Bipolar disorder causes periods of severe depression punctuated by periods of elevated mood or mania. People with the condition behave in extreme ways, experiencing extreme highs and hyperactivity followed by devastating lows and lethargy. Some estimates suggest that 30 percent will die by suicide.
One way to prevent the most extreme behaviors is to spot the symptoms as they are developing but before they manifest completely. This allows treatment to begin early. So a way of spotting these early signs automatically would have huge implications for sufferers, their families, and health-care providers.
Today, Yen-Hao Huang and pals at the National Tsing Hua University in Taiwan say they have developed a way to identify the early signs of bipolar disorder via social media. They say their method could have significant implications for the way potential patients are assessed.
The onset of bipolar disorder is characterized by symptoms such as overtalking, disturbed sleep, and rapid mood changes. And it turns out that many sufferers share details of their condition, including their diagnosis dates, on social-media platforms such as Twitter.
That gave the researchers an idea. Given that they could be sure that tweets were from people with a bipolar diagnosis, what patterns of behavior might they have demonstrated in advance?
To find out, these guys analyzed some 10,000 tweets posted between 2006 and 2016 by more than 400 people who had a diagnosis of bipolar disorder. They compared these tweets with those from a similar number of people picked at random, who acted as a control group.
The team checked the pattern of posting over time to see how it matched normal sleeping patterns. They looked at the frequency of tweets to gauge how loquacious each user was. They studied the types of words used in each tweet for sentiment and emotional content.
They also developed an entirely new phonological measure by working out the plosive energy of each word as if it were voiced. This idea was based on the thinking that people with early signs of bipolar disorder use more high-energy words.
The researchers then used a sliding-window approach to see how the content of each person’s Twitter stream changed over time, particularly as it approached the point of a diagnosis.
Finally, the team trained a machine-learning algorithm to use combinations of these features to distinguish between people with and without early signs of bipolar disorder. They achieved an identification accuracy of more than 90 percent.
The team’s new measure of the phonological energy of each word is a particularly good one. “By simply employing the phonological feature with pure text ensemble model, the classifier can achieve more than 91 percent precision,” they say.
Interestingly, Huang and co call this approach subconscious crowdsourcing. They point out that the set of tweets from a person suffering bipolar disorder can provide a rich stream of information about mental state. So these people are subconsciously providing a data set that can be mined for information.
Just how much more information can be gleaned this way isn’t clear. But bipolar disorder is unlikely to be the only mental state that can be identified.
That’s interesting work that has the potential to give people with bipolar disorder the treatment they need as soon as it is feasibly possible.
“Our experimental results demonstrate that the proposed models could greatly contribute to the regular assessments of people with bipolar disorder, which is important in the primary care setting,” they say.
And that should minimize the chances of the extreme behaviors that might otherwise result in the worst possible outcome.
Ref: arxiv.org/abs/1712.09183 : “Detection of the Prodromal Phase of Bipolar Disorder from Psychological and Phonological Aspects in Social Media”