Speech Feature-Based Depression Detection and Interpretability Analysis

Depression is a common mental disorder whose clinical diagnosis currently relies primarily on subjective scales and physician assessments, lacking objective and rapid detection methods. This paper proposes a depression detection method based on speech signal analysis. First, speech data from participants are collected, from which multiple acoustic featuressuch as Mel-Frequency Cepstral Coefficients (MFCC), fundamental frequency, energy, and formants-are extracted. Next, the XGBoost algorithm is employed to construct a depression classification model, and the model’s performance is evaluated using five-fold cross-validation. Additionally, SHAP (SHapley Additive exPlanations) is utilized to interpret the model’s predictions, revealing the importance of individual acoustic features in the decision-making process. Experimental results indicate that the proposed method achieves strong performance across various metrics, including accuracy, precision, recall, and Area Under the Curve (AUC). Furthermore, the SHAP analysis highlights that certain speech features significantly contribute to distinguishing depression. This study offers a novel approach and provides insights for early screening and intelligent diagnosis of depression using speech signals.