Background: Colposcopic examination relies heavily on the subjective experience of practitioners, leading to potential misdiagnosis of Cervical Lesions including inflammation, Low-Grade Squamous Intraepithelial Lesion (LSIL), High-Grade Squamous Intraepithelial Lesion (HSIL), cancer (Ca), and adenocarcinoma in situ (AIS). The lack of precise predictive models for these pathological outcomes underscores the need for advanced diagnostic tools. Method: This retrospective study analyzed data from 1605 patients in the colposcopy database at the Obstetrics and Gynecologic Hospital of Fudan University. Among them, 1388 were randomly divided into a training set（n=971） or internal validation group (test set, n=417), with 217 patients used for external validation. The XGBoost (Extreme Gradient Boosting, a type of machine learning model) model was utilized for optimization using 5-fold cross-validation and grid search. It incorporated variables like Liquid-based Cytology Test (LCT), human papilloma virus (HPV) status, age, and colposcopic image features to predict cervical lesions categorized as normal/inflammation, LSIL, HSIL, cancer, or AIS. Model predictions are accessible in a GitHub repository, and performance was assessed based on discriminative ability, calibration, and decision curves. Results: 1、We optimized Area Under the Curve (AUC) through L1-regularized logistic regression, selecting 35 significant variables with a final C-value of 1.101.The top 5 most significant variables identified were ’Atypical/fragile vessels, ulcers, or lump’, ’Acetate white epithelium’, ’LCT’, ’Prominent, robust columnar epithelial villi’, and ’Comprehensive HPV Type’. 2、We established a robust machine learning model-XGBoost model for cervical lesion prediction, achieving an AUC of 0.959 (0.950-0.967) on the training set, 0.930 (0.914-0.946) on the test set, and 0.982 (0.974-0.990) on the external validation set, exhibiting satisfactory calibration on the training dataset and strong performance on both the test and external validation sets. 3、The model exhibits strong performance across various classes, with high AUC values observed in training，test and external validation set: Normal (0.975/0.953/0.980), LSIL (0.928/0.901/0.964), HSIL (0.945/0.898/0.993), Ca (0.986/0.974/0.993), and AIS (0.960/0.925/0.982). The corresponding sensitivities were 0.763, 0.789, and 0.945; specificities were 0.858, 0.871, and 0.929; and accuracies were 0.839, 0.854, and 0.932 across the respective datasets: training set, testing set and external validation set. Conclusion: A novel machine learning-based predictive model was developed and validated, incorporating multiple clinically relevant variables to identify 5 different cervical lesions during colposcopic examination. This model holds significant potential for enhancing clinical guidance.
