[Construction and preliminary validation of machine learning predictive models for cervical cancer screening based on...

Area of Science:

• Oncology
• Genetics
• Bioinformatics

Background:

• Cervical cancer screening relies on human papillomavirus (HPV) detection and cytological diagnosis, which have limitations.
• DNA methylation patterns offer potential biomarkers for early cancer detection.

Purpose of the Study:

• To develop and evaluate machine learning (ML) predictive models for cervical cancer and precancerous lesion screening using human gene methylation characteristics.
• To compare the efficacy of ML models against HPV detection and cytological diagnosis.

Main Methods:

• Collected 224 cervical exfoliated cell specimens for DNA methylation analysis.
• Selected hypermethylated gene fragments using LASSO regression.
• Constructed ML models (Random Forest, Naive Bayes, Support Vector Machine) for predicting cervical intraepithelial neoplasia grade 2 (CIN2) or higher.
• Validated models using histological diagnosis as the gold standard.

Main Results:

• Seven hypermethylated gene fragments were identified and used to build RF, NB, and SVM models.
• In the validation set (80 cases), the NB model achieved an Area Under the Curve (AUC) of 0.88, outperforming HPV detection (0.68) and cytology (0.45).
• The NB model showed high specificity (93.88%) for detecting CIN2+ lesions.

Conclusions:

• Machine learning models based on DNA methylation show superior performance compared to current screening methods.
• The Naive Bayes model demonstrates significant potential for screening cervical cancer and precancerous lesions.
• Further validation is warranted for clinical application of DNA methylation-based ML models in cervical cancer screening.