Comparative evaluation of machine learning models in predicting overall survival for nasopharyngeal carcinoma using...

The objective of this study is to assess the prognostic efficacy of ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) positron emission tomography/computed tomography (PET-CT) parameters in nasopharyngeal carcinoma (NPC) and identify the best machine learning (ML) prognostic model for NPC patients based on these ¹⁸F-FDG PET/CT parameters and clinical variables.

A cohort of 678 patients diagnosed with NPC between 2016 and 2020 was analyzed in this study. The model was constructed using four advanced ML algorithms, namely Random Forest (RF), Extreme Gradient Boosting (XGBoost), Least Absolute Shrinkage and Selection Operator (LASSO), and multifactor COX step-up regression. Statistical significance of the models was assessed using Kaplan-Meier (K-M) curves, with a significance level established at P < 0.05. The prognostic efficacy of the models was evaluated through the analysis of receiver operating characteristic (ROC) curves, with the area under the ROC curve (AUC) serving as a criterion for model selection. The decision curve analysis (DCA) and concordance index (C-index) were employed to assess the precision of the optimal model.

Multivariate analysis revealed age, T stage, and metabolic tumor volume (MTV) for the primary nasopharyngeal tumor (MTVT) as significant independent prognostic factors for overall survival (OS) in NPC patients. Additionally, the LASSO model identified six key variables, including peak standardized uptake value (SUV-peak) for the primary nasopharyngeal tumor (SUV-peak(T)), MTVT, heterogeneity index for neck lymph nodes (HIN), age, pathological type, and T stage. Remarkably, the LASSO model demonstrated superior performance with a 5-year AUC of 0.849 compared to other models. Further assessment using the C-index and DCA confirmed the accuracy of the LASSO model. Subgroup analysis revealed notable risk factors, such as a high heterogeneity index (HI) for the primary nasopharyngeal tumor (HIT), MTV values for neck lymph nodes (MTVN), and HIN.

We developed a novel prognostic machine learning model that integrates ¹⁸F-FDG PET-CT parameters and clinical characteristics, significantly enhancing prognosis prediction in NPC.