Exploring the interpretability in speech-based adolescent depression detection by SHAP
Dong Wang, Qifei Li, Yingming Gao, Yong Liu, Ya Li · 2023
Adolescent depression is more harmful because teens are in a critical period of mental development. The combination of speech analysis and machine learning techniques obtains promising results in detecting depression. However, the current research mainly focuses on enhancing the performance of models and lacks explanation of the model decision process. This paper aims to investigate which features are of importance in identifying adolescent depression and how these features affect the predictions of the model. We extracted 225 acoustic features form the recordings of teenagers reading three Mandarin paragraphs, and then built an ensenble machine learning model that achieved a mean F1 score of 0.853. Combined with the model interpretation framework SHAP, we found that the dispersion of the first formant contributed most to the predictions. Our work is transparent to the process of model decision making, which may promote the the application of machine learning in healthcare fields.