Enhanced Depression Detection through Optimally Weighted Spectrogram Feature Fusion

Depression is a major psychological concern, and current research on audio-based automatic depression detection often underutilizes the full potential of both physical and spectrogram features in audio data.This paper introduces a novel architecture to bridge this gap.Audio files are preprocessed to extract two key feature sets: Mel-frequency cepstral coefficients (MFCC) and high-level spectrogram features, which are subsequently fused.The proposed Classify-Net processes these fused features layer by layer, treating depression detection as a binary classification task.To enrich the feature set, three distinct spectrograms are used to generate three fused features, with dynamically assigned weights to capture different aspects of the data.The combined probabilities and weights from Classify-Net are then applied to the classification process.To further optimize the exploration of the solution space, a "leading goat" simulated annealing algorithm is introduced, enhancing detection accuracy.Experimental results demonstrate that the proposed method achieves high accuracy and F1 scores across three public datasets, highlighting its robustness and effectiveness.