A recent study found that analyzing social media activity can help predict whether someone will continue with their treatment for substance use disorder. The study discovered that using artificial intelligence (AI) to analyze Facebook posts from individuals with substance use disorder was more effective at predicting treatment dropout than standard psychological tests. Those identified as high-risk by the AI model were more likely to drop out of treatment, while those labeled as low-risk tended to stay in treatment. The study was published in Neuropsychopharmacology.
Substance use disorder is a chronic condition characterized by excessive and compulsive drug or alcohol use, leading to significant impairment in various aspects of life. Treatment typically involves a combination of therapies tailored to the individual’s specific needs, such as counseling, support groups, and medication. However, the existing treatment procedures are now always effective. Resuming substance use (relapse) after treatment and abandoning treatment are common.
To improve treatment effectiveness, researchers are exploring ways to identify individuals who are more likely to complete treatment successfully and avoid relapse. One avenue of research focuses on analyzing social media activities of individuals with substance use disorder, as it may provide valuable insights.
However, identifying such individuals is not easy. Current mainstream methods are based on self-reports and not very good at accurately identifying and evaluating factors relevant for predicting how individuals will behave and react to treatment. This inability to capture the intricacies of daily life of each individual that are relevant for the development of substance use disorder are referred to as the ecological information gap.
One possible way to overcome the ecological information gap might be to rely on the analysis of social media activities of individuals suffering from substance use disorder, rather than on self-report assessments in clinical settings. However, unlike self-report assessments social media activity is difficult to analyze.
Study author Brenda Curtis and her colleagues wanted to explore whether AI produced digital phenotype, quantitative characterizations of an individual’s digital behavior could be used to predict the outcome of treatments for substance use disorder. They used an AI tool capable of processing textual information and natural language on Facebook posts from a group of 504 patients attending outpatient substance use disorder treatment in Philadelphia.
Of these 503 individuals, posts of 269 had sufficient language data (at least 200 words) for the AI tool to create a stable prediction of their behavior. The predictions – digital phenotypes were developed based on over 2 years of Facebook posts before the start of treatment. The AI tool analyzed status posts, “longer pieces of text users post on their Facebook timeline”, and link posts, links shared by participants together with free form text that is usually shorter than status posts.
Out of these patients, the AI tool could create stable predictions for the behavior of 269 individuals whose posts contained enough language data (at least 200 words). The predictions (digital phenotypes) were developed using over two years of Facebook posts prior to treatment. The AI tool analyzed both status posts (longer text posts on users’ timelines) and link posts (links shared with accompanying text, usually shorter than status posts).
The participants also completed assessments of addiction symptom severity (the Abridged Addiction Severity Index) and treatment outcomes (with categories – remained abstinent, relapse, dropout or, alternatively, remained abstinent, relapsed but continued treatment, relapsed and dropped out of treatment and dropped out).
The results showed that the predictions of treatment outcomes based on digital phenotypes were more accurate than those based on the standard assessment of addiction symptom severity. This difference was particularly noticeable for negative outcomes like relapse and treatment dropout. Combining digital phenotypes with addiction symptom severity assessments did not significantly improve predictions compared to using digital phenotypes alone.
“The digital phenotype extracted from language on Facebook combined with standard intake data, showed strong validity as a risk assessment tool for substance use disorder treatment dropout,” the study authors concluded. “We were able to show such data can capture many distinctions between treatment outcomes. In fact, while we were able to predict all of the three classes better with the digital phenotype than the ASI [Addiction Severity Index], relapse was significantly harder to predict and we found that dividing it into those that relapsed but remained in treatment versus those that relapsed and dropped out, resulted in greater accuracy, suggesting a potential lack of utility of focusing on relapse as a whole.”
The study makes a valuable contribution to developing novel ways to predict substance use disorder treatment outcomes. However, it should be noted that it relied solely on Facebook posts and all participants suffered from the substance use disorder. Results with data from other social media and participants with different types of disorders might not yield equal results.
The study, “AI-based analysis of social media language predicts addiction treatment dropout at 90 days“, was authored by Brenda Curtis, Salvatore Giorgi, Lyle Ungar, Huy Vu, David Yaden, Tingting Liu, Kenna Yadeta, and H. Andrew Schwartz.
