Overview
- Duke Health reported peer-reviewed results in Nature Mental Health showing that AI can estimate a child's future ADHD risk from standard electronic health records.
- The model, trained on records from more than 140,000 children, reached an accuracy score of about 0.92 by age five at a four-year prediction window.
- The system is built for clinicians to use as an early-warning screen to guide referrals, not to make an ADHD diagnosis.
- Accuracy held steady across sex, race, ethnicity, and insurance status in the study groups, suggesting a chance to narrow detection gaps.
- The team pretrained a health-records foundation model and says clinics will need prospective validation and workflow testing before adoption, with support from NIMH and NCATS grants.