ranked 23rd out of 49 HICs for stillbirth rate, with a low annual rate of reduction of stillbirth. 26 Thus, the identification of effective strategies to reduce stillbirth are urgently needed. It is clear from the data analysed here that the stillbirth rate in Iowa is decreasing, whereas this is not true for the other three states and for the USA as a whole. 27 These data suggest that lessons could be learned from Iowa, where reductions are being achieved. One such intervention is the Healthy Birth Day, and other programmes, such as the Iowa Stillbirth Surveillance Project, have also been active in this area. We have not been able to specifically determine whether individual activities are associated with a reduction in stillbirth or whether this may come from the raised awareness of stillbirth prevention per se. Large-scale intervention studies are needed and given the large number of births in the USA, with centrally collected data, this would be an ideal environment for such investigations.

AU T HOR C ON T R I BU T ION S
AEPH and JW contributed to all aspects of the study design and obtained funding. AEPH had overall responsibility for the study. FH obtained data from online sources. FH and JW analysed the data with input from AH. All authors were responsible for drafting the article. All authors approved the final version for publication.