Using Machine Learning to Predict Progression in the Gastric Precancerous Process in a Population from a Developing Country Who Underwent a Gastroscopy for Dyspeptic Symptoms

Table 2

Predictors identified based on the best subset regression with the leave-one-out method and model validation using leave-one-out cross-validation.

Outcome	Best subset model with the leave-one-out method		Validation of model identified by best subset with the leave-one-out method in the 20% test sample
	Identified predictors		Sensitivity (%) ()	Specificity (%) ()
Overall progression (general model)	Complete intestinal metaplasia	41.0	86 (19/22)	79 (31/39)
	Incomplete intestinal metaplasia
	Histological diagnosis at baseline less advanced than atrophic gastritis
	Depth of corpus inflammation at baseline
	Average density of polymorphonuclear cells in the antrum at baseline
	Alcohol intake at baseline
Overall progression (location-specific model)	Complete intestinal metaplasia	43.1	100 (21/21)	82.1 (32/39)
	Incomplete intestinal metaplasia
	Histological diagnosis at baseline less advanced than atrophic gastritis
	Average density of H. pylori infection in the corpus and the antrum
	Depth of corpus inflammation at baseline
	Intake of fried fava beans

Sensitivity and specificity were identical when eradication of H. pylori infection was forced into the model.