Early attention deficit hyperactivity disorder prediction from longitudinal electronic health records

Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental condition that can negatively impact long-term outcomes for individuals. Early diagnosis is critical, yet demographic and clinical disparities can delay detection. Using electronic health records (EHR) from a cohort of over 720,000 patients, we pretrained an EHR foundation model. We then fine-tuned it to predict the likelihood of ADHD diagnosis and timing from birth until age nine years in a pediatric cohort of over 140,000 patients. By age five years, the model achieved a time-dependent area under the receiver operating characteristic curve of 0.92 at a 4-year time horizon. Overall, the model maintained its performance across patients with differing demographics, including sex, race, ethnicity, and insurance status. Our feature importance analysis found that ADHD was strongly associated with developmental, behavioral, and psychiatric conditions. Our results suggest that EHR-based predictive models could help providers reliably identify children with ADHD in a timely manner.