Using Speech to Develop Multivariable Prediction Models for Major Depressive Disorder and Generalized Anxiety Disorder During Pregnancy.

Abstract

Objective: Depression and anxiety are common during pregnancy yet remain underdetected. In this project, we examined the usefulness of speech for predicting major depressive disorder (MDD) and generalized anxiety disorder (GAD) during pregnancy. Methods: We conducted a Prediction Model Development and Evaluation Study using data collected from July 2019 to May 2020 in British Columbia, Canada. MDD and GAD diagnoses were ascertained using the Structured Clinical Interview for DSM-5. We extracted speech features from recorded interviews that contained the voices of both patient and the interviewer (full interview), as well as from manually verified speech segments (VS) of patients. We trained several machine learning models and compared the predictive performance of speech-based models with combined models-consisting of pregnancy/sociodemographic characteristics and speech features-in multivariate analyses. Results: The total sample included 146 participants: 19 with MDD, 28 with GAD (3 had comorbid MDD and GAD), and 102 controls (no current/past known psychiatric disorders). Speech models based on the full interviews showed consistent patterns of good performance for MDD (F1-scores between 74%-77%) and GAD (F1-scores between 76%-80%). Models based on VS showed poor performance for MDD (F1-scores between 42%-45%) and acceptable performance for GAD (F1-scores between 51%-65%). Adding pregnancy/sociodemographic characteristics to the best speech models did not improve the performance of GAD models (best F1-score 78.00%±19.86) and slightly reduced that of MDD models (best F1-score 76.00%±11.45). Conclusion: The most influential speech features for prediction of MDD and GAD during pregnancy aligned with those reported among the general population.