Depressive and Mania Mood State Detection through Voice as a Biomarker Using Machine Learning

Abstract Background Both depressive and mania mood state have high prevalence and are important causes of social burden worldwide, however, there is still no objective indicator for detection. This study aimed to examine if voice could be used as a biomarker to detect these symptoms in China. Methods 1,287 voice messages from 81 subjects were classified into three groups: the depression mood state group (406 voice messages from n = 31), the mania mood state group (192 voice messages from n = 14), and the remission group (689 voice messages from n = 36), based on the scores of the MDQ, QIDS and YMRS. 34 features were extracted from voice records which is collected in real-world emotional diary. A three-group comparison was performed through analysis of Kruskal-Wallis H Test. Three feature extraction methods were adopted and four machine learning methods were performed. Results 33 voice indicators showed differences among the three groups( p < 0.05). Among the machine learning methods, the best performance was obtained using the Gate Recurrent Unit with 79.6% sensitivity, 91.1% specificity and 82.5% sensitivity, 90.7% specificity for the detection of depressive and mania mood state respectively. Conclusions This study further revealed participants with depressive or manic mood state could be accurately distinguished through machine learning. Although this study is limited by a small sample size, it is the first study on voice as a biomarker in both depressive and mania mood state which suggests the possibility of detecting these mood states through voice.