Explanation-Guided Reconstruction of Missing Clinical Features for Survival Prediction in Pancreatic Cancer

Incomplete clinical records can substantially limit survival prediction models for pancreatic ductal adenocarcinoma. We propose an explanation-guided reconstruction framework (xRF) that combines an autoencoder with an ensemble of gradient- and perturbation-based explainability methods. The framework prioritizes clinically relevant variables during reconstruction so that missing-data recovery focuses on features that matter most for downstream survival prediction. We validate xRF on 1,531 pancreatic cancer patients from the Danish Capital Region across post-diagnosis, post-metastasis, post-surgery, and post-chemotherapy survival tasks. Experiments across different levels of missingness show that the proposed approach preserves survival prediction performance better than alternative machine-learning reconstruction methods.

IEEE Journal of Biomedical and Health Informatics, 2026