When Dr. David Yusko speaks with people who've suffered traumatic experiences -- roadside explosions, gunshot wounds, rape -- he sometimes has difficulty coaxing people to open up. Patients will shrug off questions, give one-word answers or decline to discuss issues like difficulty sleeping or nightmares.
So Yusko, associate director of the University of Pennsylvania’s Center for the Treatment and Study of Anxiety, is now considering another way to diagnose post-traumatic stress disorder, one that may at first seem crassly impersonal: software.
SentiMetrix, a data analytics firm specializing in sentiment analysis, has developed a computer program that can identify symptoms of PTSD by scanning soldiers' blog posts and online musings for evidence of anxiety, depression and other traits associated with the condition.
The work has at its core a seemingly counterintuitive notion: that the impersonal, sterile combination of a screen and software could more effectively uncover someone's mental state than a face-to-face conversation between a patient and doctor.
Yet the goal, according to SentiMetrix co-founder V.S. Subrahmanian, is not to replace clinicians’ diagnoses, but to more effectively identify PTSD sufferers who may not be aware of their condition and encourage them to seek treatment.
The algorithms may eventually be able to more quickly find and process blog posts than a human physician, allowing SentiMetrix to potentially identify PTSD sufferers who have until now gone undiagnosed. A computer program's more uniform analysis could also allow clinicians to better track how their patients respond to different therapies.
“We’re not trying to say we're better than clinicians. We’re not, and I doubt if any system will be better than these people who’ve spent their whole lives looking at PTSD in great detail,” said Subrahmanian. But, he notes, “It’s difficult [for clinicians] to figure out who should be brought in for treatment if they’re not in front of them.”
The burgeoning field of sentiment analysis uses software to analyze text -- such as tweets, Facebook status updates, or online reviews -- in order to identify the intent, emotional state or even political affiliation of the person who wrote it. Companies have tapped the technology to keep tabs on whether their customers are satisfied, and it’s been pitched as a way for Wall Street traders to track attitudes toward the stock market.
Though Yusko isn't intimately familiar with SentiMetrix's research, he sees potential in its work, noting that veterans may reveal more about their mental state in blog posts or journal entries than they will when asked questions directly by a clinician. In addition, veterans’ writings, authored over a period of time, could provide a more detailed view of their health than a doctor could glean from a 50-minute session with a patient.
“What this has the ability to do is really access people’s natural language via their writing, where they’re expressing these symptoms and may be more willing to seek help and become aware of it,” said Yusko. “Sometimes when you’re asking direct questions, people aren’t ready to talk about it or do something about it, and they say, ‘no.’ But in a blog or online journal, people are more likely to express what they’re really going through in an uncensored way.”
As it scans a written passage, SentiMetrix’s algorithm looks for evidence of any one of thirteen symptoms associated with PTSD, including feeling “jumpy or easily startled,” “distant or cut off from other people,” or a “loss of interest in things that were enjoyable in the past.” Though the software can also pick up on physiological symptoms, such as sleep trouble, there are four symptoms associated with PTSD that SentiMetrix cannot yet extract, such as “feeling as though your future may be cut short,” or avoiding situations that recall a stressful experience from the past.
The SentiMetrix program rates each symptom on a zero-to-one point scale, with a higher score indicating the software has greater confidence that the text’s author suffers from that symptom.
The algorithm relies on natural language processing technology, which allows computers to deduce the meaning of someone’s writing, even if certain key terms haven’t been mentioned explicitly. For example, thanks to natural language processing, Apple’s Siri knows to deliver a weather report if asked, “Do I need a sweater today?” or “Should I bring an umbrella to Paris?” -- there's no need to ever mention "weather."
SentiMetrix has a database of “stressful past experiences” frequently discussed by PTSD sufferers, such as car accidents or explosions, that it looks for in the passages it “reads.” Subrahmanian notes that the company trained its algorithm by teaching it to mimic clinicians’ thinking: They had two clinical psychologists review several hundred publicly available -- but anonymized -- blog posts penned by veterans, which they then annotated by noting which symptoms were present.
The most recent version of SentiMetrix’s software, one of four algorithms it plans to develop for PTSD analysis, boasts an 80 percent accuracy rate, meaning its diagnosis accords with a human clinician’s in the majority of cases. The program has been in development for two years and, earlier this year, received funding from the U.S. Army Medical Research and Materiel Command.
If the technology progresses enough, Subrahmanian predicts that veterans’ groups could eventually decide to set up online forums where returning members of the military could blog and, with their permission, that writing would be monitored for signs of PTSD. SentiMetrix’s research could also be applied to identifying civilians with the condition.
Yet Yusko notes that diagnosing PTSD is only one challenge: according to a study underway at Veterans Affairs hospitals in Minneapolis and Philadelphia, a substantial number of veterans refuse treatment, even after they’ve received their diagnosis.
“The biggest difficulty has been convincing young veterans to attend to their symptoms,” said Yusko. “While finding out that they do [have PTSD] is useful information, we often can’t do anything about it.”