Performance evaluation of machine learning models in cervical cancer diagnosis: Systematic review and meta-analysis

Cervical cancer is one of the most frequent malignancies worldwide and one of the leading causes of death in women. Recently, artificial intelligence-based tools have been developed for the early diagnosis of malignancies, including cervical cancer, based on data generated in the healthcare area. This systematic review aimed to evaluate the diagnostic performance of machine learning models based on sociodemographic, epidemiologic, and clinical data for the detection of cervical cancer.

A systematic literature search was performed in PubMed, Scopus, and Embase, with no time limit until July 31, 2025. The screening process was performed by two independent authors using the Rayyan® platform, as well as the risk of bias assessment process using the QUADAS-2 tool. A third author resolved cases of dispute. Meta-analysis and sensitivity analysis were performed in Stata17®.

A total of 7 studies were included, in which the machine learning models selected several demographic and clinical variables for the classification of patients. Pooled diagnostic measures of these models were 0.97 (95 % CI 0.90-0.99) for sensitivity, and 0.96 (95 % CI 0.93-0.97) for specificity (I² = 48.63, Q = 35.22, p < 0.01). After sensitivity analysis, diagnostic measurements remained consistent.

Machine learning models are presented as a new tool with high diagnostic performance, making their introduction into cervical cancer screening programs feasible. However, this systematic review discusses some issues that should be addressed to validate and incorporate these models into clinical practice.