CINCINNATI — Researchers at Cincinnati Children’s, working with collaborators at University College London and Oak Ridge National Laboratory, have identified a practical, data-centered strategy to reduce bias in artificial intelligence (AI) systems used in children’s mental health care.
The findings, published online March 5, 2026 , in Communications Medicine , address growing concern that AI tools designed to assist clinicians may not perform equally well across patient groups. Previous studies have shown that some mental health models are less accurate for girls than boys. The new international study demonstrates that targeted improvements to training data can significantly reduce those disparities—without diminishing overall model performance.
Detecting Unequal Performance
The research team analyzed nearly 20,000 pediatric anxiety cases from electronic health records. They found that AI models were more likely to miss anxiety in female adolescents. The performance gap was most pronounced during adolescence—a period when anxiety prevalence increases substantially among girls.
“Our findings show that disparities in model performance are not inevitable,” says Julia Ive, PhD, professor at University College London and lead author of the study. “By closely examining how AI systems learn from clinical narratives, we were able to identify patterns in the training data that contributed to this bias and demonstrate concrete steps to address them.”
Why Bias Can Emerge in Mental Health AI
Unlike many medical AI systems that rely on lab values or imaging, mental health tools often analyze unstructured clinical notes to detect early warning signs. The study found that notes written about male patients were, on average, approximately 500 words longer and differed in language patterns and information density compared with notes written about female patients.
These differences were not intentional. However, because AI systems learn directly from documentation patterns, such variation can influence how models interpret symptoms and generate predictions.
“Bias in AI rarely stems from malicious intent—it reflects patterns embedded in the data,” says John Pestian, PhD, MBA, senior author of the study and co-director of the Decode Mental Health Program at Cincinnati Children’s. “In mental health care, where predictive systems rely heavily on the written clinical record, differences in how care is delivered and documented can shape how AI interprets patient information and arrives at its conclusions.”
A Data-Centered Approach
Rather than redesigning the AI system, the team focused on improving the information used to teach it. To do this the researchers used advanced language processing tools to remove less informative text that balanced the clinically meaningful information between girls and boys and replaced gender specific names and pronouns with neutral terms, while maintaining the integrity of the clinical context.
These targeted adjustments reduced diagnostic bias by up to 27%, while maintaining overall accuracy and improving confidence in predictions.
“This study shows that improving fairness does not necessarily require more complex models,” Ive says. “Careful attention to how clinical information is structured and represented can have a measurable impact.”
Clinical Implications
Anxiety disorders are among the most common mental health conditions affecting children and adolescents. They often develop gradually and may present differently across developmental stages and between sexes. Early identification and timely care are critical.
“Adolescence is a time when anxiety becomes particularly prevalent among girls,” says Jeffrey Strawn, MD, a pediatric anxiety specialist and co-author. “If AI tools are less sensitive for that population, we risk delaying recognition and treatment during a crucial developmental window.”
“As AI becomes more integrated into pediatric care, rigorous evaluation for bias is essential,” says Tracy Glauser, MD, co-director of the Decode Mental Health Program at Cincinnati Children’s. “These systems are intended to support clinicians. Ensuring equitable performance across populations is both a scientific responsibility and an ethical one.”
“Progress in AI is often measured in computational power,” Pestian says. “But its lasting impact will be measured in trust. By strengthening the data that guide these systems, we help ensure they support clinicians in ways that are equitable, reliable, and worthy of the families we serve.”
About the Study
Co-authors for this study also included Daniel Santel, PhD, Cincinnati Children’s, and collaborators with the University College London, Queen Mary University of London, Oak Ridge National Laboratory, Georgia Institute of Technology, and the Kumoh National Institute of Technology in South Korea.
This work was funded by Cincinnati Children’s Mental Health Trajectory program.
Communications Medicine
Data/statistical analysis
People
A data-centric approach to detecting and mitigating demographic bias in pediatric mental health text
5-Mar-2026