Predicting high-grade lesions and cervical cancer with supervised machine learning algorithms.

OBJECTIVE

To evaluate machine learning models for predicting high-grade squamous intra-epithelial lesions and cervical intra-epithelial neoplasia grade 2 or higher using sociodemographic and clinical parameters.

METHODS

A retrospective cohort study was conducted using data from the Subred Norte (Bogotá, Colombia), a public health network, between 2018 and 2024. Records lacking sociodemographic data were excluded. To address the 10% class imbalance, random under-sampling was applied to train balanced models, which were then compared with the original imbalanced models.

RESULTS

The final sample included 4367 patients (aged 21 to 65 years) with cytology, colposcopy, or biopsy reports. The real-world prevalence of histopathologically confirmed cervical intra-epithelial neoplasia grade 2 or higher in this cohort was 10%. Five machine learning classification methods were tested across 2 datasets: 2018-2022 (n = 3139, excluding human papillomavirus history due to under-reporting) and 2023-2024 (n = 1228, including human papillomavirus history). On the original imbalanced dataset (without human papillomavirus history), the random forest model achieved an area under the receiver operating characteristic curve of 0.91 and a specificity of 99%, but had a sensitivity of only 41%. After applying random under-sampling, the random forest model's sensitivity improved to 89%, with a precision of 84%. For the dataset including human papillomavirus history, the support vector machine model achieved the highest balanced performance, yielding similar metrics. However, because these models were developed and evaluated strictly within a single institution, these balanced metrics represent optimistic internal validation performance.

CONCLUSIONS

Machine learning models show promising performance for predicting cervical intra-epithelial neoplasia grade 2 or higher, with improved sensitivity after addressing class imbalance. These findings support further external, multi-center validation before clinical implementation.