Perinatal mortality after the Fukushima accident: a spatiotemporal analysis

After the accident at the Fukushima Daiichi nuclear power plant (FDNPP) in March 2011, little attention was paid to possible adverse effects on pregnancy outcomes in Fukushima and neighboring prefectures. The 2013 UNSCEAR report on Fukushima stated that prenatal exposures from the accident at FDNPP were not expected to increase the incidence of spontaneous abortions, miscarriages, perinatal mortality, congenital effects or cognitive impairment [1].

After the Chernobyl accident, a statistically significant 5% increase in perinatal mortality was observed in Germany in 1987, one year after the Chernobyl disaster [2, 3]. Scherb et al found an association of excess stillbirth rates in 1987 with the cesium deposition on a district level in Bavaria [3]. An analysis of monthly perinatal mortality rates found peaks of perinatal mortality rates at the beginning and end of 1987. The peaks were associated with peaks of cesium concentration in pregnant women from consumption of contaminated cow milk [2]. The estimate of the time-lag between the peaks of cesium burden and perinatal mortality was seven months. The effects were interpreted as possible teratogenic effects of radiation exposure in the period of major organogenesis, the most vulnerable period of fetal development [4].

A survey of stillbirth rates, preterm births, low birth weight, and congenital anomalies was conducted by researchers from Fukushima Medical University [5]. In women who were pregnant at the time of the Fukushima disaster no significant regional differences in stillbirth rates within Fukushima Prefecture were found. In a subsequent study, women who conceived during nine months after the disaster [6] were added. While no adverse obstetric outcomes were observed among women who conceived before the disaster, an increase in the incidence of preterm birth (less than 37 weeks) and low birth weight (less than 1500 g or less than 2500 g) was found in women who conceived within six months after the disaster.

Scherb et al reported an increase in perinatal mortality after the Fukushima nuclear accident [7]. Applying linear logistic regression, they determined a highly significant upward shift of mortality in 2012–2014 relative to the extrapolated trend before 2012. In a letter to the editor, Körblein stated that the shift in 2012–2014 is not statistically significant when a linear-quadratic temporal trend is used [8].

To overcome the uncertainty in model selection, Körblein and Küchenhoff conducted a combined regression of annual perinatal mortality data, 2002–2015, from the study region (Fukushima plus six nearby prefectures) and the control region (rest of Japan) with a common asymptotic lower limit of perinatal mortality [9]. They also found a significant upward shift in perinatal mortality in 2012–2015, but the increase was substantially smaller than reported by Scherb et al in [7].

The present study uses monthly data of perinatal mortality which allows to determine a possible seasonal pattern of the excess mortality in 2012–2017. This information may be of interest when investigating the causes of the increased perinatal mortality after Fukushima.

Numbers of live births, stillbirths and early neonatal deaths (first week), 2002 through 2017, are provided online by the Japanese government [10]. Stillbirths are defined in Japan as fetal deaths after 22 weeks of pregnancy. Perinatal mortality is defined as the number of stillbirths plus early neonatal deaths, divided by the number of live births plus stillbirths.

The prefectures Fukushima plus four neighboring prefectures (Ibaraki, Miyagi, Gunma, and Tochigi) were chosen as the study region (see figure 1). The remaining 42 prefectures of Japan were used as the control region.

Zoom In Zoom Out Reset image size

Table 1 contains the numbers of live births (LB), stillbirths (SB), and early neonatal deaths (NEO) in the study region and the control region, for two periods, 2002–2011 and 2012–2017. The study region was further divided in two regions of different mean radiation dose: (A) Fukushima Prefecture, and (B) the four neighboring prefectures (Miyagi, Gunma, Tochigi, and Ibaraki).

Statistical analysis

We compare the temporal trend in the study region with the trend in the control region conducting a spatiotemporal analysis of perinatal mortality in Japan before and after the Fukushima accident.

Our regression model for the null-hypothesis, i.e. without an effect of the Fukushima accident on perinatal mortality, has the following form.

$$\begin{eqnarray}&&E\left({y}\left(t,r\right)\right)=\alpha +\exp \left({\beta }_{1}+{\beta }_{2}\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}+{\beta }_{3}\cdot t+\beta \cdot t\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}\right).\end{eqnarray} \tag{ 1 }$$

Here, $E(y\left(t,r\right))$ is the expected value of perinatal mortality $y(t,r)$ where $t$ is time, i.e. calendar year minus 2000 in fractions of a year (e.g. mid-January 2002 means $t=2+1/24$), and $r$ is the region, i.e. study- and control region, respectively. Dummy variable ${\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}$ denotes the study region (${\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}=1$ for the study region and ${\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}=0$ for the control region). Parameter $\alpha$ is the asymptotic lower limit of perinatal mortality, and ${\beta }_{1}$ through ${\beta }_{4}$ are trend parameters.

Since we intend to compare our results with those reported in [7], we use January 2012 as the beginning of the test period for a possible level shift. An indicator variable $cp$ (change point) is defined as $cp=1$ in January 2012 through December 2017 in the study region, and $cp=0$ otherwise. The regression model has the following form:

$$\begin{eqnarray}&&\begin{array}{l}E\left(y\left(t,r\right)\right)=\alpha +\exp \left({\beta }_{1}+{\beta }_{2}\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}+{\beta }_{3}\cdot t+{\beta }_{4}\cdot t\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}\right.\\ \,\,\,\,\,+\,\left.cp\cdot \left({\beta }_{5}+{\beta }_{6}\cdot \left(t-12\right)\right)\right).\end{array}\end{eqnarray} \tag{ 2 }$$

Here, parameter ${\beta }_{5}$ estimates the size of the level shift in 2012–2017, and the interaction term ${\beta }_{6}\cdot (t-12)$ allows for a different slope in the study region in 2012–2017 than before 2012.

We determine the number of excess perinatal deaths in 2012–2017 in the study region from the difference of the sum of observed (O) cases in 2012–2017 and the sum of expected (E) cases, where E is calculated with the reduced model, i.e.

$$\begin{eqnarray*}&&E=\left(\alpha +\exp \left({\beta }_{1}+{\beta }_{2}\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}+{\beta }_{3}\cdot {\rm{t}}+{\beta }_{4}\cdot {\rm{t}}\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}\right)\right)\cdot ({\rm{L}}{\rm{B}}+{\rm{S}}{\rm{B}}).\end{eqnarray*}$$

The overall increase in perinatal mortality is sum(O) minus sum(E), divided by sum(E).

Next, we check the data for a possible periodicity of the effect in 2012–2017. To this end, we add two pairs of sine and cosine functions with periods of six and 12 months to the regression model, a method described in more detail in [11]:

$$\begin{eqnarray}&&\begin{array}{l}E\left(y\left(t,r\right)\right)=\alpha +\exp \left({\beta }_{1}+{\beta }_{2}\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}+{\beta }_{3}\cdot t+{\beta }_{4}\cdot t\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}\right.\\ \,\,\,\,\,+\,cp\cdot \left({\beta }_{5}+{\beta }_{6}\cdot \left(t-12\right)\right.+{\beta }_{7}\cdot \,\sin \left(2\pi \cdot t\right)+{\beta }_{8}\cdot \,\cos \left(2\pi \cdot t\right)\\ \left.\,\,\,\,\,+\,\,{\beta }_{9}\cdot \,\sin \left(4\pi \cdot t\right)+{\beta }_{10}\cdot \,\cos \left(4\pi \cdot t\right)\right).\end{array}\end{eqnarray} \tag{ 3 }$$

To check a possible dependency of the level shift on radiation dose, the study region is subdivided in two areas of different mean dose: (A) Fukushima Prefecture; (B) the four neighboring prefectures (Miyagi, Gunma, Tochigi, and Ibaraki). According to UNSCEAR 2013 [1] Annex A, table 5, the effective dose to adults in the first year after the accident was 1.0–4.3 mSv in Fukushima Prefecture and 0.2–1.4 mSv in the four neighboring prefectures plus two nearby prefectures (Iwate to the north and Chiba to the south of the study region). A combined regression of the data from areas A, B, and the control region with individual intercepts and level shifts ($cp$) is conducted. The model has the following form, where A, B are dummy variables for the respective regions:

$$\begin{eqnarray}&&\begin{array}{l}E\left(y\left(t,r\right)\right)=\alpha +\exp \left({\beta }_{1}+{\beta }_{2}\cdot A+{\beta }_{3}\cdot B\right)+{\beta }_{4}\cdot t+{\beta }_{5}\cdot t\cdot {\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}\\ \,\,\,\,\,+\,\left.cp\cdot \left({\beta }_{6}\cdot A+{\beta }_{7}\cdot B+{\beta }_{8}\cdot \left(t-12\right)\right)\right).\end{array}\end{eqnarray} \tag{ 4 }$$

Iteratively reweighted nonlinear regression with statistical program R [12] function $nls$ is used for the data analysis and plotting. For smoothing, moving averages of perinatal mortality are plotted. They are defined as averages of the rates in the present month, the preceding month, and the following month. Throughout the study, a two-sided $p$ value <0.05 is considered statistically significant.

A combined regression of the data from the study region and the control region (rest of Japan) with model (2) yields a significant level change in 2012–2017 (p = 0.006, $F$ test). The overall increase in perinatal mortality in 2012–2017 relative to the predicted trend is estimated as 10.6 (95% CI: 1.5, 25.2) % which translates to 195 (28, 462) excess perinatal deaths. Figure 2 upper panel shows the trends of perinatal mortality rates in the study- and region and the respective regression lines with model (2). The lower panel contains the residuals in units of standard deviations, which, in 2012–2017, stay well within the range of two standard deviations.

Zoom In Zoom Out Reset image size

Next, model (3) is applied which allows for periodic peaks in 2012–2017. Mortality peaks are determined around April, and the effect of periodicity is statistically significant (p = 0.045, F test). Figure 3 shows perinatal mortality rates from the study region in 2012–2017. The periodic peaks obtained with model (3) conform to the trend of the three month moving average of perinatal mortality rates.

Zoom In Zoom Out Reset image size

The improvement of the fit with increasing model complexity is shown in table 2. The dispersion factor, i.e. deviance divided by the degrees of freedom (dev/df), is a measure of the goodness of fit. Figure 3 shows the trend of the data.

The parameter estimates with standard errors (SE) for the three models are given in supplementary material table S1, available online at stacks.iop.org/JRP/39/1021/mmedia.

Dose dependency

To determine a possible dependency of excess perinatal mortality on mean dose in the first year, model (4) is applied which allows for individual level shifts in Fukushima Prefecture (area A) and in the four neighboring prefectures (area B). The increase is estimated as 24.8% in area A (p = 0.0009) and 8.0% in area B (p = 0.020). The improvement of the fit with model (4), relative to a model with a common level shift in areas A and B, is statistically significant (p = 0.010), and the excess perinatal mortality in 2012–2017 is 3.1 times greater in Fukushima Prefecture than in the four neighboring prefectures. The results of the regression with model (4) are contained in supplementary material table S2.

Figure 4 shows the trend of perinatal mortality rates in Fukushima Prefecture (area A) and the four neighboring prefectures (area B) with the respective regression lines. Perinatal mortality rates in Fukushima Prefecture and their moving average are shown in supplementary material, figure S1.

Zoom In Zoom Out Reset image size

This study finds a significant increase in perinatal mortality rates in 2012–2017 in a study region consisting of Fukushima Prefecture and four neighboring prefectures. Our result differs from that obtained by Fujimori in [5] who found no significant regional differences in stillbirth rates within Fukushima Prefecture. To reproduce the results, we grouped the data for stillbirths from their table 3 in three categories of increasing proximity (the proxy for dose) to the Fukushima Daiichi NPP; (1) Aizu and Kennan, (2) Kenpoku, Kenchu, Iwaki; and (3) Soso where the Fukushima Daiichi Nuclear Power Plant is located (see figure 1 in [5], available at https://jstage.jst.go.jp/article/fms/60/1/60_2014-9/_pdf/-char/en).

A Poisson regression of stillbirth rates in the six areas by dose category yields an increasing trend of stillbirth rates with proximity to the site (p = 0.057); the mean rates are 0.00064 (1/1564), 0.00274 (17/6207), and 0.00442 (4/904) in dose category 1 through 3 (see supplementary material figure S2). The stillbirth rate in Soso is 90% greater than in the remaining five areas (p = 0.24) but the effect is based on only four cases in Soso which leads to a large confidence interval for the rate ratio (RR = 1.91, 95% CI: 0.55, 5.12). Altogether their study is based on only 22 stillbirths whereas our study region includes 1629 stillbirths plus 399 early neonatal deaths in 2012–2017.

In a follow-up study [6] by the same authors, more preterm births and lowered birth weights are found in pregnancies conceived in the first six months after the nuclear disaster, i.e. in babies born between December 2011 and June 2012 (see [6] table 2). In our study we find a peak of perinatal mortality in the study region in spring 2012, which concerns pregnancies conceived 4–6 months after the disaster, but not in those conceived in the first three months after the disaster. Their table 2 also contains data for congenital anomalies (CA). The only increase in CA rates is observed in pregnancies conceived during the three months before the disaster, i.e. who were in their first trimester p.c. at the time of the disaster.

The shift in perinatal mortality rates in 2012–2017 is considerably smaller in our study than reported by Scherb et al in [7]. Their 'severely contaminated' region included the prefecture Iwate-which is not a neighboring prefecture-in addition to Fukushima and the four neighboring prefectures of our study region. A minor difference between our study and that by Scherb et al [7] is their use of a dummy variable to adjust for a possible direct effect of the tsunami in March–May 2011. In our data we find a notable increase only in March 2011 which, however, is not statistically significant (O = 40, E = 30.3, RR = 1.32, p = 0.13).

The main difference between our studies is that Scherb et al limited their analysis to the data of the study region whereas we conduct a combined regression of perinatal mortality rates in the study region with the rest of Japan as a control region.

In a recent letter [13], Scherb et al presented results for perinatal mortality in their study region for an extended study period, 2002–2017. We reproduced their results, but we showed that the effect size differs considerably (11.6% increase in 2012–2017 versus 18.6%) when a modified regression model is applied that allows for a curvilinear trend and a slope change in 2012–17 (see supplementary material figure S3). Since the two competing models fit the data equally well, there is no way to decide which model is preferable.

As mentioned above, we chose January 2012 as the starting month for testing a possible increase in perinatal mortality after Fukushima to compare our results with those reported by Scherb et al in [7]. An independent justification for this choice might be that, after Chernobyl, a first rise in perinatal mortality was observed in January 1987 in Germany [2]. Since radiation exposure from ingestion of contaminated food (e.g. cow milk and leafy vegetables) depends on the growing season, peaks in cesium burden of pregnant women should occur at similar times of the year after Fukushima and after Chernobyl.

After Chernobyl, the increase in perinatal mortality in Germany was restricted to the first follow-up year whereas the increase after Fukushima is enduring though weakening over time. This might be explained by differences in nutrition between Germany and Japan. Consumption of milk and dairy products was the main path of radio-cesium in food after Chernobyl, but the cesium concentration in milk was negligible one year after Chernobyl. In Japan, consumption of wild mushrooms might explain the enduring effects observed in our study. The importance of mushrooms for internal radiation exposure is emphasised by Orita et al [14] who calculated committed effective doses from mushroom consumption in Kawauchi village, some 30 km from Fukushima Daiichi NPP.

Measurements of radio-cesium activity in food samples were published by the Japanese Ministry of Health, Labor and Welfare after the Fukushima nuclear accident [15]. Supplementary material figure S4, upper panel, shows the cesium concentration (Cs-134 plus Cs-137) in mushrooms from the study region, 2011 through 2017 and the result of a Poisson regression with a linear-quadratic time trend plus seasonal variations. The effect of periodicity is highly statistically significant (p < 0.001). Peaks in cesium concentration are determined in August–September while the peaks in perinatal mortality were found in April (see figure S5 lower panel). The time lag of 7–8 months between the peaks of cesium concentration in mushrooms and the peaks of perinatal mortality is biologically plausible, hence our suspicion of a link between mushroom consumption and perinatal mortality. But this remains speculative since we have no data on consumption of mushrooms.

The interpretation of our results as teratogenic radiation effects is in conflict with the prevalent opinion. Publication 90 of the International Commission on Radiation Protection (ICRP) [4] states that teratogenic effects are not expected to occur below a 'practical' threshold dose of about 100 mSv, a position adopted by UNSCEAR [1] and WHO [16]. But the estimated population doses in the first year were between 1 and 4 mSv in the Fukushima study region and 0.2 mSv in Germany after Chernobyl [17], two orders of magnitude less than the threshold dose. Already in 1989, however, a review article by Michel emphasised the high radio-sensitivity of the fetus, with the first trimester of pregnancy regarded as the period of highest risk for malformation and cancer induction [18]. Studies of perinatal mortality found increases in Germany [2], in Belarus, and in Ukraine [19] after Chernobyl. The prevalence of neural tube defects plus microcephaly was significantly increased in a highly contaminated region of Ukraine [20] after Chernobyl, and a study of the effects of atmospheric nuclear weapon tests in Germany associated perinatal mortality with the strontium burden in pregnant women [21].

The main limitation of our study is inherent for ecological studies: the results are based on highly aggregated data without individual dose measurements; causes other than radiation cannot be excluded. Migration does not influence the results when residents move within the study region, and migration out of the study region will likely decrease rather than increase the observed effect.

After the accident in the Fukushima Daiichi nuclear power plant in March 2011, a statistically significant increase in perinatal mortality rates is observed in 2012–2017 in a study region including Fukushima Prefecture and four neighboring prefectures. The excess is characterised by periodic peaks with maxima around April. The effect is three times greater in Fukushima prefecture than in the four neighboring prefectures. Due to its ecological design, the results need to be interpreted with caution as other possible causes for the increased perinatal mortality cannot be excluded. Continued observation of the trend of perinatal mortality in contaminated regions of Japan is recommended.

The authors are grateful to Masao Fukumoto for providing perinatal mortality data from selected Japanese prefectures and two anonymous reviewers for valuable comments.