Assessing Bipolar Episodes Using Speech Cues Derived from Phone Calls

In this work we show how phone call conversations can be used to objectively predict manic and depressive episodes of people suffering from bipolar disorder. In particular, we use phone call statistics, speaking parameters derived from phone conversations and emotional acoustic features to build and test user-specific classification models. Using the random forest classification method, we were able to predict the bipolar states with an average F1 score of 82%. The most important variables for prediction were speaking length and phone call length, the HNR value, the number of short turns and the variance of pitch F-0.