JMIR Publications

ABSTRACT

Background:

Recent progress in identifying depression through speech patterns has accelerated because of the enhanced precision of foundation models. These models, trained on vast unlabeled speech datasets using self-supervised learning (SSL), extract powerful speech characteristics within their transformer encoder layers. The self-supervised representations of speech capture para-linguistic features that contain information about psychomotor retardation symptoms of individuals with major depressive disorder (MDD). Fully automated depression screening methods have been exploring the potential of fine-tuning such foundation models for downstream tasks such as speech emotion recognition and speech-based depression detection. However, it is vital to finetune such models with diverse datasets beyond laboratory-collected ones for the models to be robust and generalized when deployed in real-world applications.

Objective:

This study aims to introduce two novel approaches to improve automated depression screening methods, and improve the system’s robustness. The first objective is to tackle the challenge of limited data, and the second is to train the model in a way that minimizes speaker-specific biases. Achieving these two objectives is vital to ensure that models are robust and generalized when deployed in real world applications.

Methods:

The first objective is achieved by leveraging self-supervised representations and finetuning the model on a diverse real-world dataset with data augmentations. The latter is achieved by using a speaker invariant training architecture enforcing the model to learn para-linguistic features while discarding information related to speaker-specific features. This approach ensures the model discerns speech features indicative of depression, irrespective of the speaker.

Results:

The study validates the methodology across two unique language datasets using only raw acoustic components of speech. The proposed approach employs an adversarial training method, outperforms the baseline model, and achieves a macro F1-Score of 0.83, setting a new state-of-the-art (SOTA) on the publicly available DAIC-WOZ. Similarly, the Oizys Chinese corpus achieves a sensitivity of 0.8 and AUC of 0.82 with a relative improvement of 23% in sensitivity compared to the baseline model.

Conclusions:

The study advances biomedical research, enhancing automated depression screening with an efficient, effective, and simple voice interface-to-screen MDD at scale with high precision and reliability using robust vocal biomarkers. Furthermore, the fine-tuned models capture robust features to detect depression using speech with a minimal speaker bias. These self-supervised representations are objective and consistent across fundamentally different language corpuses; they therefore have the potential of being used for multilingual depression screening using speech. Clinical Trial: Our study was a prospective cohort study rather than a randomized controlled trial. The registry center encouraged but did not mandate the study’s registration. We adhered to ethical standards, meeting the ethical guidelines of and securing approval from the ethics committee of Peking University's Sixth Hospital. All participants provided written informed consent. The study was approved on December 15, 2020 (approval number: 62). The participants' privacy and confidentiality were rigorously upheld.