Detecção de depressão em gestantes através da voz com deep learning

Abstract

Depression is a mental disorder that poses a serious public health challenge, especially among pregnant women, with an estimated prevalence of 10% to 20% during pregnancy. In this context, early detection is essential to ensure appropriate interventions. This study aims to investigate the feasibility of using neural networks to identify depressive symptoms in pregnant women through voice analysis captured by smartphones. A total of 116 audio recordings from 93 patients were used, collected in uncontrolled environments. The audio data underwent preprocessing, segmentation, and conversion into spectrograms, which were then fed into a convolutional neural network. To address the small dataset size, a transfer learning technique was applied using a pre-trained model. Evaluation was performed using cross-validation, with the ROC curve and AUC as performance metrics. The model achieved an average AUC of 0.63. Although still insufficient for clinical application, the results indicate the potential of this approach for initial screening of perinatal depression using accessible devices such as smartphones. It is expected that with more data and further model refinement, diagnostic accuracy can be significantly improved.