Predictive modeling of neuroticism in depressed and non-depressed cohorts using voice features

Background

Neuroticism's impact on psychopathological and physical health issues has significant public health implications. Multiple studies confirm its predictive effect on suicide risk among depressed patients. However, previous research lacks a standardized criterion for assessing neuroticism through speech, often relying on simple features (such as pitch, loudness and MFCCs). This study aims to improve upon this by extracting features using advanced pre-trained speaker embedding models (i-vector and x-vector extractors). Additionally, unlike prior studies utilizing general population data, we explore neuroticism prediction in depressed and non-depressed subgroups.

Methods

We collected edited discourse data from clinical interviews of 3580 depressed individuals and 4016 healthy individuals from the CONVERGE study. Instead of solely extracting Low-Level Acoustic Descriptors, we incorporated i-vector and x-vector features. We compared the performance of three different features in predicting neuroticism and explored their combination to enhance model accuracy.

Results

The SVR model, combining three speech features with downscaled features to 300, exhibited the highest performance in predicting neuroticism scores. It achieved a coefficient of determination (R-squared) of 0.3 or higher and a correlation of 0.56 between predicted and actual values. The predictive classification accuracy of speech features for neuroticism in specific populations (healthy and depressed) exceeded 60 %.

Limitations

This study included only women.

Conclusion

Combining diverse speech features enhances the predictive capacity of models using speech features to assess neuroticism, particularly in specific populations. This study lays the foundation for future exploration of speech features in neuroticism prediction.