Seminar "AI Models to Diagnose Depression Using Acoustic Features"

On June 26, a scientific seminar on the topic “AI Models to Diagnose Depression Using Acoustic Features” was held at the Laboratory of Artificial Intelligence for Cognitive Sciences

The speaker was Alexandra Kovaleva (masterr's student at the HSE) under the supervision of Shalileh Soroosh (Ph.D. in Computer Science, Laboratory Head).

Annotation
Depression is one of the most widespread mental issues of the world today that affects an individual’s quality of life to a considerable extent. A lot of people tend to practice self-diagnosis avoiding doctor's consultation and try to heal themselves on their own because appointment in the hospital takes a considerable amount of time and disturbs individual's privacy. In this study we examined various Artificial Intelligence (AI) methods to detect whether a person is suffering from depression or not, using the acoustic features (such as pitch, tone, rhythm, etc.) extracted from the voices.

Assuming that the acoustic features are promising indicators of depression. We took the dataset of 346 patients from Mental Health Research Center in Moscow, RF., who were asked to record their voices while completing one of the tasks: picture description, reading IKEA instruction and telling thie personal story. To assess severity of depression doctors used two scales: Hamilton Depression Rating Scale (HDRS) and Quick Inventory of Depressive Symptomatology (QIDS) scales.  We extracted features from the audio recordings of patients and trained several different models: ranging from conventional Machine Learning (ML) models, such as ensemble learning algorithms and k-nearest neighbor to more advanced deep learning architectures, such as TabNet and Wide&Deep methods. The results of our study show that several models can achieve high accuracy in predicting depression levels with approximately 0.62 and 0.7 ROC-AUC and F1-Score respectively, using picture descriptions as a stimulus for patients. In addition, out of two scales, QIDS showed the most accurate results in terms of prediction.  

Overall, our results demonstrated that deep learning models have great potential for depression detection using extracted acoustic features; however, further research is required to improve the quality of the obtained results.

Materials:  AI Models to Diagnose Depression Using Acoustic Features.pdf