An Enhanced Depression Detection Using Bi-LSTM AI Model with Audio Data

Depression poses a significant challenge in mental health care because of delayed or missed diagnosis. This research work proposes an innovative approach for detecting depression among individuals by in-depth analysis of the audio data obtained from the patients, leveraging deep learning for early and accurate screening. Audio data in .wav format is analyzed through spectrogram-based representations to capture speech patterns' temporal and frequency features. We compare models like Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), ConvLSTM2D, and Dense Sequential models for feature extraction and classification. Bi-LSTM achieved the highest performance with 90% accuracy, precision, recall, and F1 score, effectively capturing sequential audio dependencies. ConvLSTM2D and Dense Sequential models achieved 88% and 87% accuracy, respectively, whereas traditional machine learning models proved less effective. These results highlight the potential of Bi-LSTM for noninvasive audio-based mental health screening.