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

TL;DR

This study developed machine learning models to predict the risk of high-grade cervical lesions using screening data, aiming to improve cervical cancer screening efficiency.

Contribution

The novel contribution is the development and validation of four machine learning models with varying predictors for predicting cervical lesion risk.

Findings

The models achieved AUCs ranging from 0.83–0.96 in training and test sets.

Model 1 (M1) showed the highest positive predictive value (PPV) across all prediction intervals.

The models demonstrated consistent performance in a separate validation set.

Abstract

This study aimed to improve cervical screening efficiency by developing and validating machine-learning models for predicting high-grade cervical lesions or worse (HCL) risk.

From Swedish nationwide registers, we included 474,072 women invited to cervical screening in 2016 (split into 80% training and 20% test sets) and 370,105 women invited in 2017 for validation. All women underwent index cytology and/or human papillomavirus (HPV) testing within the recommended interval after age 29. Predictors included screening results (cytology and/or HPV testing), other HPV-related factors, and demographic factors (including age). Four random forest models were trained via 5-fold cross-validation with different predictors: Model 1 (M1) (all predictors), M2 (cytology, HPV testing, age), M3 (HPV testing, other HPV-related factors, and demographic factors), and M4 (HPV testing and age). We computed area under the curves (AUCs) and created plots to depict positive predictive value (PPV) by the number of women intervened.

In training and test sets, 1-, 3-, and 5-year HCL incidence proportions were 0.25%, 0.68%, and 1.05%, respectively. Cross-validated AUCs were 0.83–0.96 (M1), 0.83–0.96 (M2), 0.91–0.94 (M3), and 0.91–0.93 (M4), depending on the prediction intervals. Similar AUCs were found in the test set. Additionally, the AUCs in the validation set were 0.85–0.95 (M1), 0.85–0.95 (M2), 0.91–0.94 (M3), and 0.92–0.93 (M4). Across all intervals, M1 consistently demonstrated the highest PPV, followed by M2, M3, and M4. For each model, PPVs were lowest for 1-year predictions but comparable at 3 and 5 years.

The models demonstrated strong predictive performance. Evaluating PPVs over the number of invited women provides the potential for risk-stratified screening and clinical utility.

Vetenskapsrådet, FORTE, Karolinska Institutet, Horizon 2020, and Cancerfonden.

Figures