Development of Electronic Health Record-Based Machine Learning Models to Predict Barrett's Esophagus and Esophageal...

Area of Science:

• Gastroenterology
• Oncology
• Medical Informatics

Background:

• Screening for Barrett's esophagus (BE) is recommended for individuals with risk factors but is underutilized.
• Existing risk prediction tools for BE and esophageal adenocarcinoma (EAC) have modest accuracy (AUROC ≤0.7) and face implementation challenges.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting BE and EAC risk using a large electronic health record (EHR) database.

Main Methods:

• Utilized a deidentified EHR database from 6 million patients.
• Identified BE and EAC cases and controls using ICD codes and natural language processing on clinical notes.
• Employed an ensemble transformer-based ML model architecture.

Main Results:

• The BE ML model achieved 76% sensitivity, 76% specificity, and 0.84 AUROC.
• The EAC ML model achieved 84% sensitivity, 70% specificity, and 0.84 AUROC.
• Identified novel risk predictors including coronary artery disease, serum triglycerides, and electrolytes.

Conclusions:

• ML models trained on EHR data demonstrate superior accuracy in predicting BE and EAC risk compared to traditional methods.
• These models hold potential for enhancing the clinical implementation of minimally invasive screening technologies for BE and EAC.