Baseline characteristics
The baseline characteristics of patients in the training and validation groups were shown in Table 1. There was no significant difference between these groups in pathologic records (P = 0.514). The incidence with actual neoplastic potential was 19.3% and 21.8% in two cohorts, respectively. This rate partly suggests that there is considerable nonessential surgery based on guidelines for surgical procedures. No significant difference was found in clinical-ultrasonic characteristics(P>0.05).
Risk Factors for Neoplastic GBPs
In the training cohort, the results showed that age, diabetes, cholelithiasis, CEA, CA199, ultrasonic diagnosis, polyp size, number, sessile and clinical symptoms were simultaneously significant predictive clinical and imaging variables for neoplastic polyps (P < 0.05 for all) (Table 2). Then, the multivariable conditional logistic regression analysis identified age, cholelithiasis, CEA, polyp size and sessile as independent factors (Table 3). According to the ROC curve analysis, we determined the optimal cut-off value of age and CEA as 58 years and 1.56 ng/ml. Considering most guidelines divided 10mm as the positive value of polyp diameter, and the ROC-induced cutoff value was 15 mm, we defined the 10mm and 15mm in polyp size as cutoff points for three-way classification.
Development and validation of the prediction nomogram
According to the results of the univariate and multivariate analyses, we developed a nomogram incorporating predictive variables to predict neoplastic risk in patients with GBPs preoperatively.
As shown in Fig. 2A, age (0, ≤ 58 years; 1, >58 years), cholelithiasis (0, negative; 1, positive), CEA (0, ≤1.56 ng/ml; 1, > 1.56 ng/ml), polyp size (0, <10 mm; 1, ≥ 10 mm and ≤ 15 mm; 2, >15 mm) and sessile (0, pedunculated; 1, sessile). The formula of the weighted value was: Y = 1.194 × [Age] +1.177 × [Cholelithiasis] +1.171 × [CEA] +1.112 × [Polyp size] +1.066 × [Sessile] -3.944.
The nomogram achieved the overall accuracy rate of 84.1%, with a sensitivity of 68.1% and a specificity of 88.2%. Among the 30 false negative cases, only 1 case was GBC. We plotted the ROC curve to compare the discrimination ability of our model to the US-reported, three different kinds of guidelines, and comparable score (Korean model) [18]. As shown in Fig.2B and summarized in Table 4, the nomogram model obtained the best discrimination ability with the AUC value of 0.846 (95% CI: 0.779, 0.913) in the training cohort. In the validation cohort, our model also yielded the highest AUC of 0.835 (95% CI: 0.765,0.905) compared with the US-reported alone (AUC: 0.659; 95% CI: 0.603, 0.716; P < 0.001), JSHBPS guideline (AUC: 0.635; 95% CI: 0.569, 0.702; P < 0.001), ESGAR guideline (AUC: 0.617; 95% CI: 0.561, 0.672; P < 0.001), CCBS guideline (AUC: 0.658; 95% CI: 0.598, 0.717; P < 0.001) and Korean model (AUC: 0.746; 95% CI: 0.663, 0.828; P < 0.001).
To evaluate prediction models, we presented the for our model, US-reported, JSHBPS guideline, ESGAR guideline, CCBS guideline and Korean model in Fig. 3. Across the reasonable threshold probability ranges in both training and validation groups, DCA graphically showed that the nomogram provided more clinical benefits in predicting malignancy in patients with GBPs than the other methods.