Global Suicide Mortality Rates (2000-2019): Clustering, Themes, and Causes Analyzed through Machine Learning and Bibliographic Data

Abstract

Suicide research is directed at understanding social, economic, and biological causes of suicide thoughts and behaviors. (1) Background: Worldwide, certain countries have high suicide mortality rates (SMRs) compared to others. Age-standardized suicide mortality rates (SMRs) published by the World Health Organization (WHO) plus numerous bibliographic records of the Web of Science (WoS) database provide resources to understand these disparities between countries and regions. (2) Methods: Hierarchical clustering was applied to age-standardized suicide mortality rates per 100,000 population from 2000-2019. Keywords of country-specific suicide-related publications collected from WoS were analyzed by network and association rule mining. Keyword embedding was carried out using a recurrent neural network. (3) Results: Countries with similar SMR trends formed naturally distinct groups of high, medium, and low suicide mortality rates. Major themes in suicide research worldwide are depression, mental disorders, youth suicide, euthanasia, hopelessness, loneliness, unemployment, and drugs. Prominent themes differentiating countries and regions include: alcohol in post-Soviet countries; HIV/AIDS in Sub-Saharan Africa, war veterans and PTSD in the Middle East, students in East Asia, and many others. (4) Conclusion: Countries naturally group into high, medium, and low SMR categories characterized by different keyword-informed themes. The compiled dataset and presented methodology enable enrichment of analytical results by bibliographic data where observed results are difficult to interpret.

Keywords: age-adjusted suicide mortality rate; association rule mining; bibliographic analysis; keyword clustering; machine learning; network analysis; recurrent neural network; suicide; text mining; word embedding.

Figure 1

Age-standardized suicide mortality rate (SMR) in selected countries from 2000–2019 showing similarity of SMR levels and trends in selected countries: Estonia is similar to Hungary, Lithuania is similar to Russian Federation, Brazil is similar to Mexico, and Lebanon is similar to Kuwait. The lower y-axis ticks represent the 10th (3.4), the 30th (6.6), the 50th (9.9) (median), the 80th (16.6), and the 90th (22.8) percentiles of the SMR across all countries. Some countries, such as Lesotho and Eswatini, have an SMR in most of the years from 2009–2019 which is considered high compared to other countries. On the other hand, the countries in which SMR in most of the years from 2000–2019 was below the value of the 10th or 20th percentile can be considered low-SMR countries compared to the other countries.

Figure 2

Histograms approximated by normal distribution of age-standardized median SMR values in countries from 2000–2019 in each cluster. Clusters with one or two elements are omitted. Legend shows parameters of the approximation by normal distribution—values of the centroids of median SMR of the countries grouped into each cluster and a standard deviation.

Figure 3

Overall average trends of suicide mortality rate dynamics (with confidence intervals) in each cluster. Cluster C4 is Eswatini, cluster C5 is Lesotho and cluster C6 contains Kiribati and Guyana. Other clusters have at least 5 countries.

Figure 4

Trends of SMR per 100,000 population of both genders from 2000–2019 in high-SMR country clusters C6, C7, C9, C10.

Figure 5

Dynamics of suicide mortality rate per 100,000 population from 2000–2019 for both genders of countries in cluster C8 that had a mix of high and medium SMR.

Figure 6

Bar plot of median suicide mortality rate per 100,000 population of years 2000–2019 for both genders marked by economy for medium-SMR countries in clusters C2 and C3.

Figure 7

SMR trends in selected countries from clusters C2 and C3.

Figure 8

Trends of suicide mortality rate per 100,000 population in years 2000–2019 for both genders in low-SMR countries from cluster C1.

Figure 9

Countries that have increasing SMR trend from 2000–2019. Uruguay comes from a high-SMR cluster. United States is from a medium-SMR cluster. Other countries are from a low-SMR cluster.

Figure 10

Network representation of the top 15 most frequent keywords and phrases (item sets) resulting from association rule mining algorithm applied on keywords of articles associated with countries in each geographical region. Highlighted are the most frequent phrases common between the regions. Depression emerges as the central theme in all regions.

Figure 11

Suicide vocabulary mapped in Keras embedding projector tool in 3D UMAP projection in which the string “suicide” is highlighted (685 matches).

Figure 12

Closest neighbors of “Afghanistan” keyword are “Afghanistan war veterans”, “Afghanistan veterans”, “veterans’ suicide”.