Disparities in the Evaluation and Treatment of Pediatric Migraine in the Emergency Department Using a Language-learning Model
Danielle Kellier1, Marissa Anto2, Mary Regina Boland5, Craig Press2, Naomi Hughes3, Svetlana Ostapenko4, John Farrar1, Christina Szperka2
1Department of Biostatistics, Epidemiology, and Informatics, University of Pennsylvania, 2Department of Neurology, 3Department of Emergency Medicine, 4Department of Biomedical and Health Informatics, Children's Hospital of Philadelphia, 5Saint Vincent College
Objective:
To examine racial/ethnic disparities in migraine diagnosis and management in the pediatric emergency department (ED) using a natural language processing model to identify migraine independent of billing diagnosis.
Background:
Children of color (CoC) received fewer migraine diagnoses in the ED leading to selection bias in studies of patients with migraine diagnoses. Less biased identification may help with health equity research.
Design/Methods:
We
developed a language-learning model to identify migraine in headache-related visits at one pediatric ED (sensitivity of model: 0.85; specificity: 0.69). This model was trained on ED clinical notes for first-time visits of patients ages 5-17 with a headache chief complaint between January 2016 and February 2020 and validated by chart review. We calculated risk ratios for migraine billing diagnosis, testing, and treatment for non-Hispanic Black (NHB) and Hispanic/Latino (HL) patients compared to non-Hispanic White (NHW) patients, adjusting for demographics and medical comorbidities.
Results:
Across 6,718 headache visits, our model identified 3,567 visits (41.3% male) for migraine although only 29.8% of those received a migraine billing diagnosis. Among those identified by the model as having migraine, a lower proportion of CoC received migraine diagnoses (NHB vs. NHW: RR 0.80 [95% CI 0.71-0.92]; HL vs. NHW: 0.72 [0.55-0.95]). A lower proportion of NHB children received any blood tests (NHB: 0.81 [0.72-0.95]; HL: 1.01 [1.00-1.19]), brain MRI scans (NHB: 0.81 [0.64-0.98]; HL: 1.29 [0.92-1.61]), or intravenous medications (NHB: 0.78 [0.72-0.85]; HL: 1.00 [0.94-1.03]).
Conclusions:
CoC presenting with migraine in the ED received fewer migraine diagnoses and lower rates of blood tests, neuroimaging, and intravenous medications compared to NHW children. This disparity in migraine diagnosis aligns with previous research based on headache diagnosis; however, our novel model classifies patients by documented clinical presentation potentially reducing selection bias. Further validation of such automated diagnostic tools will be critical for both more equitable research and clinical care.