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

BACKGROUND Screening for mental disorders in pregnant women is crucial, but traditional tools such as the Edinburgh Postnatal Depression Scale (EPDS) or Brief Symptom Inventory 18 (BSI18) lack objectivity. Voice analysis and AI show promise in identifying specific mental disorders; however, existing research focuses on general depression, neglecting other disorders in pregnant women. OBJECTIVE This study explored the use of an image classification model to detect vocal patterns and identify mental disorders in pregnant women. METHODS This prospective cohort study was conducted in Japan with 204 pregnant women attending a one-month postnatal checkup. The study involved voice recordings during interviews, supplemented by sociodemographic data and Edinburgh Postnatal Depression Scale (EPDS) scores. Voice samples were cleaned to remove environmental noise and other voices, then segmented and converted into spectrograms. Data augmentation techniques were employed to enhance training data diversity, and Efficientformer V2, pretrained with transfer learning, was used for categorization. Hyperparameters were optimized using Optuna, and an ensemble learning approach was used for final classifications, which were compared against the EPDS. RESULTS Of the 204 pregnant women enrolled, 32 were excluded due to inadequate data, resulting in 172 participants. The participants were divided into three groups: 97 for training, 32 for validation, and 43 for testing. The average audio duration was approximately 549.2±356.0 s. The segmented data yielded 2,942 training, 197 validation, and 323 test segments. After data expansion, the number of training segments increased to 14,170. Various hyperparameters were set for machine learning. The ensemble predictions displayed higher sensitivity (1.00) and recall (0.82) for vocal discrimination and higher specificity (0.97) and precision (0.84) for EPDS discrimination. The receiver operating characteristic area under the curve comparison revealed no significant difference (p = 0.759). CONCLUSIONS Voice analysis and artificial intelligence may provide better screening methods for detecting mental disorders during pregnancy.