Predicting Depression Severity from Spontaneous Speech as Prompted by a Virtual Agent

Introduction One of the major challenges in clinical psychiatry remains absence well established objective measures symptoms’ severity. Clinical insights are mainly provided through keen behavioral observation and subjective questionnaires scales. Objectives The aim this paper is to predict depression severity speech using features extracted from as by participants during a semi-structured dialogue with virtual avatar. Methods We use data subset DAICWOZ dataset consisting 142 dialogues between avatar which uses several prompts maintain conversation participant. involving topics travel, dream jobs, memorable experiences. From generated dialogue, we extract participant utterances separated prompt three sets transcripts. content transcript acoustic excerpt corresponding for question.We perform regression experiments on PHQ8 items each set Furthermore, combine transcripts compute partial spearman correlations them gender covariate. Results With our best model obtain an R2 0.1, explaining 10% variance PHQ total score. Additionally, mean absolute error 1.25, suggesting that regressor can detect more or less precision clinically meaningful differences participants. Partial score show significant dependent amount participant, along complexity syntactic structures used. Conclusions Automatic analysis spontaneous could help detection monitoring signs depression. By combining technology timely intervention strategies instance agent it contribute prevention. Disclosure Interest A. König: None Declared, M. Mina Employee of: ki:elements GmbH, S. Schäfer N. Linz Shareolder J. Tröger GmbH