910P - Machine learning-based survival prediction model for postoperative parotid mucoepidermoid carcinoma

Background

Parotid mucoepidermoid carcinoma (P-MEC) is a major histopathological subtype of salivary gland cancer, characterized by heterogeneity and complexity. Traditional clinical models and treatment plans are limited in providing personalized treatment recommendations and clinical management. This study aims to develop a machine learning-based prognostic model to improve survival rate prediction accuracy for P-MEC patients.

Methods

We used SEER database data (2004-2015) for 882 postoperative P-MEC patients. Single-factor Cox regression and four machine learning algorithms (Random Forest, LASSO, XGBoost, Best Subset Regression) were employed for variable selection and importance evaluation. The optimal variable selection model was determined using stepwise backward regression, AIC, and AUC. Bootstrap resampling was applied for internal validation.The model's prediction accuracy was assessed through C-index, ROC curve, and calibration curve. The clinical utility of the model was evaluated via DCA.

Results

The 3-, 5-, and 10-year OS rates for postoperative P-MEC patients were 0.887, 0.841, and 0.753. XGBoost, BSR, and LASSO demonstrated good predictive performance in variable selection and identified 7 independent prognostic factors: age, pathological grade, T stage, N stage, radiation therapy, chemotherapy, and marital status. A nomogram was developed based on the identified predictive variables. Internal validation with 1000 bootstrap resamples yielded consistent results for C-index (3-year, 0.841, 95%CI=0.839-0.844; 5-year, 0.842, 95%CI=0.840-0.845) and AUC (3-year, 0.858, 95%CI=0.856-0.860; 5-year, 0.872, 95%CI=0.870-0.875). The calibration plot indicated good prediction fit.

Conclusions

This study identified age, pathological grade, T stage, N stage, radiation therapy, chemotherapy, and marital status as independent prognostic factors in postoperative P-MEC patients. Machine learning algorithms outperformed traditional analysis methods in variable selection, predictive performance, and variable importance visualization. The developed machine learning-based nomogram offers high accuracy, robust calibration, and clinical application value for postoperative P-MEC patients.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.