Northwell Health researchers using Facebook data and AI to spot early stages of severe psychiatric illness

“Harnessing social media platforms could be a significant step forward for psychiatry, which is limited by its reliance on mostly retrospective, self-reported data," according to Northwell Health's Michael Birnbaum, M.D. (Northwell Health)

After going missing for three days in 2016, Christian Herrera Gaton of Jackson Heights, New York, was diagnosed with bipolar disorder type 1.

His experiences with bipolar disorder include mood swings, depression and manic episodes. During a recent bout with the illness, he was admitted to Zucker Hillside Hospital in August 2020 due to some stress he was feeling from the COVID-19 pandemic.

While at Zucker for treatment, the Feinstein Institutes for Medical Research, the research arm of New York’s Northwell Health, approached him to join a study about Facebook data and psychiatric conditions.

The goal of the study was to use machine learning algorithms to predict a patient’s psychiatric diagnosis more than a year in advance of an official diagnosis and stay in the hospital.

Michael Birnbaum, M.D., assistant professor at Feinstein Institutes’ Institute of Behavioral Science, saw an opportunity to use the social media platforms that are a part of everyday life to gain insights into the early stages of severe psychiatric illness.

“There was an interest in harnessing these ubiquitous, widely used platforms in understanding how we could improve the work that we do,” Birnbaum said in an interview. “We wanted to know what we can learn from the digital universe and all of the data that's being created and uploaded by the young folks that we treat. That's what motivated our interest.”

RELATED: Brigham and Women's taps mental health startup to use AI to track providers' stress

After Gaton, a former student at John Jay College of Criminal Justice, was discharged from the hospital, he shared almost 10 years of Facebook and Instagram data with the Feinstein Institutes. He uploaded an archive that contained pictures, private messages and basic user information.

“It's been a difficult experience to deal with [COVID] and to go through everything with the hospitals and losing friends because of doing stupid things during manic episodes,” Gaton told Fierce Healthcare. “It's not easy, but at least I get to join this research study and help other people.”

The study, conducted along with IBM Research, looked at patients with schizophrenia spectrum disorders and mood disorders. Feinstein Institutes researchers handled the participant recruitment and assessments as well as data collection and analysis. Meanwhile, IBM developed the machine learning algorithms that researchers used to analyze Facebook data.

Results of the study, called “Identifying signals associated with psychiatric illness utilizing language and images posted to Facebook,” was published Dec. 3 in Nature Partner Journals (npj) Schizophrenia.

How machine learning helps analyze social media data

Feinstein Institutes and IBM researchers studied archives of people in an early treatment program to extract meaning from the data to gain an understanding of how people with mental illness use social media.

“Essentially, at its core, the machine learns to predict which group an individual belongs to, based on data that we feed it,” Birnbaum explained. “So, for example, if we show the computer a Facebook post and then we say to the computer, based on what you've learned so far and based on the patterns that you recognize, does this post belong to an individual with schizophrenia or bipolar disorder? Then the computer makes a prediction.”

Birnbaum added that the greater the predictions and accuracy, the more effective the algorithms are at predicting which characteristics belong to which group of people.

Feinstein and IBM took care to anonymize the social media data, according to Birnbaum. They stripped out names and addresses from written posts. “Words essentially using language-analytic software become vectors,” Birnbaum said. “The actual content of the sentences, once they're parsed through the software, often becomes meaningless.”

Michael Birnbaum, M.D. (Northwell Health)

In addition, the machine learning software does not analyze participants’ images closely. Instead, it focuses on shape, size, height, contrast and colors, Birnbaum said.

“We did our best to ensure that we identified the data to the extent possible and ensured the confidentiality of our participants because that's one of our top priorities, of course,” Birnbaum said.

Studying communication patterns of research subjects

The study analyzed Facebook data for the 18 months prior to help predict a patient’s diagnosis or hospitalization a year in advance.

Researchers used machine learning algorithms to study 3.4 million Facebook messages and 142,390 images for 223 participants for up to 18 months before their first psychiatric hospitalization. Study subjects with schizophrenia spectrum disorders and mood disorders were more prone to discuss anger, swearing, sex and negative emotions in their Facebook posts, according to the study.

RELATED: Northwell Health research arm develops AI tool to help hospital patients avoid sleepless nights

Birnbaum sees an opportunity to use the data from social media platforms to gain insights to deliver better healthcare. By using social media, such as analyzing Facebook status updates, researchers can gain insights on personality traits, demographics, political views and substance use.

“Harnessing social media platforms could be a significant step forward for psychiatry, which is limited by its reliance on mostly retrospective, self-reported data,” the study stated.

Gaton believes that he could have avoided time in the hospital if he received an earlier diagnosis. Like with other subjects in the study, Gaton can sense the warning signs of an episode when he starts to post differently on Facebook.

From analyzing the data, researchers were able to study who would use more swear words compared with healthy volunteers. Some participants would use words related to blood, pain or biological processes. As their conditions progressed and patients neared hospitalization, they would use more punctuation and negative emotional words in their Facebook posts, according to the study.

Future use of social media data in psychiatric care

Other organizations are also turning to artificial intelligence to monitor mental health. Researchers at Brigham and Women's Hospital are using AI technology from startup Rose to monitor the mental well-being of front-line workers during the COVID-19 pandemic. Meanwhile, the Feinstein Institutes recently developed an AI tool that can help patients get better sleep in the hospital.

Researchers see a use for social media data for patients that could be similar to the vital data they pull from a blood or urine sample, according to Birnbaum. “I could imagine a world where people go see their psychiatrists and provide their archives in the same way they provide a blood test, which is then analyzed much like a blood test and is used to inform clinical decision-making moving forward,” he said.

RELATED: The unexpected ways AI is impacting the delivery of care, including for COVID-19

“I think that is where psychiatry is heading, and social media will play a component of a much larger, broader digital representation of behavioral health.”

Guillermo Cecchi, principal research staff member, computational psychiatry, at IBM Research, also sees a use for social media data as a common way to evaluate patients.

“Our vision is that this type of technology could one day be used in a non-burdensome way, with patient consent and high privacy standards, to provide clinicians with the most comprehensive and relevant information to make treatment decisions, including regular clinical assessments, biomarkers and a patient’s medical history,” Cecchi told Fierce Healthcare.

Researchers hope that the Facebook data can inform future studies.

“Ultimately, the language markers we identified with AI in this study could be used to inform future work, shaped with rigorous ethical frameworks, that could help clinicians to monitor the progression of mental health patients considered at-risk for relapse or undergoing treatment,” Cecchi said.

Gaton said he would like to see the technology get more accurate. “I just hope that with my contributions to the study, the technology gets more accurate and more responsive and can be something that doctors can use in the near future—with patient consent, of course,” he said.