Combining Statistical Modeling and Machine Learning for Prognostic Feature Selection in Cervical Cancer: A Retrospective Study Based on SEER 2004 to 2015 Data.

OBJECTIVES: Cervical cancer is a leading female malignancy with high global morbidity/mortality, and remains high recurrence risk after standard treatment. Accurate prognostic feature identification is critical for personalized therapy and patient survival improvement, while traditional indicators and single biomarkers lack sufficient accuracy in prognostic prediction. METHODS: In this population-based retrospective study, we analyzed 5392 patients with cervical squamous cell carcinoma from the surveillance, epidemiology, and end results (SEER) database between 2004 and 2015. Multivariable logistic regression and machine learning models were used to evaluate the associations between sociodemographic and clinical variables and overall survival. RESULTS: The analysis identified that marital status, median household income, tumor grade (Grade Recode 2017), disease stage, and tumor size were collectively related to prognosis. Patients with lower socioeconomic status and residing in nonmetropolitan areas were found to have worse survival outcomes. A predictive model incorporating these variables, along with age and race, demonstrated acceptable discriminatory performance with an area under the ROC curve (AUC) of 0.70 in prognostic prediction. CONCLUSIONS: These findings suggested that both clinical and sociodemographic factors contributed meaningfully to prognosis in cervical squamous cell carcinoma, providing a basis for integrating these 2 types of factors in the development of stratified risk prediction tools that improve survival and quality of life.