A texting-based crisis line is using big data and artificial intelligence to flag its higher-risk users—even by looking at the emoticons they use, Brian Resnick reports for Vox.
Crisis Text Line (CTL) started in 2013 with the goal of helping people who were experiencing a mental health crisis but didn't feel comfortable calling a counselor. Today, CTL has around 10 so-called "active rescues" per day, during which CTL staff call emergency services to intervene when someone actively attempting suicide.
The system allows people to text a number to be connected with a counselor. But CTL can't always quickly connect with those who reach out. Bob Filbin, the chief data scientist at CTL, said after certain high-profile events—such as the death of comedian and actor Robin Williams or a terrorist attack—demand can double. That can cause long wait times and prompt tough questions about who connects with a counselor first.
Leveraging big data
To help answer those questions, CTL turned to data and machine learning to triage and prioritize the most at-risk texters, Resnick reports.
The crisis line has collected massive amounts of anonymized, encrypted data from the 30 million texts it has exchanged with users, along with metadata (such as when the text came through and the counselor with whom they texted) and survey responses. Users ultimately have control over the data, and can erase any existing data by texting "loofah," Filbin said.
To mine the existing data, CTL trained an algorithm to identify language used in instances when CTL launches an active rescue. The results, Filbin said, were surprising.
CTL initially thought words such as "die" and "cut" associated with an active rescue. But the algorithm found "thousands of words and phrases indicative of [requiring] an active rescue that are actually more predictive," Filbin said. For instance, words like "Advil" and the names of other common household drugs are 14 times more predictive of an attempted death by suicide than the words they had initially hypothesized, according to Filbin.
"Even the crying face emoticon—that's 11 times as predictive as the word 'suicide' that somebody's going to need an active rescue," Filbin said. The data also showed anxiety levels are at their highest on Wednesday, and crises involving self-harm often happen late at night.
Data also has helped CTL evolve as an organization. For instance, it found that just 3 percent of its users use 34 percent of its counselors' time. Filbin said in some cases, those users are using the service as a form of therapy rather than a crisis hotline. "The data really exposed that difficult question and forced us as an organization to make a philosophical stand that we're here for people in crisis," Filbin said.
Making it work
As CTL feeds more data into its algorithms, Filbin expects they will improve: The crisis line eventually hopes it will be able to use the data not just to help those in crisis, but to help predict and prevent instances of self-harm from occurring in the first place. And everything is aimed at building upon what crisis centers are already doing. "We're never rolling out a product that decreases performance," he said.
Resnick notes that some people in the mental health profession prefer using intuition and experience to solve problems over data analysis. But Filbin argued that's a false distinction because data analysis is just a way to learn from experience at scale. "We're saying, no human can understand the scope and reflect on the performance and service as a whole, so we need data to allow us to reflect and improve."
At the same time, Filbin said while CTL has learned a tremendous amount from data, it doesn't want to take humans out of the loop when prioritizing users. "We flag a texter who appears to be circling," he said. "Then the supervisor makes the call. We always have the human making the decision" (Resnick, Vox, 3/30).
Here are the 2017 telehealth industry trends
Join our population health experts on Tuesday, April 25th to get the latest updates on the current state of the telehealth market, get guidance on common planning pitfalls to avoid, and see how leading provider organizations have managed to achieve success using virtual care delivery platforms.