Somatisation Disorder Detection via Speech: Introducing a Self-Supervised Learning Model.

Area of Science:

• Psychiatry and Computational Linguistics
• Application of Artificial Intelligence in Mental Health Diagnosis

Background:

• Somatisation disorders are underdiagnosed due to their subtle presentation.
• Accurate speech recognition for somatisation disorder identification requires large, labeled datasets, which are scarce.
• Existing speech recognition methods heavily depend on supervised learning and ample labeled data.

Purpose of the Study:

• To investigate the efficacy of self-supervised learning (SSL) models for identifying somatisation disorders in speech with limited labeled data.
• To compare the performance of three SSL models: Contrastive Predictive Coding (CPC), wav2vec, and wav2vec 2.0.
• To establish a more efficient method for developing diagnostic tools for somatisation disorders.

Main Methods:

• Utilized three pre-trained self-supervised learning models: CPC, wav2vec, and wav2vec 2.0.
• Applied these models to a few-labeled somatisation disorder speech dataset.
• Compared the performance of SSL models against a supervised learning benchmark.

Main Results:

• The wav2vec 2.0 model achieved the highest performance with 77.0% unweighted average recall.
• The wav2vec 2.0 model significantly outperformed the CPC model (p < .005).
• The best SSL model surpassed the performance of the traditional supervised learning model.

Conclusions:

• Self-supervised learning, especially wav2vec 2.0, is effective for somatisation disorder speech recognition with limited labeled data.
• This approach can significantly reduce the burden of manual data labeling for clinical applications.
• The developed model offers a valuable tool to assist psychiatrists in the clinical diagnosis of somatisation disorders.