Enhancing Depression Detection through 1D Convolutional Neural Networks on DAIC-WOZ Dataset for Investigation of Visual Cues

Abstract

With the increasing prevalence of depression globally, there is a growing demand for advanced, standardized diagnostic tools that can assist in early identification and intervention. This work utilises deep learning algorithms to address the growing demand for standardised and reliable diagnostic tools in depression identification. In particular, we investigate how 1D Convolutional Neural Networks (CNNs) can be utilised to analyse visual characteristics from the extensive DAIC-WOZ dataset, a collection of clinical interview sessions. The proposed architecture utilises a Deep CNN designed to discern intricate patterns in voice acoustics and facial expressions, aiming to achieve state-of-the-art precision and effectiveness. To improve the model's performance, we tested several dropout rates (0.3 and 0.5) and learning rate (0.001 and 0.0001) setups. The setup with a learning rate of 0.0001 and a dropout rate of 0.5 had the best overall performance, according to the results, with a ROC AUC of 0.79 and macro-average precision, recall, and F1-score of 0.79 across classes. According to this ideal setup, a higher dropout rate combined with a lower learning rate improves the model's ability to generalize, most likely by avoiding overfitting and enabling it to more successfully identify minute patterns in the data.