Categorization of disaster-related deaths in Minamisoma city after the Fukushima nuclear disaster using clustering analysis

The medical situation during disasters often differs from that at usual times. Disasters can lead to significant mortality that can be difficult to monitor. The types of disaster-related deaths are largely unknown. In this study, we conducted a survey to categorize the disaster-related deaths caused by a radiation disaster. A total of 520 people living in Minamisoma City, Fukushima Prefecture, at the time of the Fukushima Daiichi Nuclear Power Plant accident, who were certified to have died due to disaster-related causes were surveyed. We divided the participants into those who were at home at the time of the earthquake and those who were in hospitals or facilities when the disaster struck and conducted a hierarchical cluster analysis of the two groups. Disaster-related deaths could be divided into seven groups for those who were at home at the time of the disaster and five groups for those who were in hospitals or facilities at the time of the disaster. Each group showed different characteristics, such as "the group with disabilities," "the group receiving care," and "the group with depression," and it became evident that not only uniform post-disaster support, but support tailored to the characteristics of each group is necessary.


Clustering background information
To determine how disaster-related deaths could be classified according to the 24 types of background information, excluding the direct causes of death, a hierarchical cluster analysis was performed on the background information using the Ward method 36 with square Euclidean distance.Previous studies in Fukushima have revealed differences in the characteristics of those who died at home compared to those who died in hospitals or institutions 14,15,37 .Therefore, statistical analysis was performed to compare the background information between those who died at home and those who died in a hospital or institution, and significant differences were found (Supplemental Table 1).Hierarchical cluster analysis was conducted separately for those who were at home and those who were in hospitals or institutions at the time of the earthquake.One person whose residence at the time of the earthquake was unknown was excluded from the analysis.While conducting the analysis, there were 17 items of binary data and continuous variables: sex, age at the time of death, whether or not the person was evacuated, number of times evacuated, number of times moved, situation at the time of move, location of move, family situation (which meant single at the time of death or not), disability, whether or not the person was receiving nursing care at the time of the disaster, increased drinking opportunities and alcohol consumption, insomnia, depressed mood, development or severity of dementia, decreased social participation and activities, decreased communication, and date since disaster.Missing values were added using the multiple assignment method.Each variable was standardized (Z-score) to ensure that the data scale was uniform.The number of clusters was determined by combining the tree diagrams (Fig. 2), which were created when the Ward method was implemented, and the interpretability of the results was good.After clustering, color coding was performed for each cluster class on the dimensionality-reduced data using the t-SNE 38 algorithm to visualize the results (Fig. 3).

Analysis of cluster results and their factors
To clarify the relationship between the background information and the degree to which the results of the cluster performed earlier were influenced by the factors, a table summarizing the relationship between each class and variable was created (Tables 2 and 3).Categorical variables were counted by category, and continuous variables were derived as means and variances.The number of participants before cluster analysis was sufficient, but some clusters had few participants, and these were considered the minority.In addition, to clarify the variables that contributed to the cluster results, the Fisher's exact method was applied to the cluster results for categorical variables, one-way analysis of variance (ANOVA) was applied to continuous variables, and effect sizes were derived.V ≥ 0.50 and η 2 ≥ 0.14 were determined to have a large effect size 39 .Fisher's exact test or one-way ANOVA was performed only for groups with 30 or more participants.In contrast, descriptive statistics were used for   minority groups with less than 30 participants.Finally, the experts discussed the results and created a graph to interpret the relationship between the cluster results and background factors.(Fig. 4).These analyses were performed using python version 3.9.7.

Ethics declarations
This study was approved by the Minamisoma Municipal General Hospital Ethics Committee (approval number:2-21) and the Fukushima Medical University Ethics Committee (reference number:2020-297).As data were evaluated retrospectively and pseudonymously and were solely obtained for treatment purposes, the requirement for informed consent was waived by the ethics committees of Minamisoma Municipal General Hospital and the Fukushima Medical University.The study was conducted in accordance with the tenets of the Declaration of Helsinki.
Factors that largely distinguished the characteristics of each group were the presence or absence of a disability, receiving care certification, evacuation, and depression (Fig. 4).

Discussion
In this study, the average age of those identified as disaster-related deaths was 82.7 ± 11.9 years, which is consistent with previous reports that these deaths tend to occur in people in their 70 s and 80 s, with no significant sex bias 7,10,40 .Among people who were at home at the time of the disaster, only three of the seven groups received nursing care, whereas among those in a hospital or facility, only one of the five groups did not receive nursing care.This analysis is significant, as it helped extract a population that did not receive nursing care from a population with many care recipients and vice versa.
This study showed that disaster-related deaths in a radiation disaster can be divided into seven groups for those who were at home at the time of the disaster and into five groups for those who were in a hospital or at a facility at the time of the disaster.Clear differences were evident in the groups' causes of death.The people at home had more deaths from circulatory diseases and malignant neoplasms, which accounted for approximately half of all deaths.While those in a hospital or facility had more deaths due to respiratory diseases and senility, which also accounted for one-third of all deaths.The two groups also differed in terms of the diseases to which they were prone.
However, the two groups also shared some similarities.Regardless of the location of the disaster, people in the groups that had a longer period between the disaster and their death experienced depressive feelings.Most of the individuals belonging to the people at home and in a hospital or facility exhibited depressive feelings after an average of 100 and 170 days, respectively, suggesting that they may have been more likely to become depressed 3 months following the disaster.This is possibly because long-term evacuation prolongs the effects of environmental changes, which consequently increases the psychological impact on the survivors, thus raising the risk of death.This is consistent with previous studies which reported that long-term depressive feelings after disasters are more likely to lead to death 25,26,41,42 .
Notably, that the time to death in clusters I4 and H5, which included people with disabilities, was approximately 610-670 days, which was longer in comparison to that in other clusters, suggesting that the characteristics of disaster-related deaths among people with disabilities may differ from those of other groups.The reason for this is thought to be that disaster victims weaken over the long term due to prolonged evacuation and changes in post-disaster medical services that gradually affect their lives.Disabilities that did not have Vol:.(1234567890  www.nature.com/scientificreports/ a significant impact on them before the disaster gradually begins to affect their lives.Previous studies have reported that people with disabilities take longer to recover after disasters 8,43 and there have been cases of people dying from long-term effects of disasters 37 .These findings are consistent with the present study results, which showed that people with disabilities succumb to long-term effects.Interview research to clarify how people with disabilities are affected and weakened is warranted in the future.People living in institutions or hospitals at the time of the disaster were divided into three groups according to the duration of time before their death.Individuals belonging to one group died within 2 months, another within 3-6 months, and the other within 6 months due to the acute, subacute, and chronic effects after the earthquake, respectively.Previous similar studies have shown that people in facilities die within one to two months after a disaster, primarily due to evacuation 44 .However, in this study, we found that some people died before evacuation, for example class I 2 , and others died after six months, such as class I4 or I5, which had not been found in previous studies.Future studies are required to accumulate more research evidence regarding such groups. This study revealed that people without receiving and not receiving nursing care who were living at home at the time of the disaster and had depressive tendencies could be divided into two main categories: those with reduced social activity and communication with others (such as class H5) and those without these characteristics (such as class H3).In the first category, it is considered that the environmental factors due to prolonged evacuation resulted in reduced social activity and caused depressive tendencies.However, in the second group, it is contemplated that these characteristics were not related to the disaster, but occurred because the person originally showed depressive tendencies, such as mental illness and cancer.Previous studies [45][46][47][48] have suggested that people with depression were affected by changes in the social environment after the earthquake.However, some people who have depression without such effects, which seems to be a new finding.
An important finding is that while some debilitated people who originally had serious illnesses died within three months of the disaster, others lingered beyond three months and developed depression.Moreover, these long-term survivors not only included originally healthy individuals, but also those with disabilities and preexisting medical conditions.These facts suggest that post-disaster support for disaster victims may require not only uniform support, but also group-specific continuous support.In addition, mental health care was required 3 months after the disaster.These are important findings on chronic disaster care for future disaster medicine.

Limitations
However, this study had a few limitations which need consideration.First, the classifications are used to observe group tendencies and do not consider every individual's behavior.This aspect should be considered when conducting qualitative surveys in the future to determine individual characteristics.Furthermore, a more detailed analysis with the population stratified by various demographic factors, such as age and sex, is necessary.Second, data were registered based on applications from family members of those who died in disaster-related deaths, which were classified by the authors.This may have resulted in some bias because some parts were not based on objective indicators.However, the judgments by multiple raters would have reduced this bias.Third, there was a large amount of missing data regarding questions such as motivation to engage in social activities and depressive tendencies, with sleep disorders at the top of the list.These were complemented by the multiple assignment method, which has limited accuracy.However, when a similar analysis was conducted using only data that did not include missing values, the participants were classified into similar clusters, suggesting that the multiple assignment method is a reasonable complement.

Conclusion
Radiation disaster-related deaths can be divided into seven categories for those who were at home at the time of the disaster and into five categories for those who were in hospitals or facilities at the time of the disaster.These clusters can be classified based on "residential status at the time of the disaster, " "receiving long-term care certification, " "presence of disabilities, " "presence of depression, " and "experience of evacuation." Moreover, these clusters included "people who did not evacuate after the disaster and stayed at home, " "people with disabilities, " "people receiving care," and "people with depressive feelings."Some of the individuals belonging to these categories originally had serious illnesses and died within three months of the disaster, while others survived until after three months and began to experience depressive feelings.In addition, each category had distinct diseases that were likely to be the cause of death, such as "the group with more deaths from malignant neoplasms, " "the group with more deaths from respiratory diseases, " and "the group with more deaths from debilitating diseases".The prolonged group not only included those who were originally healthy, but also those with disabilities and pre-existing medical conditions.These findings suggest that assisting disaster victims not only require uniform support, but also individual support based on the group and category they belong to.

Figure 1 .
Figure 1.Location of Minamisoma City and Fukushima Daiichi Nuclear Power Plant along with the "Preparation Zone for Lifting Evacuation Order, " "Restricted Population Zone, " and "Difficult-to-Return Zone" in Minamisoma City from April 1, 2012, to July 12, 2016.This map was created with ESRI's Arc GIS pro version 3.1.(https:// www.esrij.com/ produ cts/ arcgis).

Figure 2 .
Figure 2. The tree diagrams of hierarchical clustering used to determine the number of clusters.The clustering distance is the distance produced using the Ward method with square Euclidean distance.(a) The tree diagram of hierarchical clustering of those who were at home at the time of the earthquake.(b) The tree diagram of hierarchical clustering of those who were in the hospital or facility at the time of the earthquake.

•Figure 3 .
Figure 3. Results of dimensional compression using the t-SNE method and visualization of cluster results in a plan view.(a) Visualization of cluster results of those who were at home at the time of the earthquake.(b) Visualization of cluster results of those who were in the hospital or facility at the time of the earthquake.

Figure 4 .
Figure 4. Main factors revealed by cluster analysis and the classification algorithm graphs that used them.(a) Classification algorithm graph for those who were at home at the time of the earthquake.(b) Classification algorithm graph for those who were in the hospital or facility at the time of the earthquake.

Table 1 .
Summary of background data of the study population.

Table 3 .
Results of cluster analysis of those who were in hospital or facility at the time of the earthquake.*Fisher's test for categorical variables, one-way ANOVA otherwise.† Cramer's V for categorical variables, η 2 otherwise.Statistical analysis is performed only for groups with n ≥ 30.