Using topic modeling to infer the emotional state of people living with Parkinson's disease.

Abstract

Individuals with Parkinson's disease (PD) often exhibit facial masking (hypomimia), which causes reduced facial expressiveness. This can make it difficult for those who interact with the person to correctly read their emotional state and can lead to problematic social and therapeutic interactions. In this article, we develop a probabilistic model for an assistive device, which can automatically infer the emotional state of a person with PD using the topics that arise during the course of a conversation. We envision that the model can be situated in a device that could monitor the emotional content of the interaction between the caregiver and a person living with PD, providing feedback to the caregiver in order to correct their immediate and perhaps incorrect impressions arising from a reliance on facial expressions. We compare and contrast two approaches: using the Latent Dirichlet Allocation (LDA) generative model as the basis for an unsupervised learning tool, and using a human-crafted sentiment analysis tool, the Linguistic Inquiry and Word Count (LIWC). We evaluated both approaches using standard machine learning performance metrics such as precision, recall, and $F_1$scores. Our performance analysis of the two approaches suggests that LDA is a suitable classifier when the word count in a document is approximately that of the average sentence, i.e., 13 words. In that case, the LDA model correctly predicts the interview category 86% of the time and LIWC correctly predicts it 29% of the time. On the other hand, when tested with interviews with an average word count of 303 words, the LDA model correctly predicts the interview category 56% of the time and LIWC, 74% of the time. Advantages and disadvantages of the two approaches are discussed.