Development of a machine learning-based model for predicting positive margins in high-grade squamous intraepithelial lesion (HSIL) treatment by Cold Knife Conization(CKC): a single-center retrospective study.

Objectives: This study aims to analyze factors associated with positive surgical margins following cold knife conization (CKC) in patients with cervical high-grade squamous intraepithelial lesion (HSIL) and to develop a machine-learning-based risk prediction model.

Method: We conducted a retrospective analysis of 3,343 patients who underwent CKC for HSIL at our institution. Logistic regression was employed to examine the relationship between demographic and pathological characteristics and the occurrence of positive surgical margins. Various machine learning methods were then applied to construct and evaluate the performance of the risk prediction model.

Results: The overall rate of positive surgical margins was 12.9%. Independent risk factors identified included glandular involvement (OR = 1.716, 95% CI: 1.345-2.189), transformation zone III (OR = 2.838, 95% CI: 2.258-3.568), HPV16/18 infection (OR = 2.863, 95% CI: 2.247-3.648), multiple HR-HPV infections (OR = 1.930, 95% CI: 1.537-2.425), TCT ≥ ASC-H (OR = 3.251, 95% CI: 2.584-4.091), and lesions covering ≥ 3 quadrants (OR = 3.264, 95% CI: 2.593-4.110). Logistic regression demonstrated the best prediction performance, with an accuracy of 74.7%, sensitivity of 76.7%, specificity of 74.4%, and AUC of 0.826.

Conclusion: Independent risk factors for positive margins after CKC include HPV16/18 infection, multiple HR-HPV infections, glandular involvement, extensive lesion coverage, high TCT grades, and involvement of transformation zone III. The logistic regression model provides a robust and clinically valuable tool for predicting the risk of positive margins, guiding clinical decisions and patient management post-CKC.