Introduction

The distinction between eosinophilic esophagitis (EoE) and other causes of dysphagia in clinical practice can be difficult without histology. In addition, failure to collect esophageal biopsies represents one of the reasons for a delayed diagnosis in patients with EoE. Current guidelines recommend multiple esophageal biopsies in all patients reporting dysphagia although this may not be cost-effective.

Aims & Methods

To optimize resources allocation in the diagnostic workup of EoE, we developed and thereafter validated on an independent cohort of patients two point-of-care machine learning (ML) tools to predict a diagnosis of EoE: 1) based on clinical history data alone and 2) based on clinical-endoscopic findings, prior to the results of esophageal biopsies. This was a two-step multi-center international study conducted in three European tertiary referral centers for EoE (London, Pisa, and Padua). We build predictive ML models based on retrospectively extracted clinical data alone (age at diagnosis, sex, presence or absence of heartburn, chest pain, postprandial nausea/dyspepsia, history of food impaction requiring endoscopic removal, allergic rhinitis, asthma, nasal polyposis) and on clinical and upper endoscopy (EGDS) (exudates, rings, edema, furrows, stricture) data collected from 273 EoE and 55 non-EoE dysphagia (NED) patients in London. We performed an external validation of the models on an independent cohort of 93 prospectively enrolled patients reporting dysphagia at Pisa and Padua University Hospitals. The performance of the models was described with area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). We then developed a freely downloadable software for use in clinical practice at the point-of-care.

Results

The machine learning model trained on clinical data alone showed AUC of 0.82, and sensitivity, specificity, PPV, and NPV of 0.75, 0.82, 0.82, 0.75, respectively, for the diagnosis of EoE in the external validation cohort. Among clinical data alone, the patients’ age at diagnosis and sex had the highest predictive ability. The machine learning model trained on a combination of clinical and endoscopic data showed AUC of 0.93, and sensitivity, specificity, PPV, and NPV of 0.88, 0.95, 0.95, 0.87, respectively, in the external validation cohort. Among clinical and endoscopic data, the patients’ age at diagnosis, esophageal rings at endoscopy, and sex had the highest predictive ability. The models’ performance in the development and external validation cohorts is reported in Table 1.

Conclusion

The international external validation of our ML tools, as well as the freely available software, allow generalizability of our findings to other settings. Our software can be easily used at the point-of-care to optimize resources allocation in the diagnostic workup of EoE and 1) Support the clinical suspicion of EoE even in the absence of endoscopic findings, 2) Support decision making in performing repeat EGDS in patients in whom biopsies had been omitted, 3) Support decision making in performing repeat EGDS in patients with persistent EoE symptoms who had non-diagnostic levels of eosinophils at a previous EGDS performed while on proton-pump inhibitors, 4) Avoid unnecessary biopsies in very low-risk patients.