Combining scales to assess suicide risk

Abstract

Objectives

A major interest in the assessment of suicide risk is to develop an accurate instrument, which could be easily adopted by clinicians. This article aims at identifying the most discriminative items from a collection of scales usually employed in the assessment of suicidal behavior.

Methods

The answers to the Barrat Impulsiveness Scale, International Personality Disorder Evaluation Screening Questionnaire, Brown–Goodwin Lifetime History of Aggression, and Holmes & Rahe Social Readjustment Rating Scale provided by a group of 687 subjects (249 suicide attempters, 81 non-suicidal psychiatric inpatients, and 357 healthy controls) were used by the Lars-en algorithm to select the most discriminative items.

Results

We achieved an average accuracy of 86.4%, a specificity of 89.6%, and a sensitivity of 80.8% in classifying suicide attempters using 27 out of the 154 items from the original scales.

Conclusions

The 27 items reported here should be considered a preliminary step in the development of a new scale evaluating suicidal risk in settings where time is scarce.

Introduction

Suicide is a major health issue. One suicide is completed every 40 seconds, leading to approximately one million deaths every year worldwide (WHO, 2002). Moreover, suicide is the third most important cause of death worldwide among people aged 15–44 (Holmes et al., 2007). Notwithstanding human costs, the economic burden of suicidal behavior has been estimated annually in $33 billion in the United States (Coreil et al., 2001). Fortunately, suicidal behavior might be prevented to a great extent (Jamison, 2000). Treating subjects at risk with the appropriate preventive measures, such as cognitive behavior therapies (Brown et al., 2005) can reduce suicide rates up to 25% (Isaacson, 2000). More recently, a 75% reduction of suicide rates has been reported in a large depression care program (Hampton, 2010). In order to detect subjects at risk, researchers have investigated the factors underlying suicidal behavior. The most relevant risk factors are major depression (Mann et al., 1999b), high impulsiveness (Patton et al., 1995), aggressiveness (Mann et al., 1999b), personality disorders (Mann et al., 1999a), life events (Kolves et al., 2006), and social-demographic factors (Smith et al., 1988).

Unfortunately, most of these studies did not measure the effectiveness of the risk factors to identify subjects at risk. They just tested if there was a statistically significant relationship between the studied variable (e.g. high impulsiveness) and suicidal behavior. Therefore, the clinical usefulness of these studies is limited. One notable exception is the seminal Pokorny's article (Pokorny, 1983). Pokorny applied discriminant analysis to several features including, among others, socio-demographic variables, the 24 items of the brief psychiatric rating scale, and the items included in the nurses' observation scale for inpatient evaluation. Although it was an innovative approach, his results showed a weak performance, with accuracy, sensitivity and specificity levels below 70%. More recently, Hendin (Hendin et al., 2010) slightly improved these results achieving an accuracy of 71.67% with a specificity of 74% and a sensitivity of 60%. The improvement was basically due to the use of a different set of predictive variables, as they used a simple classifier consisting on the sum of the individual scores associated to each variable. This study suggested that the use of a more suitable set of predictive variables together with the use of more sophisticated classifiers might improve the classification accuracy of people at risk of suicidal behavior. This intuition was confirmed by us (Delgado-Gomez et al., 2011) in a study aimed at discriminating between suicide attempters (SA) and non-SA. In this study, we used two personality scales as predictors, and a collection of modern classification techniques such as linear discriminant analysis, Fisher linear discriminant analysis, boosting, and support vector machines (SVM). The best results were obtained with SVM, which achieved a classification accuracy of 80.3%, with a specificity of 86.8% and a sensitivity of 76.1%. Recently, Stefansson et al. (Stefansson et al., 2010) have shown that the prediction of suicide can be improved by means of an appropriate selection of the items. However, their results were obtained ad hoc. Therefore, we have applied the Lars-en algorithm in order to automatically select the most discriminative items (Delgado-Gomez et al., 2012). Using the selected items of scales measuring life events and personality disorders, it was possible to achieve a classification accuracy of 83%. A question that remains open is the accuracy that could be obtained if the Lars-en algorithm were applied to a set of items assessing the most relevant risk factors for suicidal behavior.

The present study extends our previous findings and is conceived as a further step toward the development of more precise and reliable measures of suicide risk (García-Nieto et al., 2012). The major goal is to maximize the classification accuracy applying simultaneously Lars-en to sociodemographic factors and items from four scales measuring some of the most relevant risk factors for suicidal behavior (impulsive aggression, life events, and personality disorders). As a by-product, we developed a small set of items to classify subjects as SA or non-SA. This set of items might help to develop a tool to support clinical decisions with regard to suicide risk in settings where time is scarce.