ASPO Abstracts
Using natural language processing to determine predictors of healthy diet and physical activity behavior change in ovarian cancer survivors
Category: Behavioral Science & Health Communication
Conference Year: 2020
Abstract Body:
Purpose of the study: To explore the use of speech technology and natural language processing in evaluating
language and vocalics as predictors of behavior change in ovarian cancer survivors participating in a lifestyle
intervention.
Methods: Recorded telephone coaching sessions from women participating in the Lifestyle Intervention for
Ovarian cancer Enhanced Survival (LIVES) study were used for this analysis. LIVES is testing whether women
randomly assigned to a lifestyle intervention promoting a high vegetable, fruit and fiber and low-fat diet and
increased physical activity will have increased progression free survival as compared to women assigned to an
attention control. Motivational interviewing, a directive, patient-centered approach, is used to elicit behavior
change. A 10% random sample of call recordings were scored for protocol fidelity. Three automated speech
recognition programs, Google cloud to speech, AWS transcriber, and Watson speech to text were tested. The text
transcriptions were analyzed by a natural language processing expert for how well they retained the information
necessary for evaluating fidelity of the motivational interview to the LIVES protocol. Using the OpenSMILE
acoustic feature extraction library, the audio was analyzed by a speech technology expert for how well different
aspects of the speech signal (e.g., pitch, spectral energy) reflect low vs. high participant achievement.
Results: The three automated speech recognition programs accurately detected between 72 and 76% of text, with
Google cloud to speech performing the best. Additionally, the text was correctly attributed to the speaker 68% of
the time (32% DER score). Analysis of the transcriptions suggests that the Google output recovers the critical
words and phrases for five of six different measures of fidelity to the LIVES protocol. Analysis of the audio
suggests that high-achieving participants show more variable pitch than low-achieving patients.
Conclusions: Next steps will include analysis of the more than 33,000 recorded hours of LIVES calls for language
and sentiment in relation to diet and physical activity behavior change. Speech technology and natural language
processing hold high potential for identifying characteristics of language used in coaching calls
Keywords: speech technology, natural language processing, ovarian cancer, sentiment