Automatic Suicidal Risk Assessment based on Clinician-Patient Verbal and Nonverbal Behavior Master Thesis

Suicide is the third leading cause of death in the United States of America for ages 13 to 18 and in some countries youth suicide is even the second leading cause of death for ages 10-24 according to Centers for Disease Control and Prevention and WHO. Hence, there is a need for a valid, rapid and accurate way to identify suicidal risk; it is especially important to be able to predict potential repeat occurrences of attempted suicides. In this thesis the conversation dynamics, verbal information and acoustic features of 30 suicidal and 30 nonsuicidal patients of ages 13 to 18 interviewed at the Cincinnati Children's Hospital Medical Center Emergency Department and the single clinician were statistically evaluated. Beside the identification of non-suicidal vs. suicidal patients, data of 17 suicidal non-repeaters and 13 repeaters was investigated. A hierarchical classification was used to confirm the discriminative faculty of the features to characterize the psychological states of the adolescents. Ensemble classifiers and SVM classifiers were used and tested with data from entire interviews and with data from the first five open-ended ubiquitous questions and answers. I was able to show that the analysis of verbal and nonverbal interactions between patient and clinician can identify the suicidal risk of adolescents. The investigation of the first five questions of the interview yielded similar promising results in comparison to the information observed in the entire interview. The backchannel of the clinician which lasted less than 700ms was useful to identify the suicidal risk including the patients' features. The hierarchical SVM-ensemble classification yielded an accuracy of 78.3% using data from entire interviews. Thus, the suicidal risk assessment of non-suicidal adolescents, suicidal non-repeaters and suicidal repeaters using clinician's and patient's verbal and nonverbal behaviors yielded promising results.