Development of machine learning models for survival prediction in nasopharyngeal carcinoma using population-based data.

backgroundDespite significant improvements in treatment efficacy, nasopharyngeal carcinoma (NPC) remains one of the most common and threatening head and neck cancers. This study sought to build a risk stratification model that categorizes NPC patients into distinct prognostic groups based on readily available demographic and clinical variables, with the goal of enhancing patient care and improving survival rates.

methodsData for patients diagnosed with NPC between 2010 and 2018 were obtained from the SEER program. Kaplan-Meier analysis and Cox proportional hazards regression were applied to evaluate the effects of treatment modalities on overall survival (OS) and cancer-specific survival (CSS). Prognostic variables from multivariable Cox models were used to construct nomograms for predicting 1-, 3-, and 5-year OS and CSS. Several survival-focused machine learning algorithms, including the random survival forest (RSF), were trained and compared using the concordance index (C-index) and integrated Brier score. An interactive web-based calculator was developed based on the best-performing model.

resultsThe analysis included 9,816 patients diagnosed between 2000 and 2020. Combined radiotherapy and chemotherapy significantly improved survival in locoregionally advanced and advanced NPC compared with other treatment regimens. Cox-based nomograms achieved a C-index of 0.71 (95% CI: 0.70-0.72), outperforming the conventional staging system. The RSF model demonstrated the highest prognostic accuracy, with C-indices of 0.73 (95% CI: 0.71-0.74) for 3-year OS and 0.75 (95% CI: 0.73-0.76) for 5-year OS. The RSF algorithm was implemented as a free online survival prediction tool.

conclusionMachine learning based models, particularly the RSF algorithm, provide improved individualized survival predictions for NPC. While performance was promising, further external validation and inclusion of additional clinical and molecular predictors are needed before clinical adoption. We also developed an interactive web-based survival prediction tool, enabling clinicians and researchers to obtain individualized survival probabilities in real time.