ProtoSurv-X: Explainable Brain Tumor Survival Prediction

Abstract

Accurate survival prediction for patients with high-grade glioma is important for prognostic assessment and personalized treatment planning; however, substantial intratumoral heterogeneity, complex multimodal MRI patterns, and limited interpretability challenge existing deep learning approaches. This study aims to develop ProtoSurv-X, an explainable and uncertainty-aware framework for MRI-based glioma survival prediction that integrates probabilistic tumor habitat modeling, prototype-guided learning, evidential prediction, and evidence-grounded clinical explanations. A unified cohort was constructed from the BraTS 2019 and BraTS 2020 datasets by removing duplicate subjects, resulting in 369 unique subjects, including 118 gross total resection patients with complete survival annotations. The proposed framework uses diffusion-enhanced SwinUNETR for tumor segmentation, probabilistic habitat construction, and fusion of radiomic, deep imaging, habitat, and age-related features. Prototype learning and survival-aware contrastive learning generate prognostic representations, while Evidential Deep Learning estimates risk and predictive uncertainty alongside continuous survival regression. For explanation generation, seven open-source large language models were benchmarked using structured model evidence, and a Meta-Llama-3.1-8B-Instruct model was fine-tuned using QLoRA. In five-fold cross-validation on the unified survival cohort, ProtoSurv-X achieved a mean absolute error of 145.2 ± 6.5 days, an RMSE of 189.4 ± 8.3 days, a C-index of 0.745 ± 0.008, and an integrated Brier score of 0.124. The segmentation module achieved Dice scores of 91.24%, 87.38%, and 81.76% for whole tumor, tumor core, and enhancing tumor, respectively, on BraTS 2019. The fine-tuned explanation model obtained a BERTScore of 0.915 and a hallucination rate of 2.6%. These findings indicate that ProtoSurv-X offers a unified computational framework for accurate, uncertainty-aware, and evidence-grounded glioma prognostic modeling, while further clinical and multicenter validation remains necessary