Using topic modeling to infer the emotional state of people living with Parkinson's disease.
Andrew P Valenti, Meia Chita-Tegmark, Linda Tickle-Degnen and 2 others
PMID 31194649WHAT IT FOUND
For short speech turns, an automated topic model predicted emotional content better than a standard word-count tool.
For longer interviews, the word-count tool was more accurate. Neither method is ready for clinical use, but the topic model may help caregivers interpret brief exchanges where facial masking obscures true feelings.
Key findings
01The LDA model outperformed the LIWC model for short text (average 13 words), with an F1 score of 0.80 versus 0.44.
02For longer text (average 303 words), the LIWC model achieved higher accuracy (76%) and F1 score (0.74) compared to the LDA model.
03The LDA model is less sensitive to language and transcription errors, making it potentially more suitable for spoken dialogue with people with Parkinson's disease.
STILL TO COME
How it was doneWhat they foundWhat it means for SLPs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The study used a small dataset of transcripts, which limits the generalizability of the machine learning models. The LDA model's topics are not interpretable by humans, so it cannot explain why a text was classified as positive or negative. The models were tested on written transcripts of spoken interviews, not real-time speech, so accuracy may drop with transcription errors from speech recognition. Participants were screened to exclude depression, so results may not apply to people with Parkinson's who also have mood disorders.
Declared interests
The research was supported by the National Institute of Health and the National Institute on Disability, Independent Living, and Rehabilitation Research.
The easy way to misread this
Do not interpret the LDA model's success on short text as evidence that automated emotion detection is ready for clinical use. The study used pre-processed transcripts from controlled interviews, not real-time speech, and the models had significant error rates even in their best conditions.