A hybrid model for depression detection using deep learning

Abstract

Millions of people are suffering from mental illness due to unavailability of early treatment and services for depression detection. It is the major reason for anxiety disorder, bipolar disorder, sleeping disorder, depression and sometimes it may lead to self-harm and suicide. Thus, it is a very challenging task to recognize people who are suffering from mental health disorders and provide them treatments as early as possible. Conventionally, depression detection was done through patients' interviews and PHQ scores, accuracy of conventional methods is very less. In this work, a hybrid model is proposed for depression detection using deep learning algorithms, which mainly combines textual features and audio features of patient's responses. To study behavioral characteristics of depressed patient's, DAIC-WoZ database is used. Proposed method consists of three components; first, a textual CNN model in which a CNN model is trained with only text features, second, an audio CNN model in which CNN model is trained with only audio features and third, a combination of audio and textual model named as hybrid model in which LSTM algorithms are applied. An improved version of LSTM model named as Bi-LSTM model is also used in the proposed work. In results, training accuracy, training loss, validation accuracy and validation loss is calculated for all the mentioned models. The results shows that deep learning is a better solution for depression detection in which accuracy of textual CNN model is 92% whereas accuracy of audio CNN model is 98% and loss of textual CNN is 0.2 whereas loss of audio CNN is 0.1. These results show that audio CNN is a good model for depression detection. It performs better as compared to textual CNN model. It is also observed that Bi-LSTM has better learning rate as compared to other models with accuracy 88% and validation accuracy 78%. There are some parameters such as precision, F1-score, recall and support are found for evaluation of models. In results, graphs for training loss, validation loss, training accuracy and validation accuracy are plotted. At last, by using confusion matrix depression can be detected for textual CNN Model, audio CNN model, LSTM model and Bi-LSTM against true label and predicted label.

Keywords

• Deep learning
• Convolution neural network
• Textual CNN
• Audio CNN
• LSTM
• Bi-LSTM