A speech-based approach for depression detection

Abstract

Depression is a widespread mental health disorder with severe impacts on quality of life. Current screening relies mainly on psychiatrist evaluations, lacking technological support. Advances in AI now enable rapid pattern recognition, allowing researchers to leverage speech analysis for depression detection. As a result, the present thesis aims to a) compare different machine learning models in terms of their classification accuracy in separating depressed and non-depressed people on voice data obtained from the DAIC-WOZ dataset, b) detect significantly important features that contribute to speech patterns related to depression, and c) fine-tune the optimal machine learning model found in stage a) ensuring accurate and robust results for real-world data. The used dataset is part of the Distress Analysis Interview Corpus (DAIC) from the University of Southern California, which aids in diagnosing depression. It consists of 189 English interview sessions with pre-extracted voice features, processed with the Cooperative Voice Analysis Repository (COVAREP) toolbox. In the present thesis, these features were processed via Python programming with the Anaconda Distribution package. The dataset was originally divided into training, validation, and testing sets, but a new split of 80% training and 20% testing was chosen. The 74 pre-extracted COVAREP features represent time-series data, which created a large matrix that posed computational challenges. Two approaches were employed to address these: (1) selecting core features (7000 middle rows of each session) and (2) aggregating time series data into four statistical features per feature, reducing it to one row per participant. Both approaches included preprocessing steps to handle missing or infinite values and standardization. In the first approach, two dataset versions were tested: one non-balanced and one balanced through row deletion (392000 rows for each class). Models evaluated included Neural Networks, Convolutional Neural Networks, Long Short-Term Memory, AdaBoost, Multilayer Perceptron, and Decision Tree. In the second approach, the data balanced with the SMOTE technique was used to evaluate the Support Vector Machine (SVM) algorithm on the aggregated features. The SVM demonstrated the best performance across scenarios, achieving 81% accuracy, 79% precision, 90% recall, 74% F1-score, and 72% specificity. Interpretability tools (LIME, SHAP, PDP) identified three key features contributing to the model’s predictions.