S389 Development and Validation of a Clinical Tool to Identify Patients Who Do Not Need Biopsies When Eosinophilic Esophagitis Is Suspected

Abstract

Introduction: There is a low threshold for esophageal biopsy in all patients with esophageal symptoms to assess for eosinophilic esophagitis (EoE). As predictive models of EoE may not fully rule-in EoE, we aimed to develop a reverse model that reliably predicts against a diagnosis of EoE to eliminate unnecessary esophageal biopsies. Methods: In this two-center study, a predictive model was developed (Mayo Clinic [Mayo]) and then validated (University of North Carolina [UNC]). At both centers, cross-sectional data from consecutive adult patients without prior EoE who underwent index endoscopy (EGD) with esophageal biopsies were used. EoE cases were diagnosed per consensus guidelines; controls did not meet these guidelines. A priori, we sought to have at least a 1:2 ratio of EoE:non-EoE. Data were collected on patient demographics, clinical characteristics, and endoscopic findings. Multiple variable logistic regression was used to identify associations with non-EoE status using backward selection method to identify a parsimonious predictive model, while maintaining a specificity ≥95%. Results: The Mayo and UNC cohorts consisted of 345 (EoE=94, non-EoE=251) and 297 patients (EOE=84, non-EoE=213), respectively. In the Mayo cohort, non-EoE compared to EoE patients were significantly more likely to be older and women, and less likely to report solid food dysphagia, prior food impactions, family history of EoE, or have esophageal rings, strictures, edema, furrows, or exudates at index endoscopy. The resulting multiple variable model (Table, Figure) was predictive against a diagnosis of EoE (c-statistic=0.92, 95% CI:0.88-0.96). At a probability of ≥0.91 with this model, the specificity was 95.0% (89/94), meaning that choosing to biopsy patients with a model predicted probability of < 0.91 resulted in biopsying 89 of the 94 patients with EoE. This resulted in a sensitivity of 64.5% (162/251), meaning that choosing not to biopsy patients with a model predicted probability of ≥0.91 results in correctly not biopsying 162 of 251 patients that did not have EoE. The model was validated using the UNC cohort (c-statistic=0.87, 95% CI:0.82-0.92). At a probability of ≥0.91 with this model, the specificity was 93.8% and sensitivity 57.4%. Conclusion: This reverse model accurately identifies the large group of patients with a low likelihood of EoE where unnecessary biopsies can be avoided, potentially resulting in cost and time savings, and lower risk.Figure 1.: ROC curves for the logistic regression model to predict against a diagnosis of eosinophilic esophagitis based on the Mayo and UNC cohorts. Table 1. - Multiple variable logistic regression model to predict against a diagnosis of eosinophilic esophagitis Mayo Cohort UNC Cohort Odds Ratio (95% CI) p-value Odds Ratio (95% CI) p-value Age at index EGD, per 5 years 1.31 (1.17-1.47) p< 0.01 1.25 (1.08-1.47) p< 0.01 Sex Male Female 1.0 (reference)2.63 (1.30-5.36) p=0.01 1.0 (reference)1.0 (0.64-3.41) p=0.38 Presence of any atopic disorder/food allergy No Yes 3.88 (1.83-8.21)1.0 (reference) p< 0.01 1.67 (0.59-4.24)1.0 (reference) p=0.27 Presence of dysphagia No Solid foods only Solids+liquids/unspecified 4.03 (1.40-11.54)1.0 (reference)4.16 (1.44-12.02) p=0.01p=0.01 2.95 (0.40-60.8)1.0 (reference)0.99 (0.41-2.32) p=0.35p=0.98 History of food impaction No Yes 7.47 (1.88-29.68)1.0 (reference) p< 0.01 6.23 (2.65-15.4)1.0 (reference) p< 0.01 Presence of rings, strictures, edema, furrows, or exudates on index EGD No Yes 16.28 (7.76-34.18)1.0 (reference) p< 0.01 10.6 (3.97-33.2)1.0 (reference) p< 0.01