Text-Based Depression Detection from Clinical Interview Transcripts Using a Bidirectional Recurrent Neural Network

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
The threshold of talking to a mental health professional can be too high for several reasons, which include the stigma on mental health, limited accessibility of mental health professionals and a lack of self-awareness. As a result, many mental health patients remain unnoticed and untreated, leading to negative rami cations in the long term. Yet, how can we ensure that we could screen patients in a timely and non-intrusive manner to limit the negative consequences? To address this problem, a research pipeline (the detect-react pipeline) was created with the bachelor thesis of G.W.F.X. Peters Rit. This thesis focuses on the detect segment, aiming at quantifying the presence of depression using machine learning techniques. Peters Rit's thesis focuses on developing a chatbot that can respond appropriately to the quanti ed level of depression. A Bidirectional Long Short-Term Memory Recurrent Neural Network (BiLSTM RNN) was developed to quantify the participants' level of depression based on of clinical interview transcripts. Accuracy, precision, recall and F1-scores were achieved at 0.85, 0.87, 0.82 and 0.84 respectively. Even though the performance metrics were relatively high given the sparse dataset, the research proposes some critical issues in uencing the clinical deployment of the model. These issues concern the values of bias, explainability and safety of the network. Further research needs to be conducted to strengthen the scienti c foundation and reliability of answering the research question. Keywords { depression detection, deep learning, LSTM, recurrent neural networks, computational linguistics, value-sensitive design
Faculteit der Sociale Wetenschappen