| dc.description.abstract | Depression is a potentially life threatening mood disorder which affects many people. Two thirds of the people with depression don‟t realize that depression is a treatable illness, only 50% of people diagnosed with major depression receive any kind of treatment, and only 20% of them get treatment. Depression can lead to suicidal behavior. It is very important to realize that depression is a treatable disorder and suicide is a preventable act. A recent research study reported a frightening result which was that 48% of patients who have suicidal ideations and 24% of those who have committed suicide did not receive any care or even perceive the need for care. Therefore, it is very important to evaluate a patient‟s psychological state and to evaluate a depressed patient‟s risk of committing suicide, since suicide may be prevented by the psychiatric help. A unique challenge is discriminating the high risk suicidal (HR) patients from the depressed (DP) patients and this dissertation is focused on tackling this challenge.
In this dissertation, two different types of audio recordings from the depressed patients (diagnosed with depression), the high risk suicidal patients (diagnosed with high risk suicide), and the remitted patients (diagnosed with remission from the depression) were gathered and analyzed. One type is audio recordings that were gathered from the clinical interviews (interview session); the other one is gathered while the patients were reading a predetermined passage (reading session).
This dissertation presents three different studies. In the first study, mel-frequency cepstral coefficients (MFCCs) are used to estimate suicidal risk using different numbers of MFCCs with and without environmental compensation. A different approach is proposed to maximize the classification rates of discriminating the high risk suicidal patients from the depressed patients using fewer coefficients. The aim of this research is estimating the suicidal risk using MFCCs with high classification rates and the results show that the MFCCs are useful indicators for DP-HR discrimination.
In the second study, we propose various approaches to maximize the classification rates of discriminating the high risk suicidal patients from the depressed patients using power spectral density features. In earlier studies, 4 fixed energy bands which are uniformly placed band edges in the 0-2000 Hz frequency range (0-500 Hz, 500- 1000 Hz, 1000-1500 Hz, 1500-2000 Hz) were analyzed. In this study, various optimization techniques are used which are increasing the number of energy bands, increasing the energy band range, increasing the energy band number & range, exponential band edges, exponential band edges & increasing the energy band range, non-uniform band edges, and finally non-uniform band edges & increasing the energy band range. It is found that these approaches provide better classification rates for discriminating the high risk suicidal patients from the depressed patients.
In the last study, gender specific differences on optimized energy bands are investigated. There exist statistically significant gender differences in the Depressed (DP) and the High Risk Suicidal (HR) pairwise group during the interview and reading sessions. 14 statistically significant features are found during the interview session, and 4 statistically significant features are found during the reading session. There are no statistically significant gender differences in the High Risk Suicidal-Remitted (HR-RM) pairwise group during the interview session and during the reading session. There exist statistically significant gender differences in the Depressed (DP) and the Remitted (RM) pairwise group during the interview and reading sessions. 26 statistically significant features are found during the interview session, and 2 statistically significant features are found during the reading session. Spontaneous speech (interview session) is more effective for revealing gender differences than the controlled reading speech (reading session). | |
