The researchers used a form of AI called machine learning to analyze more than 80 baseline factors - ranging from psychological and sociodemographic to health and lifestyle variables - for 126 anonymized individuals diagnosed with GAD. The data came from the U.S. National Institutes of Health’s longitudinal study called Midlife in the United States, which samples health data from continental U.S. residents aged 25 to 74 who were first interviewed in 1995-96. The machine learning models identified 11 variables that appear most important for predicting recovery and nonrecovery, with up to 72% accuracy, at the end of a nine-year period. The researchers published their findings in the March issue of the Journal of Anxiety Disorders.
"Prior research has shown a very high relapse rate in GAD, and there’s also limited accuracy in clinician judgment in predicting long-term outcomes," said Candice Basterfield, lead study author and doctoral candidate at Penn State. "This research suggests that machine learning models show good accuracy, sensitivity and specificity in predicting who will and won’t recover from GAD. These predictors of recovery could be really important for helping to create evidence-based, personalized treatments for long-term recovery."
The researchers ran the baseline variables through two machine learning models: a linear regression model that examines the relationship between two variables and plots data points along a nearly straight line, and a nonlinear model that branches out like a tree, splitting and adding new trees and plotting how it self-corrects prior errors. The models identified the 11 variables key to predicting recovery or nonrecovery over the nine-year period, with the linear model outperforming the nonlinear model. The models also identified how important each variable was compared to the others for predicting recovery outcomes.
The researchers found that higher education level, older age, more friend support, higher waist-to-hip-ratio and higher positive affect, or feeling more cheerful, were most important to recovery, in that order. Meanwhile, depressed affect, daily discrimination, greater number of sessions with a mental health professional in the past 12 months and greater number of visits to medical doctors in the past 12 months proved most important to predicting nonrecovery. The researchers validated the model findings by comparing the machine learning predictions to the MIDUS data, finding that the predicted recovery variables tracked with the 95 participants who showed no GAD symptoms at the end of the nine-year period.
The findings suggest that clinicians can use AI to identify these variables and personalize treatment for GAD patients - especially those with compounding diagnoses, according to the researchers.
Nearly 50% to 60% of people with GAD have comorbid depression, said Michelle Newman, senior author and professor of psychology at Penn State. She explained that personalized treatments could target that depression as well as treat anxiety.
"Machine learning not only looks at the individual predictors but helps us understand both the weight of those predictors - how important they are to recovery or nonrecovery - and the way those predictors interact with one another, which is beyond anything a human might be able to predict," Newman said.
The researchers noted that the study could not determine the duration of GAD over the nine-year period, as it’s a chronic condition and periods where symptoms manifest strongly come and go. The work, however, lays the groundwork for more tailored treatments, they said.
"This work helps us begin to understand more ways in which treatment could be personalized for specific individuals," Newman said.
The U.S. National Institutes of Health, through the National Institute of Mental Health, supported this research.
Basterfield C, Newman MG.
Development of a machine learning-based multivariable prediction model for the naturalistic course of generalized anxiety disorder.
J Anxiety Disord. 2025 Mar;110:102978. doi: 10.1016/j.janxdis.2025.102978