PhD Scientific Days 2021

Budapest, 7-8 July 2021

MH_I_L: Mental Health Sciences I. Lectures

Examination of Acoustic Features in Depression, Developing an Automatic Decision System for Discriminating Speech Pathology

Bálint Hajduska-Dér1, Gábor Kiss2, Klára Vicsi2, Lajos Simon1
1 Department of Psychiatry and Psychotherapy, Semmelweis University, Budapest
2 Department of Telecommunications and Media Informatics, University of Technology and Economics, Budapest

Text of the abstract

Introduction: Major depression has an increasing prevalence throughout the world, causing a huge burden in the economy. With the increasing number of patients and high suicide rate in depression, there’s a need of easy and cheap diagnostical methods in medical practices, without overwhelming the health care systems. Speech can be a viable biomarker in the diagnostical process of depression. With computer science and machine learning, we can examine not only the segmental level of the acoustic features, but the suprasegmental phonemes as well. Machine learning can be a great tool in the automatization of the diagnostical process and can speed up the adequate therapy.
Aim: The purpose of our study is developing an automatic decision system based on speech, using machine learning, in order to estimate the severity of depression in patients.
Method: The severity of depression in patients included in our study, was assessed by Beck Depression Inventory II (BDI) and Hamilton Depression Scale (HAMD). Antipsychotic use was an exclusion criterion due to the extrapyramidal side effects, which can alter the speech. Speech recordings contained a read text called “The North Wind and the Sun”. The samples were segmented on phonema level and support vector regression was used in the automatic decision system.
Results: Currently we have 218 speech samples in our database, 175 speech samples with BDI scores only, 43 speech samples with HAMD and BDI scores too. With the use of HAMD in the training process of our automated decision system, the prediction and recognition of depression was improved by 17%, compared to the model trained only with BDI. The accuracy of the classification was improved by 10% (from 81% to 84%)
Conclusion: Our results show that speech analysis and the acoustic biomarkers can show great promises in the diagnostical process of depression. With the use of a more objective assessment like the HAMD, we can improve the accuracy of our automatic decision system. With this new and innovative diagnostical tool, the early and proper diagnosis will be more accessible in the general medical practices, more patients can get adequate treatment in time, which can lower the suicide rate as well.
Funding: Support was provided from the National Research, Development and Innovation Fund of Hungary, financed under the K_18 funding scheme (Project no. K1285968).

University and Doctoral School

Semmelweis University, Doctoral School of Mental Health Sciences