Voice Analysis and Deep Learning for Detecting Mental Disorders in Pregnant Women: A Cross-sectional Study

Abstract

Introduction: Perinatal mental disorders are common, affecting 10–20% of pregnant women. Traditional screening tools, such as the Edinburgh Postnatal Depression Scale (EPDS), have subjective limitations, and healthcare providers often face challenges in screening owing to time constraints. Therefore, there is a need for more objective screening methods. Voice analysis has shown promise in detecting mental disorders; however, research on pregnant women is limited. This study aimed to develop a machine learning model that analyzes the voices of pregnant women to screen for mental disorders using a balanced data approach. Methods: In this cross-sectional study, we collected voice samples from 204 pregnant women during one-month postpartum checkup. We preprocessed the audio data, segmented it into 5000 ms intervals, and converted it into melspectrograms using a short-time Fourier transform with different window widths. We applied data augmentation techniques, including TrivialAugment and context-rich minority oversampling, to enhance the training data. We employed transfer learning using the Efficientformer V2-L model pretrained on ImageNet for classification. We optimized the hyperparameters using Optuna to improve the generalization. We combined these predictions using ensemble learning for the final predictions. Results: We included 172 participants in the analysis (149 without mental disorders and 23 with mental disorders). The voice-based model demonstrated higher sensitivity (1.00) and recall (0.82), whereas the EPDS showed higher specificity (0.97) and precision (0.84). The area under the receiver operating characteristic curve revealed no significant difference (P = 0.759) between the two methods. Discussion: Our study demonstrates the potential of voice analysis and deep learning as objective screening tools for perinatal mental disorders. The voice-based model performed comparably to the EPDS, with higher sensitivity and recall, indicating its potential to identify more women at risk for mental disorders. Conclusion: Voice analysis and deep learning show promise as innovative, objective screening tools for perinatal mental disorders.