IDENTIFYING MIGRAINE IN PRIMARY CARE WORKFLOWS: PREVALENCE, RISK SIGNALS, AND PREDICTIVE TOOLS
Author(s)
Taylor P. Ikner, BA, MS, Richard W. Hass, PhD.
College of Population Health, Thomas Jefferson University, Philadelphia, PA, USA.
College of Population Health, Thomas Jefferson University, Philadelphia, PA, USA.
OBJECTIVES: Migraine is a leading neurological disability; however, it is often underdiagnosed in primary care, leading to delays in treatment and an increased burden on patients. This study aimed to identify demographic, clinical, and social risk factors, and to evaluate predictive models for earlier detection in primary care.
METHODS: We conducted a cross-sectional analysis of de-identified EMR data from 59,088 adult outpatient visits across Jefferson Health clinics. Predictors included demographics, comorbidities, vital signs, clinic location, and health-related social needs. Two models, an L1-regularized logistic regression and a Random Forest model, were trained with a 70/30 split and were evaluated with ROC-AUC, Precision-Recall AUC, calibration, Brier Score, and threshold-based diagnostic metrics.
RESULTS: Overall migraine prevalence was low (2.53%, n = 1,495) and higher among younger adults and females. Prevalence increased slightly (3.17%) among patients with one or more social needs. Psychiatric and pain-related conditions were strongly linked to migraine and ranked among the top predictive features along with age and sex. Both models displayed consistent but modest discrimination (ROC-AUC: 0.705-0.745) and low performance on precision-recall metrics (PR-AUC: 0.055-0.101), particularly due to class imbalance. Logistic regression demonstrated smoother calibration, and more reliable probability estimates than Random Forest. At a typical clinical threshold (0.05), specificity and negative predictive value stayed high (0.997 and 0.97, respectively), but sensitivity was low (<0.03).
CONCLUSIONS: Results indicate that the models are most appropriate for conservative screening rather than for diagnosis. Including clinical and social factors slightly enhances risk stratification, but identifying positives remains challenging due to low prevalence. These findings indicate that EMR-based prediction tools may identify at-risk patients who require follow-up. To improve performance, more detailed clinical data and class imbalance strategies are needed.
METHODS: We conducted a cross-sectional analysis of de-identified EMR data from 59,088 adult outpatient visits across Jefferson Health clinics. Predictors included demographics, comorbidities, vital signs, clinic location, and health-related social needs. Two models, an L1-regularized logistic regression and a Random Forest model, were trained with a 70/30 split and were evaluated with ROC-AUC, Precision-Recall AUC, calibration, Brier Score, and threshold-based diagnostic metrics.
RESULTS: Overall migraine prevalence was low (2.53%, n = 1,495) and higher among younger adults and females. Prevalence increased slightly (3.17%) among patients with one or more social needs. Psychiatric and pain-related conditions were strongly linked to migraine and ranked among the top predictive features along with age and sex. Both models displayed consistent but modest discrimination (ROC-AUC: 0.705-0.745) and low performance on precision-recall metrics (PR-AUC: 0.055-0.101), particularly due to class imbalance. Logistic regression demonstrated smoother calibration, and more reliable probability estimates than Random Forest. At a typical clinical threshold (0.05), specificity and negative predictive value stayed high (0.997 and 0.97, respectively), but sensitivity was low (<0.03).
CONCLUSIONS: Results indicate that the models are most appropriate for conservative screening rather than for diagnosis. Including clinical and social factors slightly enhances risk stratification, but identifying positives remains challenging due to low prevalence. These findings indicate that EMR-based prediction tools may identify at-risk patients who require follow-up. To improve performance, more detailed clinical data and class imbalance strategies are needed.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
RWD51
Topic
Real World Data & Information Systems
Disease
SDC: Neurological Disorders