College of Computing and Technology, Lipscomb University, 1 University Park Dr., Nashville, TN, USAUnited States
College of Computing and Technology, Lipscomb University, 1 University Park Dr., Nashville, TN, USA; Vanderbilt University Medical Center, Nashville, TN, USAUnited States
College of Computing and Technology, Lipscomb University, 1 University Park Dr., Nashville, TN, USA; Office of Research, Middle Tennessee State University, 1301 E., Main St., Murfreesboro, TN, USAUnited States
Also available as PDF (figures and tables appear in the PDF).
The goal of this research is to identify and promote awareness of prominent demographic risk factors to predict individuals at risk for suicide and aid in the prevention within the USA. This would support the Education Development Center’s Zero Suicide initiative and provide strategies and tools to health and behavioral health systems to reduce suicide mortality. The research presented in this paper focuses on the hypothesis that demographic variables available in the Center for Disease and Prevention (CDC) mortality data sets should be integrated into initiatives to identify and prevent suicide mortality. A comprehensive analysis of the CDC mortality data from 2003 through 2013 was extracted, transformed, loaded and analyzed utilizing Python, R Scripting, RStudio and Tableau. The CDC mortality data was subsided into a data frame of 17 variables from the original 75 variables that indicated the most statistical significance as a function of the respective suicide ICD10 codes. Education attainment levels of a 12th grade education emerged as one of the most statistically significant variables that contributed to suicidal deaths; this observation is consistent with initial observations of the 2013 CDC mortality data analyzed in our previous studies.1 Based on this unique finding of education emerging as a strong and consistent variable in the comprehensive analysis of CDC data over an 11 year period, the authors hypothesize education attainment level segments are the most significant demographic predictor variable of suicide and a systematic approach to targeting continuing educational opportunities to patients with low education attainment levels paired with other high risk segments including but not limited to age, race, ethnicity, marital status and gender.
Keywords: Center for Disease and Prevention (CDC); Data Science; Demographics; Education; Suicide; Suicide risk factors; Mortality data; Educa- tion
Developing a zero-suicide culture in a healthcare system requires the application of data driven quality improvement to identify at risk patients and measure the outcomes of preventative care to eliminate suicide as a leading cause of death in the United States. Variability in patient behavior and health care provider observations and assessments lead to subjective measures of risk factors that create data quality issues impeding the ability of healthcare systems to standardize data collection to discern meaningful information from their data.
Problem Space • Suicide rates continue to rise in the United States • Suicide ideation, attempt and mortality data quality reporting integrity • Systematic adoption of suicide risk identification and prevention protocols • Significant monetary impact to the United States Rising Suicide Rates A data driven quality improvement to identify suicidal patients and implement preventative measures has been difficult by healthcare systems. Despite the efforts of the organizations to prevent suicide, the CDC data revealed 400,349 suicide related deaths, marking a 2.8% increase per year 2003 to 2013. Amongst the reported deaths, the epidemic affects white, middle-aged males with a 12th grade education attainment level the most out of the United States population. The subset of 17 variables from the CDC revealed the following notable information (Table 1).
Attainment of beneficial and accurate suicide ideation, attempt and mortality data is a barrier to healthcare systems to truly capture the scope of the suicide epidemic in the United States.
Prevention Program that outlines some of the barriers impacting research, identification and prevention efforts. These are2: • Provide an incomplete picture of the problem of suicidal behavior: Most suicide attempts do not result in death and are not included in mortality data • Despite better reporting than morbidity data, not all suicides are reported: Sometimes there is not enough information to determine intent. Without conclusive evidence, potential suicides may be recorded as unintentional or undetermined on death certificates • Less completely reported: While psychologically serious, many suicide attempts are not medically serious enough to require medical attention and do not get reported/coded • Captures a biased view of the suicide injury problem: Hospital datasets are more accessible for public health surveillance than data from private physicians, clinics, and health main tenance organizations. However, hospital data may under- or over-represent certain sub-groups • More difficult to accurately collect data about the way people feel or think versus how they behave • Subject to reporting biases. For example, high school students are asked on the youth risk behavior survey if they ever seriously considered suicide. This question is subject to recall bias (not all people will remember), social desirability bias (not all will want to admit suicidal feelings, even on an anonymous survey), and to definition issues. After all, what is meant by “seriously” considered suicide?
Suicide is a sensitive subject for many people suffering from a mental illness, a traumatic event, friends and families with suicidal loved ones, school systems and even healthcare providers. The solution to treating the United States suicide epidemic will not be a simple solution. However, research suggests that a zero-suicide culture adopted from a systematic approach can reduce and even eliminate suicide related deaths, as proven with Henry Ford Health System’s award winning Depression Care Program and Centerstone’s Crisis Care Services program.3 Positive results have been identified in these two healthcare facili ties; however, the United States healthcare system is extremely large, highly regulated and financially burdened. The implementation of programs like the EDC’s Zero Suicide initiative would require a leadership-driven culture change with how suicide prevention training is conducted within healthcare systems. 4 Nonhealthcare organizations and systems may not be able to take the same approach; however, as the United States begins to act to remediate the closeted conversation of mental illness and suicide, additional systematic approaches can be developed for a greater national impact.
Suicide related deaths take an enormous toll on society in the United States. Aside from the emotional ramifications, the CDC reported in 2010 that suicide related death cost $44.6 billion USD for ages 10 and older. The seemingly large figure is calculated upon the combination of medical and work lost cost. Each suicide related death is reported to impact society at a rate of $1,164,499 USD. The application of this figure to the 400,349 suicidal related deaths from 2003 to 2013 amounts to a cumulative cost of $466,206,010,151.
Table 1. Data Subset Summary Statistics.
Suicide deaths by gender Male 315,175 Female 85,174 Suicide deaths by race (Top four of fifteen)
American Indian 4,720 Suicide deaths by education attainment level High School Diploma 70,195 Some college, Bachelor and Advanced Degrees 57,198 Page 3 MOTIVATION The CDC recognizes suicide is one of leading cause of death worldwide, and the United States is not an exception, account ing for 42,773 deaths in just 2013. As of 2015, suicide was the 10th highest causes of death. For years, healthcare professionals have been fighting relentlessly regarding this issue. According to Dinah Miller,5 of Psychology today suicide rates are increas ing every year. In Simon’s 6 article Suicide risk assessment: Is clinical experience enough? He states, “Accurate and defensible risk assessment requires a clinician to integrate a clinical judgment with the latest evidence-based practice, although accurate prediction of low base rate events, such as suicide, is inherently difficult and prone to false positives.”
According to contributors to the Assessment of Suicide Risk,7 effective suicide risk assessment, “...should distinguish between acute and chronic risk. Acute risk might be raised because of recent changes in the person’ s circumstances or mental state, while chronic risk is determined by a diagnosis of a mental illness, and social and demographic factors. Suicide risk assessments are currently conducted with the following assessments:” • The Scale for Suicide Ideation (SSI) • The Modified Scale for Suicide Ideation (MSSI) • The Suicide Intent Scale (SIS) • The Suicidal Affect Behavior Cognition Scale (SABCS) • The Suicide Behaviors Questionnaire (SBQ) • The Life Orientation Inventory (LOI) • The Reasons for Living Inventory (RFL)
suicide risk noted that, “There are risks and disadvantages to both overestimation and under-estimation of suicide risk. Oversensitivity to risk can have undesirable consequences, including inappropriate deprivation of patients’ rights and squandering of scarce clinical resources. On the other hand, underestimating sociality because of a dismissive attitude or lack of clinical skill jeopardizes patient safety and risks clinician liability.” As suicide rates continue to rise, there is reason to believe the assessments that are currently in use are not comprehensively capturing the motivations and severity of being able to properly classify the suicide risk level of patients.
The purpose of this paper is to explore social and demographic factors that are currently not being utilized in assessments that could help identify the risk of suicide in people. By uncovering these indicators, we speculate insights can be gained to aid in the improvement of suicide risk assessments being utilized by healthcare professionals to positively impact the efficacy of the efforts to decrease suicide rates in the United States. If the correlation is discovered between suicide and previously ignored risk factors, actionable programs could be developed and targeted for high risk groups. These programs could be designed to help patients seeking clinical care and, as well as those at high-risk who are not actively seeking clinical help.9
While the CDC has gathered an extraordinarily large data set on suicides in the United States, the variables have not revealed clear motivators as to why suicides rates continue to rise. Per Miller,5 The CDC reports revealed the following: “The news from the Centers for Disease Control shows a striking increase in suicide rates. Among those ages 35 to 64 years old (the baby boomers), there is a 28% increase in suicide rates from 1999 to 2010. It holds for males (up 27%), females (up 31%), and across different regions of the country. The peaks were seen in men in their 50’ s and women in their early 60’ s. The gender difference continues to show that men die of suicide at three times the rate of women, and suicide is now the 4th cause of death for that age group. More people die of suicide than car accidents. The rise is most striking in non-hispanic whites and native american alaskan indians, groups that have always had the highest rates. The suicide rate is now 17 per 100,000, up from 13 per 100,000. And while we worry more about homicide, suicide rates are twice the homicide rates. Marriage is protec tive, as is a college education, and in fact the suicide rate in college-educated women went down.”10-12 Further, according to Milner et al13 in a study published in the British Journal of Psychology, not all variables are known that are strong indicators of suicide.
“This study confirms that certain occupational groups are at elevated risk of suicide compared with the general employed population, or compared with other occupational groups. At greatest risk were laborers, cleaners and elementary occupations (ISCO major category 9), followed by machine operators and ship’ s deck crew (ISCO major group 8). The greater risk of suicide in lower skilled occupational groups may be symptomatic of wider social and economic disadvantages, including lower education, income and access to health services.”
In addition to the non-monetary impacts of suicide, according to the CDC’s,14 there are significant financial impacts to society: • Suicide costs society over $44.6 billion a year in combined medical and work loss costs • The average suicide costs $1,164,499 The majority of the individual suicide cost is a result of the work-loss amounting in $1,160,655 and the total amount increases once the average medical cost of $3,646 is calculated. 15 The monetary impact to society was explored in Suicide and Suicidal Attempts in the United States: costs and Policy Impli cations. Sheppard et al calculated the cost to society in millions Table 2: Suicide Breakdown Cost.
Medical cost Suicides $121 $26 $146 0.3 Nonfatal suicide attempts $1,149 $388 $1,537 2.6 Total (all self-inflicted injuries) $1,270 $413 $1,684 2.9 Indirect economic cost Suicides $43,589 $9,458 $53,047 90.8 Nonfatal suicide sttempts $3,196 $518 $3,714 6.4 Total (all self-inflicted injuries) $46,785 $9,976 $56,761 97.1 Total economic cost Suicides $43,710 $9,483 $53,193 91.0 Nonfatal suicide attempts $4,346 $906 $5,251 9.0 Total (all self-inflicted injuries) $48,056 $10,392 $58,445 100 Source: Author’s calculation.
ªItems may not sum to totals due to rounding.
Table 3. Analysis of CDC 2013 Death Data where Education Level was Made Available.
by component and age range in the following table 2 and 3.16 The analysis of the CDC suicide data performed by Brandt et al1 made a case that a sample data set which only uses records where the education level of the deceased is known can be used to identify trends. The data matched national statistics around sex, marital status (Table 3) and location leading to a conclusion that these records would also reflect the education levels of the larger population.
This paper is organized as follows. Section 1 is the introduc tion of the paper, containing the problem space, motivation and related work. Section 2 describes the materials and data preparation involved in analyzing and drawing meaningful conclusions. Section 3 is a discussion of the results of the various data science techniques. Section 4 discusses potential challenges. Section 5 presents the possible future work for this project. Section 6 offers a conclusion based on the results. Section 7 lists the references used in the research paper. Finally, section 8 contains tables, code, and charts of interest.
The data was collected from The Centers for Disease Control and Prevention (CDC), http://www.cdc.gov/nchs/data_access/ vitalstatsonline.htm#Mortality_Multiple. The files contain all the death records of all known deceased individuals from 2003- 2013 across the United States. The data is gathered through the CDC’s National Vital Statistics Systems, with the exception of the ICD10 Codes that are sourced from the World Health Or ganization (WHO).17 The data files report on the 75 applicable variables for each record in respect to the deceased individual’s Page 4
known demographic data, reported death indicators and ICD10 Codes.
Statistical Data Analysis • Step 1: The 2003-2013 death data file was downloaded in a DUSMCPUB format • Step 2: The file required to be parsed and converted into a CSV format to be imported into RStudio. To accomplish this task, a modified python script (Appendix B) available on Git Hub, was utilized to parse the DUSMCPUB data into a CSV file.
• Step 3: After parsing file and converting into a CSV file, the data was imported into RStudio.
• Step 4: As suicide risk could be correlated to objective demographic variables, a subset of 18 variables from the original 75 variables were identified to create a subset for statistical analysis of unknown variables to suicide risk. The variables identified as potentially significant are listed in Appendix A. • Step 5: To aid in correlation analysis and, ultimately, a linear regression model of the most significantly correlated variables, variables were assigned factor levels, except for age, as described in the Rscript in Appendix B.
• Step 6: The original data contained all causes of death in the manner of death variable; however, the research is focused only on variables correlated with the suicide value. As described, in the Rscript in Appendix A subset of data was created to contain only suicide related death.
• Step 7: According to aforementioned studies, the Education variable was likely to be highly significant to suicide risk. As described in the Rscript in Appendix C, the records without Education data were removed to clear the data set.
• Step 8: After removing the records without a reported Education level and NA’s, the subset of data contained 158,970 records.
• Step 9: A combination of RStudio and a Tableau were used to visualize the data for analysis.
As can be seen in the Table 4 below, the original death data file contained 27,224,858 rows and 75 variables (i.e., columns). This data was reduced to contain only death labeled as suicided, bringing the record count to 400,349. Some of these rows of data were missing education levels; these records were removed, leaving a sample size of 158,970 and 17 variables that represent demographic data that could be valuable to the research.
Potential data science approaches being explored are clustering, regression and hypothesis testing to identify any significance the 18 selected variables will have to predict the probability a subject would commit suicide. Cluster analysis or clustering is the task of grouping a set of objects in such a way that objects in the same group (called a cluster) are more similar (in some sense or another) to each other than to those in other groups (clusters). It is the main task of exploratory data mining, and a common technique for statistical data analysis, used in many fields, including machine learning, pattern recognition, image analysis, information retrieval, bioinformatics, data compression, and computer graphics. 18 Regression is defined as a technique in which a straight line is fitted to a set of data points to measure the effect of a single independent variable. The slope of the line is the measured impact of that variable. 19 Hypothesis testing is the use of statistics to determine the probability that a given hypothesis is true. The usual process of hypothesis testing consists of four steps namely null hypothesis, test statistic, p value and comparison of p value to an acceptable significance value called α-value to see if the effect is statistically significant, then the null hypothesis is ruled out, and the alternative hypothesis is valid.20
The rationale for applying a data science approach to the 2003-2013 CDC death dataset is the success achieved in genome sequencing using data science. In the Center for Disease Control and Prevention blog, Khoury states, “Genome sequencing of humans and other organisms has been a leading contributor to Big Data, but other types of data are increasingly larger, more diverse, and more complex, exceeding the abilities of currently used approaches to store, manage, share, analyze, and interpret it effectively. We have all heard claims that Big Data will revolutionize everything, including health and healthcare.”21 By discovering associations and understanding patterns and trends Table 4: Observation and Variable Counts.
Original death data set 27,224,858 75 Original by suicide 400,349 18 Suicide by education 158,970 18 Page 5 within the data, big data analytics has the potential to improve care, save lives and lower societal impact.
The following graphs show the findings from the analysis of the 11-year period 2003-2013 (Figure 1, 2 and 3).
Rationale for using Data Set The CDC’s22 defines suicide as, “Death caused by self-directed injurious behavior with intent to die as a result of the behavior.” It is believed the appropriate data to find unknown suicide risk indicators from the large number variables available and data published by the CDC. In addition to the large number of variables and records, their data is reliable. According to the CDC’s,7 their suicide data is gathered through the resources:
• National Inpatient Sample (NIS) • National Violent Death Reporting System • The National Vital Statistics System • WISQARS • Youth Risk Behavior Surveillance System (YRBSS)
Figure 2. This Graph Supports the Idea that Education Level could be a Key Risk Factor in Suicide Prevention.
Figure 1. Suicide Continues to Rise Year Over Year.
Figure 3. Suicide Ratio of Male Versus Females is 4 to 1 but the Average Age is Very Similar for both.
• National Survey on Drug Use and Health (NSDUH) Non-Federal Data Sources • Pan American Health Association, Regional Core Health Data Initiative • The American Association of Sociology • WHO Statistical Information System (WHOSIS)
There has been significant research conducted in the rise of suicide in the United States; however, the published research has been unable to produce a solution to the rise of suicide. Within the scope of this research, the potential challenges have been identified by the research team: • The majority of the research team is new to data science and is considered non-experts in the domain. The ability to find and use the correct data science method could prove difficult • The data set made available only accounts for the deceased and does not include any living patients. Additional data sets may be needed for a conclusive study, which may be protected information • Managing false positives and mis-classifying someone with a high risk or low risk of committing suicide • If an unknown variable is identified, how would information is provided to the proper people in a timely manner to help with prevention?
In the course of this research, the authors met with several individuals to gather a better understanding of the current work being done in the field of suicide prevention as well as insight to where the discovery could be used as a mechanism for change. Scott Ridgeway, Executive Director of the Tennessee Suicide Prevention Network, highlighted the need for a more rigor around how the data is collected not only across counties in Tennessee but across the nation, meeting notes in Appendix D. Jennifer Lockman, Program Evaluator and PhD candidate from Centerstone Research Institute, conveyed the fact that there needs to be more analysis of the data and echoed the need for more standardiza tion of data collection, meeting notes in Appendix E. The data presented at the Tennessee Suicide Prevention Network Advisory Council Retreat reinforced what the CDC data is showing, suicide rates are increasing, meeting notes in Appendix F. These meetings showed the need for more work in Tennessee and across the nation. The next steps include getting data specific to Tennessee suicides and comparing to what is being found on the national level. Concluding if education level is a factor that can be used in prevention and if so, help support initiatives that ensure the correct data is being collected.
Using the added data, information from experts, and the updated analysis publishing again to have the highest impact is a consideration. According to Journal Selector, the top 3 journals that would have the highest impact using the abstract of this paper are: Pediatrics, American Journal of Preventive Medicine, and American Journal of Public Health.
Due to the increasing rise of suicide in the United States, research was initiated with a data science approach to identify previously unknown indicators that could lead to the prevention of suicide in the US. Previous research has been conducted to determine indicators of suicidal deaths; however, the research was based on subjective analysis of a suicidal subject’s likelihood to commit suicide. This research sought to focus on indicators that were objective characteristics so the risk assessments conducted on suspected suicidal patients could potentially increase the accuracy of the risk assessment study. The studies reviewed prior to forming the research question did not utilize data science approaches to reach their conclusions that lower education levels and labor intensive occupations lead to a higher suicidal risk. It is believed that a linear regression model can be formed to fit the variables identified in the Center for Disease Control’s death datasets from 2003-2013 that are the most significantly correlated with reported suicidal death. If the model proves accurate, subjects of the populations fitting the criteria of high risk characteristics could be introduced to potentially life-saving preventative actions to reduce the probability the subject’s cause of death would be suicide.
The Authors acknowledge the CDC and the WHO for the making the data available to public and for analysis, to Scott Ridgeway for taking time to meet with the authors to discuss about TN state suicide rate and Jennifer Lockman for her insights on how data is currently being collected and suggestions of where change is needed.
The authors have no conflicts of interest.