One can forecast the worsening of major depressive disorder symptoms by examining a person's word choice. Researchers evaluated written responses using ChatGPT and human evaluators, and they discovered that both methods could reliably forecast the severity of depression weeks later.
Traditional language analysis tools, such as LIWC, could not accurately capture the emotional tone; however, ChatGPT was able to do so by using word order and phrase meaning. This discovery may open the door for AI-assisted mental health assessments, giving medical professionals more resources for diagnosing and forecasting mental health outcomes.
Predicting Depressions through Words
Researchers used human investigators and the big language model ChatGPT to show that written answers to open-ended questions may be used to predict who would have worse depressive symptoms weeks later.
The findings, released Sept. 16 in the Transactions of the National Academy of Sciences, show that automated approaches to assessing language use can complement and enhance mental health assessments.
An increasing amount of studies have found a connection between a person's language use and depression. For example, people who are depressed tend to use more negative, emotive language in texts and on social media. Furthermore, a person's word choice affects how effectively they respond to treatment.
How to Predict Depression through Words
A new way to use the vast amount of language data already available in the clinical setting to better understand mental health is through artificial intelligence tools like ChatGPT.
A technology called ChatGPT uses artificial intelligence to simulate speech used in conversation. As such, word order and the meaning inside and between phrases are considered differently than they would be if typical language analysis methods such as LIWC were used.
According to the researchers, ChatGPT versions 3.5 and 4.0 predicted future changes in depression intensity like that of human raters when they were asked to rate the participants' responses' positive and negative tones.
Testing of Predicts Changes in Depression
Participants in two studies (N = 467) answered neutral, open-ended questions by describing parts of their lives that were related to depression, such as motivation, mood, and sleep.
To monitor mood dynamics, participants also completed a risky decision-making task that involved periodic assessments of momentary happiness and the Patient Health Questionnaire (PHQ-9) to measure depression symptoms.
Three methods were used to assess the sentiment of written responses: the Linguistic Inquiry and Word Count (LIWC) tool, large language models (LLMs; ChatGPT 3.5 and 4.0), and human raters (N = 470).
At a three-week follow-up, alterations in depressed symptoms were predicted by linguistic sentiment assessed by human raters and LLMs but not by LIWC.
Key Facts
- Future depressive symptoms were accurately predicted by ChatGPT and human raters.
- The predictive power of conventional word-counting instruments, such as LIWC, was reduced.
- AI language analysis could enhance clinicians’ abilities to assess mental health.