Voice biomarkers of perinatal depression: cross-sectional nationwide pilot study report

TL;DR

This study demonstrates that machine learning models using 2-minute unstructured speech samples can feasibly and accurately predict perinatal depression in pregnant women, offering a potential digital biomarker for early detection.

Contribution

First to develop and validate voice biomarkers specifically for predicting perinatal depression in pregnant women using machine learning on unstructured speech data.

Findings

Achieved ~71% balanced accuracy with PHQ-8

Achieved 80% balanced accuracy with EPDS

Demonstrated feasibility of large-scale voice-based depression screening

Abstract

Perinatal depression (PND) affects 1 in 5 mothers, with 85% lacking support. Digital health tools offer early identification and prevention, potentially reducing PND risk by over 50% and improving engagement. Despite high interest, user retention needs improvement for maximum benefit. Voice biomarkers have emerged as a possible digital alternative to collecting mood questionnaires to inform individualized user adherence and retention. However, no voice biomarkers have been developed specifically for pregnant mothers. We aimed to test an unstructured speech sample for the feasibility of yielding a machine learning (ML) model to predict PND. We surveyed 446 women nationwide at 22 weeks of gestation. We measured PHQ-8 and modified EPDS and four voice samples. Participants were classified as depressed if their PHQ-8 or EPDS were equal or higher than 10. In the present report, we focus on a 2-minute unstructured speech prompt. A standardized cleaning pipeline was applied to the audio files, including a uniform format conversion, segment filtering for background noise, extraneous acoustic events, and file size. This resulted in 112 high-quality voice recordings, i.e., a rate of 25%. The ML modeling strategy utilized the Allie ML framework on 80% of the data for training on the binary classification task (depressed or not). The analysis included steps for feature selection using referenced speech audio embeddings, followed by a TPOT classifier for ML. We report the validation results for the 20% held-out data. For PHQ-8, we achieved a balanced accuracy of ~71%, precision of 73%, recall of 68%, and AUC of 0.71. For EPDS, we achieved a balanced accuracy of 80%, precision of 92%, recall of 80%, and AUC of 0.80. In conclusion, we demonstrate the feasibility at scale and a promising accuracy for detecting PND using 120 seconds of unstructured speech audio data in expectant mothers.