A Multi-head Attention Transformer Framework for Oesophageal Cancer Survival Prediction

Oesophageal cancer is one of the malignant tumours with poor prognosis and accurate survival prediction is important to guide clinical treatment decisions. However, existing prediction methods are difficult to effectively capture the complex interactions among clinical features and lack sufficient interpretability. To address these issues, we propose a survival prediction model for oesophageal cancer based on multi-head attention Transformer, which achieves adaptive capture and efficient representation of interactions between features through a 4-head attention mechanism and a 64-dimensional feature embedding space. Experiments on the TCGA-ESCA dataset (n=286) show that the model achieves 84.6%, 89.1% and 83.7% in the key metrics of accuracy, AUC-ROC and F1 score respectively, which are significantly better than existing methods. Characteristic importance analysis revealed that the age at first diagnosis, number of lymph nodes examined, and KPS score were the key factors affecting prognosis, providing an interpretable basis for clinical decision-making.