A Multi-Stage Multi-Level Attention-Based Feature Learning Framework for Multi-Endpoint Survival Prediction In Head and Neck Cancer
Abstract
Purpose
Accurate patient stratification is critical for personalized radiation therapy in head-and-neck cancer (HNC), yet it remains challenging due to disease heterogeneity, high-dimensional feature spaces, and limited sample sizes for specific clinical endpoints. This study aims to develop and evaluate a novel multi-task, attention-based deep learning framework, DeepCox_MTL, for the joint prediction of multiple survival endpoints, including Overall Survival (OS), Local Failure-Free Survival (LFFS), Regional Failure-Free Survival (RFFS), and Distant Failure-Free Survival (DFFS).
Methods
The proposed DeepCox_MTL framework consists of three integrated modules: (a) a Multi-Modal Feature Extraction (MMFE) module, which fuses low-level semantic radiomic features with high-level deep imaging representations (extracted via an attention-guided global module) and clinical variables; (b) a Multi-Stage Feature Selection (MSFS) module, which utilizes an Enhanced Genetic Algorithm (EGA) with a regularized fitness function to identify a compact and informative predictor set; and (c) a Multi-Endpoint Survival Prediction (MESP) module. This module employs a shared backbone to capture complex non-linear representations across related endpoints, along with task-specific heads and an adaptive weighted loss function to dynamically balance learning based on individual task performance.
Results
The framework was validated on a public dataset of 606 patients with oropharyngeal cancer. DeepCox_MTL demonstrated strong and consistent performance across all endpoints. Specifically, it achieved high predictive accuracy for well-represented endpoints, with C-indices of 0.7770 for OS and 0.8018 for LFFS. Furthermore, it showed notable performance gains for sparse or imbalanced endpoints, yielding C-indices of 0.7507 for RFFS and 0.7179 for DFFS.
Conclusion
By integrating multi-level features (radiomics, deep learning embeddings, and clinical data) within an adaptive multi-task learning architecture, DeepCox_MTL effectively captures inter-task relationships to provide comprehensive survival predictions. The framework's ability to improve accuracy for both primary and sparse endpoints underscore its significant potential for clinical translation and personalized treatment planning in complex oncological scenarios.