A data-centric approach to detecting and mitigating demographic bias in pediatric mental health text

Healthcare Artificial Intelligence (AI) offers transformative potential but often inherits biases from training data, worsening disparities. While bias mitigation has focused on structured data, mental health relies on unstructured clinical notes, where linguistic differences and data sparsity pose challenges. This study aims to detect and reduce non-biological textual bias in AI models supporting pediatric mental health screening.

We analyzed ~20,000 pediatric anxiety cases and matched controls (ages 5-15) from Cincinnati Children's Hospital records, where gender prevalence transitions from male-dominant in early childhood to female-dominant in adolescence. Anxiety prediction models were fine-tuned using a Transformer architecture optimized for computational efficiency. Classification parity across sex subgroups was evaluated, and we also verified that the model relied on clinically relevant words (using the LIME tool). Bias was mitigated through informative term filtering and systematic gender-biased text replacement.

Here, we show systematic under-diagnosis of female adolescents, with 4% lower accuracy and 9% higher false-negative rates compared to male patients. Notes for male patients are on average 500 words longer, and linguistic similarity metrics reveal distinct word distributions between sexes. Applying our de-biasing framework reduces diagnostic bias by up to 27%, improving equity in model performance.

We develop and evaluate a data-centric de-biasing framework to address gender-based disparities in clinical text arising from non-biological differences, such as reporting practices and documentation styles. Our method selectively de-biases data by neutralizing biased language and normalizing information density while preserving clinically relevant content. Further validation across different models is essential before clinical deployment.