By Susan Roy
70 percent of people have experienced a traumatic event. Only some of them go on to develop symptoms of post-traumatic stress disorder (PTSD) such as flashbacks, nightmares, severe anxiety, and insomnia. The condition can be debilitating and is associated with higher risks of suicide and substance abuse.
To diagnose PTSD, mental health providers have traditionally relied on their own more subjective interviews of patients and patients’ self-reporting. Now, machine learning technologies may enable quicker and more accurate diagnosis and treatment. The following three examples explore how.
More Immediate Intervention
Research has found that symptoms of PTSD can emerge in the first days and weeks after an experience of trauma—and timely intervention can improve treatment outcomes. The challenge for clinicians, though, has been predicting a person’s risks of PTSD in the early aftermath of a traumatic event. That has begged the question of whether it’s even possible to identify who is more likely to develop PTSD, based on their very first post-trauma symptoms.
The answer seems to be “yes”—with machine learning (ML), according to a July 2019 study in the journal JMIR Mental Health. It found that supervised ML, using data collected from smartphone surveys and self-reported symptoms, was able to identify statistically relevant correlations between observable symptoms shortly after a trauma and those that emerged one month later. The resulting prediction algorithm could then accurately target for early intervention those who were at highest risk of PTSD.
More Accurate and Objective Diagnosis
Misdiagnosis of PTSD can be common. One reason is the problem of bias that is unavoidable in a more subjective interview and assessment of symptoms. That has motivated researchers to look for more objective markers for PTSD, by using ML technologies that can analyze voice and text data for meaningful patterns.
A type of statistical ML program known as “random forests,” which can be trained to learn how to classify individuals with increasing accuracy as training data accrues, was remarkably precise at analyzing speech samples to determine the presence of PTSD, in an April 2019 study in the journal Depression and Anxiety. The program was 89.1 percent accurate in objectively identifying veterans with PTSD and those without.
How might voice and speech patterns signal PTSD? An article in Neuroscience News summarizing the findings shed some light:
- A voice can reveal a lot about the speaker, like their mental and emotional state, health, and communication abilities.
- Anecdotal observations have linked PTSD to less clear speech and a lifeless tone.
- And trauma is believed to cause changes in brain circuits that process emotion and muscle tone, and these effects register in the voice.
These PTSD-specific changes in the voice, researchers believe, are a measurable and objective marker for PTSD. (Specific voice features have reportedly been linked to other mental health conditions, too—for example, pressured speech in the case of bipolar disorder and monotone speech in the case of depression.)
The researchers began the study by recording hours-long diagnostic interviews of veterans who had and had not developed PTSD from serving in the military. These recordings were processed by voice software that analyzed words along with other qualities, such as frequency, rhythm, and tone. After the software had produced more than 40,000 speech-based features, the ML program sifted through them to identify patterns. The program was so successful at learning and interpreting these patterns that it soon was accurately diagnosing PTSD most of the time.
Text-based diagnoses are also achievable and accurate most of the time, thanks to ML technologies. After training a ML model, University of Alberta researchers found it could detect PTSD with 80 percent accuracy based on text data alone. (Their findings appeared in an April 2022 synopsis on the University of Alberta’s website.)
In this case, the researchers wanted to see if it would be possible to identify people with PTSD by analyzing the emotional content of interview data. A process known as “sentiment analysis” allowed them to categorize this data into subsets. For example, one subset consisted of those who expressed positive thoughts, while another subset comprised those who expressed negative thoughts. From there, the researchers were able to determine from predominantly negative or neutral speech patterns who had PTSD.
More Individualized Interventions
Research at USC is now underway to explore how ML models might identify individual behavioral triggers for mounting PTSD symptoms and respond preventatively with tailored interventions. The focus of this study is veterans with untreated PTSD who are self-medicating with cannabis. If they can be helped by ML probability models that predict and target escalating symptoms, the study authors believe, less vulnerable populations can be served by the same technique, too.
Machine learning could revolutionize how we diagnose and treat PTSD, by enabling more immediate intervention, more objective and accurate diagnosis, and more personalized interventions. ML models don’t just offer more efficiency at interpreting and categorizing data. They are a whole new way of ordering, perceiving, and understanding the complex relationships hidden within all that data. The three applications featured here only hint at the potentially profound and life-changing implications of machine learning for those with PTSD and their loved ones.
Susan V. Roy is Chief Strategy Officer at the national behavioral health provider FHE Health.
This post has been sponsored by FHE Health
Digital Health Buzz! aims to be the destination of choice when it comes to what’s happening in the digital health world. We are not about news and views, but informative articles and thoughts to apply in your business.