Prediction models for high-grade cervical lesions or worse using machine learning.

Area of Science:

• Oncology
• Medical Informatics
• Public Health

Background:

• Cervical cancer screening efficiency can be enhanced using predictive models.
• Machine learning (ML) offers a promising approach for identifying women at high risk of high-grade cervical lesions (HCL).

Purpose of the Study:

• To develop and validate ML models for predicting HCL risk.
• To assess the predictive performance of different ML models using various combinations of predictors.

Main Methods:

• Utilized Swedish nationwide registers with data from 474,072 women (2016) for training and 370,105 women (2017) for validation.
• Trained four random forest models (M1-M4) using predictors including cytology, human papillomavirus (HPV) testing, HPV-related factors, and demographic data.
• Evaluated models using area under the curves (AUCs) and positive predictive values (PPVs) across 1-, 3-, and 5-year prediction intervals.

Main Results:

• Models demonstrated strong predictive performance with cross-validated AUCs ranging from 0.83 to 0.96 and validation AUCs from 0.85 to 0.95.
• Model 1 (M1), incorporating all predictors, consistently showed the highest PPV across all prediction intervals.
• Positive predictive values were lowest for 1-year predictions but comparable for 3- and 5-year predictions across models.

Conclusions:

• Developed ML models exhibit significant potential for improving cervical screening efficiency.
• The models' strong predictive performance supports their utility in risk-stratified screening approaches.
• Evaluating PPVs relative to the number of women screened highlights the clinical applicability of these predictive tools.