The emotional tone, or sentiment, conveyed in writing can be predictive not only of an individual’s current mood but also future changes in mood, according to a report published today in PNAS. Further, both human reviewers and AI tools were able to predict future depressive symptoms based on a sentiment analysis of written statements.
Jihyun K. Hur, M.A., of Yale University, and colleagues developed nine open-ended questions based on the nine-item Patient Health Questionnaire (PHQ-9), which screens for depressive symptoms. The questions were phrased to remove any negative framing. For example, the question related to fatigue was: “Sometimes we feel tired and exhausted, and sometimes we feel full of energy. How would you describe your energy level in the past 2 weeks?”
Hur and colleagues then invited 467 participants to answer the open-ended mood questions; the participants also completed the standard PHQ-9 a day before and three weeks after they offered their written narratives. The researchers also enrolled 470 additional participants to act as sentiment raters; each rater was given a random subset of responses and asked to rate the positivity and negativity of each statement from zero to 10 using predefined criteria. The sentiment of written responses was also assessed by Linguistic Inquiry and Word Count (a text assessment tool) and ChatGPT (a large language learning model) using the same criteria.
Not surprisingly, individuals with higher baseline PHQ-9 scores were more likely to write narratives with high negative sentiment across all three rating methods. However, even after factoring in individuals’ baseline mood symptoms, high negativity ratings as calculated by human raters or ChatGPT were reliably associated with worsening depressive symptoms at the three-week follow-up.
“These findings underscore the potential of using automated text analysis tools to improve psychological assessments at minimal cost. Our analysis expenses were almost free, less than 0.1 cents per individual based on current GPT-3.5 prices,” Hur and colleagues wrote. “Future research is needed to examine whether our approach can be used to monitor and predict symptoms in clinical populations diagnosed with depressive disorders including those with severe symptoms.”
For related information, see the Psychiatric Services article “Machine Learning–Based Evaluation of Suicide Risk Assessment in Crisis Counseling Calls.”
(Image: Getty Images/iStock/PeopleImages)
Don't miss out! To learn about newly posted articles in Psychiatric News, please sign up here.