Objective  To establish a diagnostic model for gastric precancerous lesions (GPL) based on multiple hematological indicators, and to explore its clinical diagnostic efficiency.

Methods  A total of 1 142 subjects diagnosed as GPL by upper gastrointestinal cancer screening were enrolled, and 1 222 healthy subjects or patients with superficial gastritis were enrolled as controls. All the subjects were divided into training group (n=1 655, including 779 GPL cases and 876 controls) and validation group (n=709, including 363 GPL cases and 346 controls) at a ratio of 7 to 3 by Rand random function. Clinical characteristics and laboratory parameters were compared between GLP patients and controls in the training group; the receiver operator characteristic curve (ROC) was drawn to obtain the diagnostic efficacy and optimal threshold valuesfor GPL related risk factors, and the diagnostic model was established. The Hosmer-Lemeshow test was used to evaluate the fit of the model, and internal validation was performed through the validation group; the discrimination of the diagnostic model was evaluated by area under the curve (AUC).

Results  In the training group, patients with gender of male, a history of smoking, a history of drinking and a history of H.pylori infection were more likely to develop GPL, while the controls were faster in eating, had more digestive symptoms and history of stomach disease (P < 0.05). The optimal cut-off values of monocyte count, standard deviation of red blood cell distribution width (RDW-SD), monocyte-to-lymphocyte ratio (MLR), and serum gastrin-17 (G-17) were 0.34×10⁹/L, 46.55 fL, 0.23, 3.98 pmol/L, respectively in the GPL group, which were higher than those of the control group (P < 0.05). The optimal cut-off value of pepsinogen Ⅰ-to-pepsinogen Ⅱ ratio (PGR) was 11.80 in the GPL group, which was significantly lower than that in the control group (P < 0.05). Multivariate regression analysis showed that a history of H.pylori infection, RDW-SD>46.55 fL, PGR < 11.80 and G-17>3.98 pmol/L were independent risk factors for GPL (P < 0.001). In the training cohort, the diagnostic model had an AUC of 0.762 (95%CI, 0.739 to 0.785; P < 0.001), sensitivity of 72.2%, and specificity of 68.5%. Validation of this model using the validation cohort showed that the model had an AUC of 0.719 (95%CI, 0.681 to 0.756; P < 0.001), sensitivity of 81.3%, and specificity of 52.6%, suggesting that the model had good discriminatory ability. The Hosmer-Lemeshow test indicated that the model had a good fitting.

Conclusion  The diagnostic model for GPL established based on RDW, PGR, G-17 and a history of H.pylori infection has good discrimination and calibration, and is a useful tool for the early identification of GPL patients.