(Ames) Iowa State University scientists say their research could lead to better methods of recognizing when someone is experiencing mental health struggles based on their online presence, making it easier to offer resources to those in need.
ISU assistant professor of information systems and business analytics Wenli Zhang, in collaboration with researchers at universities in Delaware and Rhode Island, has developed a “deep learning model” for detecting mental health symptoms on social media, The model factors in more than just positive and negative sentiments on a subject that a person might post online.
This method could and would not replace medical diagnoses, Zhang said. But in the future, she said, the learning model could be used as a tool to provide additional information and resources to those expressing symptoms of mental health issues and to study population-level data over a period of time.
“Maybe we can use big data and machine learning to help … this process,” Zhang said. “Because nowadays, not just in the United States or other countries but in the entire world, mental disorders, especially depression, are underdiagnosed and undertreated.”
The team used a large dataset sourced from Reddit by a different group and offered to researchers for their own work, Zhang said, but the learning model her group created allowed them to dive deeper than others.
Zhang said the goal behind deep learning is to use data to identify patterns in order to detect connections and predict outcomes. As social media data is much more than just numbers in rows and columns, she said the model used to sift through it needed to be intricate.
“When we have unstructured data, like natural language or images, the patterns are very complicated,” Zhang said. “The size of the inputs are very large, so we need a very complex model to … identify the patterns.”
The difference between the team’s new deep learning model and others is what data it pulled as evidence of possible mental health symptoms, Zhang said.
While previous research has relied on “sentiment scores” calculated by finding language expressing someone’s positive or negative emotions in their posts, Zhang said they found little connection between these scores and depression diagnoses. Posters could be referring to something specific they did or didn’t like, she said, like movies or music, which is not what researchers were looking for.
Instead of trying to employ the same method, Zhang said the team turned to language used in more clinical settings when it comes to research and diagnosis, then tried matching it to more natural text and speech.
One example she gave was the term “attempted suicide” — found in clinical writing, but not often on social media. The team used “natural language processing techniques” to link informal language from Reddit to formal terms, Zhang said, then looked at how many times the poster would “mention the symptoms or treatments or major life events that may exacerbate mental disorders.”
Zhang said her experience with modeling such as this comes from her background working in health care analytics during her Ph.D. studies. She used machine learning to track social media data to detect symptoms of asthma and other conditions.
When she came to ISU and began teaching, Zhang said she found depression and other mental health conditions to be a big issue. She said she hopes that in the future, the model can be used by social media platforms as a tool to support at-risk people, like providing a pop-up message with relevant resources and the statement that they are not alone.
“Maybe they can use our tools to identify, maybe, the teenagers that need more help, or maybe senior people that need more help,” Zhang said.
