UNDERPIN Mental Disease Dialogue Corpus and Mental Disease Classification Using Audio and Linguistic Features with Feature Importance Analysis

Abstract

We have been creating our UNDERPIN mental disease dialogue corpus, which includes more than 1000 hours of recorded voice. Our corpus consists of patients' information (disease name, drugs, etc.), disease test results, and dialogue data. We classified the disease types (bipolar disorder, schizophrenia, dementia, depression, anxiety) versus healthy people, using audio and linguistic features extracted from the corpus. We achieved around 75-91 points in f-score depending on the disease types, which feature importance suggested that formants, fillers, laughs, and questions are important indicators to predict mental diseases.