Predicting Efficacy of Internet-Based Cognitive Behavioral Therapy for Depression through Speech Acoustic Analysis: A Machine Learning Approach (Preprint)

Abstract

Internet-based cognitive behavioral therapy (ICBT) is an effective remote psychological treatment option. However, its effectiveness is usually assessed based on subjective symptom scales, which may lead to inaccurate reflection of the true symptoms or progress in treatment. Using both subjective symptom scales and objective speech acoustic features to evaluate ICBT has the benefits of being non-invasive and convenient. Utilizing reliable and valid measures of patient response based on objective speech indicators would aid in the advancement of ICBT.

The objective of this study was to evaluate the effectiveness of ICBT using speech acoustic features for treatment monitoring and predicting treatment response. We examined changes in symptoms and speech, and developed a machine learning-based classification model that could monitor treatment progression and predict treatment outcome based on speech acoustic features.

A four-week randomized controlled trial was conducted to study the use of ICBT in college students with depression. Speech samples and clinical symptoms were collected at the beginning and end of treatment, and the extracted acoustic features were compared between the ICBT and wait-list groups and analyzed for correlations. An artificial neural network (ANN) was also created to predict the efficacy of ICBT and classify treatment response.

In comparison to the wait-list group, the first formant bandwidth of speech significantly changed in the ICBT group, along with improvements in depressive symptoms following treatment. The efficacy of ICBT and treatment response were predicted using speech features such as the difference in the first and third formants and first formant bandwidth. There was a significant correlation (r=.452, P=.004) between the predicted and true values of the change in PHQ-9 scores from baseline to week 4 of ICBT. Additionally, the classification model built by ANN to identify treatment response and nonresponse had an accuracy rate of 78.37%.

This study identified speech formant as objective biological markers of speech that are closely related to depression and the effectiveness of ICBT. The research also showed that classification models based on key speech acoustic features can be a useful method for tracking progress in psychotherapy and predicting efficacy.

The study was registered at ClinicalTrials.gov (ChiCTR2100045542).