Development and validation of an endoscopic diagnostic model for sessile serrated lesions based on machine learning algorithms

Background and aims Sessile serrated lesions (SSLs) are morphologically subtle and often misclassified as hyperplastic polyps (HPs), increasing colorectal cancer risks. We developed a machine learning (ML) model to improve endoscopic SSL diagnosis. Methods Three hundred and eighty-six colorectal polyps (135 SSLs, 251 HPs) with histologically confirmed were retrospective analyzed and divided into a training set and a test set. Multiple ML classification models were applied for a comprehensive analysis. SHapley Additive exPlanations (SHAP) for model contribution were plotted, and the model results were interpreted by calculating the contribution of each feature to the prediction results. Results Comparative analysis revealed that the shrinkage method based on penalisation and post-estimation model fit (R 2 Shrinkage) model demonstrated superior performance in the SSL diagnostic task, with an average accuracy of 84.7% ± 7.7, a specificity of 71.2% ± 15.0, a sensitivity of 92.7% ± 4.1 and F 1 -score of 88.5% ± 6.2. The results revealed that the area under the curve (AUC) values based on both the validation and test sets eventually stabilized at approximately 0.90, indicating the reliable predictive performance of the model. By constructing individualized SHAP plots, we established quantitative diagnostic criteria: when the lesion size was >8 mm, there was a mucus cap, the lesion was located in the right half of the colon, SSL was predicted with a probability of more than 85%; otherwise, HP tended to be diagnosed. Conclusion This study represents the first application of an ML algorithm techniques to the endoscopic classification of serrated polyps. The lesion size, mucus cap and lesion location are key features for the endoscopic diagnosis of SSL.