Genetic algorithm-optimized neural network outperforms TNM staging in predicting rapidly progressive nasopharyngeal
Li-Ting Ling1, Wang-Jian Li1, Yue Yao1
1Department of Radiation Oncology, Guangxi Medical University Cancer Hospital, Nanning, China.
Purpose:
To establish machine learning-based predictive models for rapidly progressive nasopharyngeal carcinoma (RP-NPC), defined as disease progression within 24 months post-initial treatment, and to assess differential survival benefits of adjuvant chemotherapy (AC) following concurrent chemoradiotherapy (CCRT) in RP-NPC versus Non-RP-NPC subgroups.
Methods:
This retrospective cohort study analyzed 716 NPC patients (2007-2012). Five machine learning models were constructed using independent risk factors, including: Genetic algorithm-optimized neural network (GNN), Standard artificial neural networks (ANN and BPNN), eXtreme Gradient Boosting (XGBoost), and Logistic regression (LR). Predictive performance was rigorously evaluated using ROC curve analysis (quantified by area under the curve, AUC) for discrimination and calibration plots for reliability estimation, with both internal validation (bootstrap resampling) and external validation procedures. Stratified survival analysis was performed using Cox proportional hazards models to compare CCRT-AC versus CCRT alone in both machine learning-predicted and clinically defined RP-NPC and Non-RP-NPC subgroups.
Results:
Five independent predictors emerged for rapid disease progression (T/N stage, age, alkaline phosphatase, lactate dehydrogenase). The AUC value of the GNN, ANN, BPNN, XGBoost, LR model, and TNM stage in predicting RP-NPC was 0.777 vs 0.792 vs 0.774 vs 0.841 vs 0.735 vs 0.688 (training cohort), and 0.782 vs 0.734 vs 0.606 vs 0.698 vs 0.711 vs 0.687 (validation cohort), respectively. After propensity score matching, AC demonstrated no survival benefit for patients with RP-NPC, regardless of whether they were identified by the GNN model or clinically defined criteria.
Conclusion:
The GNN demonstrated superior predictive capability and enhanced generalizability over conventional TNM staging for identifying RP-NPC. Critically, RP-NPC patients derived no survival benefit from AC supplementation after CCRT.
