Abstracts Submitted: 250
Number of Users: 295
View Abstracts Submitted
Back to home Page
Major Depressive Disorder, also known as clinical depression, is characterized by a prolonged feeling of sadness and worthlessness in an individual. Depression is a common precursor to suicide, unfortunately it often goes undetected and untreated. Timely detection of depression and an accurate measure of its severity is crucial for the patient and the family alike. To date, clinicians rely on subjective measures that are guided by their interaction with the patients. The diagnosis relies on interviews based on the measures like the Hamilton Rating Scale for Depression or self reporting tools like the Beck's Depression Inventory or the Patient Health Questionnaire. These measures are unreliable as they are biased and are guided by the patients' cooperation, which is often minimal. To be able to overcome these limitations, objective measures for depression detection have being explored by scientists over the past years to improve the performance of the depression detection system. To increase the accuracy and hence the dependability of the depression detection system, many multimodal techniques have been proposed that involve study of two or more modalities (videos, speech and text) together. Techniques based on the observation of facial expressions captured through video recordings of the patients, the key words spoken that are related to intensity of sadness and the speech characteristics are some of the modalities that are being investigated. Though the performance in terms of the ability to detect depression improves in the multimodal system, the complexity in acquiring the features as well as the complexity of the decision system are a deterrent in the development and the evaluation of a depression detection system. Speech is a noninvasive relevant measure which exhibits strong correlation with depression symptoms and also with suicidal tendencies. Noticeable acoustic changes are measured as a result of cognitive and physiological changes that mark depression. The tenseness of the vocal chords and the vocal tract leads to a monotone and hollow voice that marks the presence of the depressed state of mind. Speech based depression detection systems are leveraged by the low cost of acquiring data and the low complexity of the learning model. As opposed to the multimodal systems, an unimodal system based on speech alone is a strong contender for depression detection. In literature, various features from the speech are extracted and investigated and their relevance judged for depression detection. But, presence of irrelevant features inhibits the efficiency of the learning system. It is imperative to reduce the dimensionality of the data using appropriate techniques of feature selection, which retain the most relevant features of speech that are correlated with depressed speech. Furthermore, the optimum combination of feature subset selection method and the classifier is essential for a robust speech based depression detection system. In this paper, we have explored the potential of the acoustic features, with the aid of feature selection and classification methods. Three well known univariate filter methods have been investigated and four classifiers have been used to develop the decision models. The univariate filter methods investigated are: the Pearson Correlation method, the Mutual Information method and the Fisher Discriminant Ratio. The classifiers used for building the decision model are: the k-Nearest Neighbour, the Linear Discriminant, the Tree based classifier and the Support Vector machine. The Experiments have been performed on both Gender Independent and Gender dependent data of the DAICWOZ dataset. Feature selection techniques to select a subset of relevant set of features have given high performance in terms of low error rates and high f1-score of depression in comparison to the related work in this area. Based on the results, we propose the use of the acoustic features for depression detection as the modality, which is easy to acquire and low on complexity, and at the same time gives high performance in terms of error, f1-score of depressed speech and the number of features involved.
© Copyright 2017 All Rights Reserved