Changes in cerebrovascular disease–related deaths and their location during the COVID-19 pandemic in Japan

Objective The COVID-19 pandemic placed an enormous strain on healthcare systems and raised concerns for delays in the management of patients with acute cerebrovascular events. In this study, we investigated cerebrovascular excess deaths in Japan. Study design Vital mortality statistics from January 2012 to May 2022 were obtained from the Japanese Ministry of Health, Labour and Welfare. Methods Using quasi-Poisson regression models, we estimated the expected weekly number of cerebrovascular deaths in Japan from January 2020 through May 2022 by place of death. Estimates were calculated for deaths in all locations, as well as for deaths in hospitals, in geriatric health service facilities, and at home. The age subgroups of ≥75 and <75 years were also considered. Weeks with a statistically significant excess of cerebrovascular deaths were determined when the weekly number of observed deaths exceeded the upper bound of 97.5% prediction interval. Results Excess deaths were noted in June 2021 and became more pronounced from February 2022 onward. The trend was notable among those aged ≥75 years and for those who died in hospitals. With respect to the location of deaths, the excess was significant in geriatric health services facilities from April 2020 to June 2021, whereas no evidence of excess hospital deaths was observed during the same period. Conclusions Beginning in the late 2021, excess cerebrovascular deaths coincided with the spread of the Omicron variant and may be associated with increased healthcare burden. In 2020, COVID-19 altered the geography of cerebrovascular deaths, with fewer people dying in hospitals and more dying in geriatric health service facilities and at home.


Introduction
The COVID-19 pandemic has had significant impacts on access to and quality of healthcare across the globe, and the management of cerebrovascular diseases is no exception. In addition to a purported association between COVID-19 and neurologic conditions such as cerebral infarction, 1 the early stage of the pandemic was characterized by decreased stroke consultations 2 and higher rates of mortality among acute ischemic stroke patients. 3 In Japan, significant excess all-cause deaths, the net difference between the number of deaths observed and expected on the basis of past trends, were not observed during 2020 4 but were temporarily recorded in 2021 5 in line with the surge of infection.
However, trends in cerebrovascular deaths during the pandemic era remain unclear, particularly during the recent spread of the Omicron variant. In the present study, we aimed to assess changes in cerebrovascular deaths from before and after the start of the COVID-19 pandemic using vital statistical data. We also aimed to assess the changes in the reported location of death among those who died due to cerebrovascular causes.

Data source and population
We used national mortality (vital statistical) data obtained from the Ministry of Health, Labour and Welfare between January 2012 and May 2022. Death certificates in Japan are prepared by a physician within one week of an individual's death and contain information regarding the place of death and underlying cause of death (based on the International Classification of Diseases, 10th Revision). All persons with a certificate of residence who died in Japan, regardless of nationality, are captured by the mortality data. Those whose place of residence or date of birth was unknown, those who died abroad, and those who stayed in Japan for a short period (without a residence card) were not included in the present study. Daily data were converted to weekly data.
We analyzed deaths where the underlying cause of deaths was a cerebrovascular disease (International Classification of Diseases, 10th Revision, classification I60eI69), including but not limited to cerebral infarction, subarachnoid hemorrhage, and intracerebral hemorrhage (see Supplementary Table 1). In other words, it should be noted that our analysis did not include deaths of patients with cerebrovascular disease who contracted COVID-19, became severely ill, and subsequently died (in this case, at least during the study period, the underlying cause of death would be recorded as COVID-19). In this study, 'all places' refers to total deaths across all places of death. Place-specific estimations were performed for hospitals, geriatric health service facilities (which are intermediate facilities that link hospitals and homes with the aim of supporting the independence of older people with disabilities and returning them to their homes; and the length of stay is also limited), and homes, where the number of weekly deaths was sufficient to avoid instability in the model. With respect to age, analyses were conducted for all ages in addition to those aged !75 and <75 years.

Statistical analysis
The Farrington algorithm, which is a variant of the quasi-Poisson regression model, was used to estimate the expected number of deaths for any given week between January 2018 and May 2022. Briefly, this algorithm uses historical data to construct a baseline from a five year moving window along with parameters to control for seasonality and then estimates the expected number of deaths and the corresponding two-sided 95% prediction intervals for any given week. The method is also described elsewhere. 5 We estimated the expected number of deaths per week for the study period, adjusting for linear trends and seasonality using the Farrington algorithm. 6 The expected number of deaths for a certain week, t, was calculated using data from weeks t À w and t þ w of years h À b and h þ b, where w and b are specified parameters and h is the year of t, referred to as the reference period. This method is intended to limit the data used for estimation. Thus, logðEðY t ÞÞ ¼ a þ bt þ f T ðtÞg f ðtÞ can be used to define the Farrington algorithm, where Y t is the number of deaths in a given week, and t is assumed to follow a quasi-Poisson distribution with a dispersion parameter. As shown in the previously mentioned equation, a and b are regression parameters, f ðtÞ is a vector of dummy variables that evenly splits time points outside the reference period, and g f ðtÞ is a regression parameter vector that represents seasonality. To regulate seasonality, the present study separated data for one year period that were not part of the reference period into nine evenly distributed segments, as done in prior studies. 7 We considered data up to five years prior (b ¼ 5) to the reference period and data from three weeks (w ¼ 3) before and after a certain week in the reference period, in line with previous studies. 7 For all estimates of expected weekly mortality, we calculated 95% two-sided prediction intervals.
The discrepancy between the observed and expected death toll was used to compute the number of excess deaths. We also defined percent excess as a relative measure of the magnitude of the excess, which was calculated as the number of excess deaths divided by the expected number of deaths. Weeks in which the observed number of deaths exceeded the upper bound or fell below the lower bound of the 97.5% prediction interval were considered weeks of statistically significant excess deaths or exiguous deaths, respectively. All analyses and visualization were conducted in R version 4.1.0 (R Core Team, Vienna, Austria).

Results
In 2020, no excess cerebrovascular deaths were observed, but in 2021 and 2022, consecutive weeks with excess deaths were noted from May 31 to June 13, 2021, and from February 7 to March 13, 2022 (Fig. 1A). A similar trend was observed among those aged !75 years (Fig. 1B) and in-hospital deaths (Fig. 1D). Excess home deaths were reported from the week of January 31 to February 13, 2022 (Fig. 1F). There were no consecutive weeks of excess deaths among those aged <75 years in 2022 (Fig. 1C). Among those who died in geriatric health service facilities, weeks of excess deaths were observed from April 13, 2020, to May 16, 2021 (Fig. 1E). For inhospital deaths, weeks of exiguous deaths were observed from January 20 to November 29, 2020 (Fig. 1D).
With respect to the percentage of the number of excess deaths divided by the expected number of deaths (percent excess deaths), many weeks showed a negative value in 2020 (mean À2.97%, standard deviation [SD] 3.57%; Supplementary Fig. 1A, Supplementary Table 2). Percent excess deaths began to uptrend in 2021, with positive values predominating in May 2021 (2021: mean 2.96%, SD 3.64%; 2022: mean 5.78%, SD 4.72%). Similar findings were noted among those aged !75 years (Supplementary Fig. 1B) and those who died in hospitals (Supplementary Fig. 1D). In contrast, among those who died in geriatric health service facilities ( Supplementary Fig. 1E), positive values of percent excess deaths began manifesting around April 2020 to June 2021, then reverted to primarily negative values in late 2021.

Discussion
In Japan, we found evidence of excess cerebrovascular deaths at the beginning of June 2021 and more predominantly from February 2022 onward. Excesses were particularly notable among those aged !75 years and those who died in hospitals. Notably, these excesses coincided with the beginning of the spread of the Omicron variant in Japan.
Increased cerebrovascular mortality postpandemic has been noted elsewhere. 3 The excesses in cerebrovascular deaths observed in Japan were noted slightly later than those in Western countries, possibly because of superior control of COVID-19 transmission and relatively low case counts until late 2021. During the spread of the Omicron variant, increased healthcare burden and the resultant effect on the quality of cerebrovascular disease management may have disproportionately affected older patients even more than during the initial COVID surges. 8 With respect to the location of deaths, our findings also suggest that during 2020, COVID-19 altered the geography of cerebrovascular deaths, with fewer people dying in hospitals and more dying in geriatric health service facilities. These periods largely corresponded with state of emergency declarations implemented throughout Japan, which encouraged Japanese residents to minimize outings and social interaction. A US study suggested that decreased healthcare-seeking behavior or fear of presenting to clinical spaces may have contributed to worse cerebrovascular disease outcomes. 9 Previous research has shown that Japanese residents avoided clinical spaces generally during the early phases of the pandemic, 10 which likely contributed to changes in the geography of cerebrovascular deaths. The reversion of mortality from non-hospital to hospital settings occurred in early 2022, at which point no states of emergency had been declared.
The present study has limitations. First, we specifically assess changes in cerebrovascular deaths; we do not assess changes in the incidence of cerebrovascular disease and cannot determine changes in case fatality rates for such patients. Because the COVID-19 pandemic led to widespread changes in healthcare-seeking behavior, excesses or deficits in cerebrovascular deaths may not exactly mirror changes in the incidence of disease. Second, the Farrington algorithm captures gradual changes in trends with the use of historical data; abrupt and pinpoint changes, such as those caused by new policy implementation, may not be fully absorbed by the model when constructing baselines for mortality prediction.
Author statements