A New Method Based on Feature Alignment for Multi-Center Machine Learning Survival Prediction in Nasopharyngeal Cancer Patients

Abstract

Purpose In the survival prediction of multi-center nasopharyngeal carcinoma patients, the inconsistency of data characteristics across different centers significantly affects the predictive performance of data-driven machine learning methods. This study explores how to improve the effectiveness and consistency of multi-center data through feature alignment and fusion techniques, thereby optimizing the performance of survival prediction models.

Methods We collected two sets of nasopharyngeal cancer data sets.After feature alignment, LASSO regression was used to screen the features in the aligned latent space to select the most discriminative features. Subsequently, we introduced 15 mainstream machine learning algorithms for survival prediction model training, selecting multidimensional metrics such as AUC, F1-score, Accuracy, and Recall for evaluation.Additionally, to validate the superiority of the method, we applied Pearson correlation and the Boruta algorithm to train the aforementioned 15 machine learning models separately on both datasets and compared the performance results. Combining the encoder weights of an Autoencoder, the SHAP importance is back-projected from the latent space to the original feature variables to analyze the factors influencing the survival of NPC patients.

Results All evaluation metrics (AUC, F1 score, precision) exceeding 0.97.This approach successfully mitigated multi-centre data heterogeneity while validating the robust generalisation capability and stability of these models within multi-centre data environments.SHAP interpretability analysis revealed that the feature combinations of gender + time, age + time, and primary site + age exhibited the broadest SHAP value distributions, representing the most influential feature combinations affecting survival risk prediction.

Conclusion The data heterogeneity issue associated with multi-center NPC survival prediction was effectively resolved by combining autoencoder-based feature alignment with adversarial training.