Algorithm Classifying Depressed Patients’ Speech Samples Using Deep Learning

This research uses deep learning techniques to create a highly accurate model for identifying audio recordings, with an emphasis on female speakers and their mental health difficulties. Based on gender, the audio recordings are categorized into three categories: Remitted, Depressed and High-risk for suicide. We studied and compared different deep learning algorithms in depth. Neural networks include 1D and 2D Convolutional Neural Networks (CNNs), Support Vector Machines (SVMs), and Recurrent Neural Networks with Long Short-Term Memory (RNN-LSTM). Specifically, between Remitted vs. Depressed and Remitted vs. High risk for suicide, utilizing 45-second and 2 -minute speech segmented data, respectively. In this category, the CNN1D model achieved an accuracy of 0.92 and 0.98, accompanied by robust precision-recall scores of $0.92-0.98$ and $0.92-0.98$, respectively. The CNN2D model performed even better with an accuracy of 0.95 and 0.99, along with equally commendable robust precision-recall scores of $0.95-0.99$ and $0.95-0.99$, respectively. However, the LSTM model struggled in this category, managing only an accuracy of 0.98 and 0.95 with a precision of 0.89 and 0.85 and a recall of 0.89 and 0.85. Conversely, the SVM model excelled with an accuracy of 0.83 and 0.98, boasting a precision of 0.93 and 0.95 and a recall of 0.93 and 0.97. All precision and recall accuracy of the model were calculated from each number of FN and FP. We are excited to publish our findings and hope to make a substantial contribution to the knowledge and treatment of female depression. We are glad to provide the outcomes of our study.