PREDICTING CLINICAL SYMPTOMS OF DEPRESSION FROM ACOUSTIC SPEECH SIGNALS USING NEURAL NETWORKS

Abstract

Approximately 280 million people suffer from depression, a disabling illness. Early diagnosis and effective monitoring are known to reduce adverse effects. Still, they require extensive clinical resources, thus motivating considerable work in automatic detection of depression, including from acoustic speech signals, with some recent success using deep learning. Much less work has been done for automated assessment. We make progress towards automated assessment by presenting the first approach to use acoustic features of speech to predict responses for individual items on validated clinical assessment tools and demonstrate results better than a majority-based baseline on many of the items. We achieve this using CNN, and LSTM architectures whose inputs are speech signals’ acoustic features and outputs are distributions over individual item responses corresponding roughly to presence/absence of each such symptom. This approach provides valuable explanatory power as it inherently predicts which symptoms might lead to the overall assessment score.