The Discriminatory Power of Vocal Features in Detecting Mental Illnesses Under Complex Context

Background: Vocal features have been proposed as a way to identify depression by distinguishing depression from healthy controls, but while there have been some claims for success, the degree to which changes in vocal features are specific to depression has not been systematically studied. In particular, it is not clear whether vocal features are characteristic of mental ill health in general, rather than characteristic of different psychiatric diagnoses. We examined the performance of vocal features in recognizing three diseases (depression, schizophrenia and bipolar disorder) in comparison with healthy controls and in pairwise comparison with each disease in turn. Methods: : We sampled 32 bipolar disorder patients; 106 depression patients, 114 healthy controls and 20 schizophrenia patients. We extracted i-vectors from MFCCs features, and built logistic regression models with ridge regularization and 5 fold cross validation on the training set, then applied models to the test set. Results: : Our results showed that AUC score for classifying depression and bipolar disorder is 0.5 (F-score = 0.44). For other comparisons, AUC scores range from 0.75 to 0.92 (F-score ranges from 0.73 to 0.91). The performance (AUC) of depression and bipolar disorder classification model is significantly worse than the performance of bipolar disorder and schizophrenia classification model (corrected P < 0.05). We found no significant difference in pairwise ROC difference tests among the remaining classification tasks. Conclusions: : Vocal features have robust discriminatory power not only in classifying depression and health, but also in pairwise classification among different mental illness.