Machine Learning-based Development and Validation of a Scoring System for Screening High-Risk Esophageal Varices.

BACKGROUND & AIMS: Many patients with cirrhosis who undergo esophagogastroduodenoscopy (EGD) screening for esophageal varices (EVs) are found to have no or only small EVs. Endoscopic screening for EVs is therefore a potentially deferrable procedure that increases patient risk and healthcare cost. We developed and validated a scoring system, based on readily-available data, to reliably identify patients with EVs that need treatment.

METHODS

We collected data from 238 patients with cirrhosis undergoing screening EGD from January 2016 through December 2017 at 3 separate hospitals in Los Angeles (training cohort). We abstracted data on patient sex, age, race/ethnicity, platelet counts, and levels of hemoglobin, serum sodium, aspartate aminotransferase, alanine aminotransferase, total bilirubin, international normalized ratio, albumin, urea nitrogen, and creatinine. We also included etiology of cirrhosis, presence of ascites, and presence of hepatic encephalopathy. We used a random forest algorithm to identify factors significantly associated with the presence of EVs and varices needing treatment (VNT) and calculated area under the receiver operating characteristic curve (AUROC). We called the resulting formula the EVendo score. We tested the accuracy of EVendo in a prospective study of 109 patients undergoing screening EGDs at the same medical centers from January 2018 through December 2018 (validation cohort).

RESULTS

We developed an algorithm that identified patients with EVs and VNT based on international normalized ratio, level of aspartate aminotransferase, platelet counts, urea nitrogen, hemoglobin, and presence of ascites. The EVendo score identified patients with EVs in the training set with an AUROC of 0.84, patients with EVs in the validation set with and AUROC of 0.82, and EVs in patients with cirrhosis Child-Turcotte-Pugh class A (n = 235) with an AUROC of 0.81. The score identified patients with VNT in the training set with an AUROC of 0.74, VNT in the validation set with and AUROC of 0.75, and VNT in patients with cirrhosis Child-Turcotte-Pugh class A with and AUROC of 0.75. An EVendo score below 3.90 would have spared 30.5% patients from EGDs, missing only 2.8% of VNT. The same cutoff would have spared 40.0% of patients with Child-Turcotte-Pugh class A cirrhosis from EGDs, missing only 1.1% of VNT.

CONCLUSIONS

We algorithmically developed a formula, called the EVendo score, that can be used to predict EVs and VNT based on readily available data in patients with cirrhosis. This score could help patients at low risk for VNT avoid unnecessary EGDs.