Early Detection of Postpartum Depression Using Deep Learning Analysis of Maternal Speech Signals

Postpartum depression (PPD) affects approximately ۱۰-۱۵% of new mothers worldwide, yet it remains underdiagnosed due to stigma, limited screening resources, and lack of objective biomarkers. This study proposes a novel deep learning approach for early PPD detection using acoustic analysis of maternal speech. A dataset of ۴۵۰ postpartum women (۲۲۵ with PPD diagnosed by Edinburgh Postnatal Depression Scale, ۲۲۵ healthy controls) was collected. Participants completed a ۵-minute structured interview describing their daily experiences. Audio recordings were segmented and ۸۸ acoustic features were extracted, including Mel-frequency cepstral coefficients (MFCCs), pitch, jitter, shimmer, harmonics-to-noise ratio, and formants. A hybrid deep learning architecture combining bidirectional long short-term memory (BiLSTM) networks with attention mechanism was developed for PPD classification. The proposed model achieved an accuracy of ۸۷.۳%, sensitivity of ۸۵.۶%, specificity of ۸۹.۱%, and area under the curve (AUC) of ۰.۹۲. The most discriminative features included MFCCs ۲-۵, pitch variability, and speech rate. This proof-of-concept study demonstrates that vocal biomarkers can serve as an accessible, non-invasive, and objective screening tool for PPD, enabling early intervention and improved maternal mental health outcomes.

کلیدواژه ها:

Postpartum Depression ، Deep Learning ، Speech Signal Processing ، Acoustic Analysis ، Maternal Mental Health ، Digital Biomarker