Development and validation of hybrid machine learning approach for predicting survival in patients with cervical cancer: a SEER-based population study.

BACKGROUND

Accurate survival prediction in cervical cancer is crucial for personalized therapy, particularly in high-risk groups where early intervention might enhance results. The study aims to create a hybrid survival model that integrates Cox Proportional Hazards (CoxPH) with Elastic Net regularization and Random Survival Forest (RSF) to improve prediction accuracy and interpretability.

METHODS

Data from the SEER database (2013-2015) were pre-processed through normalization and encoding. RSF recorded non-linear interactions between covariates, while the CoxPH Elastic Net Regularization model provided linear interpretability and identified key variables. Model parameters were optimized using cross-validation, and final performance was assessed on an independent test set using metrics including C-index, Integrated Brier Score (IBS), AUC-ROC, and calibration plots.

RESULTS

The hybrid model outperformed the individual models with an Integrated Brier Score (IBS) of 0.13 and a concordance index (C-index) of 0.82. With an AUC-ROC of 0.84, the model provided robust calibration and classification performance on the independent test set, effectively separating between individuals at high and low risk.

CONCLUSION

The hybrid model provides a promising tool for personalized risk stratification in cervical cancer based on survival probability. Further testing in varied clinical categories is recommended to confirm its efficiency in precision oncology.