Real-time Acoustic based Depression Detection using Machine Learning Techniques

Depression disorder is predicted to rise to the second leading cause of disability by 2030 as per the identifications of the World Health organization (WHO). Though well trained clinicians, medical and psychological treatments are available for depression treatment, persons or families are reluctant to speak out/reach doctors about this disorder for various social reasons. Diagnosis of depression disorder includes numerous interviews with patient and family, clinical analysis, questionnaires which is time consuming and also demands well trained clinicians. In the present era of Machine learning, automation of depression detection is not complicated and can easily be deployed. However, the automation should use fewer resources, provide accurate results with more reachability. In this paper, acoustic features are used to train a classification model to categorize a human as Depressed or not-Depressed. DIAC-WOZ database available with AVEC2016 challenge is considered for training the classifiers. Prosodic, Spectral and Voice control features are extracted using the COVAREP toolbox and are feature fused. SMOTE analysis is used for overcoming the class imbalance and 93% accuracy is obtained with the SVM algorithm resulting in Depression Classification Model (DCM). An android application cureD Deployed on Cloud is developed to self assess depression using DCM and PHQ-8 questionnaire. The application is tested on real time data of 50 subjects under the supervision of a qualified psychiatrist and an accuracy of 90% is obtained.