FA-DeepMSM: a few-shot adapted interpretable multimodal survival model for improved prognostic prediction in glioblastoma

Survival prediction in IDH-wildtype glioblastoma is an inherently time-dependent challenge that remains clinically critical, yet limited by inter-institutional heterogeneity and insufficient labeled data. Existing deep learning prognostic models often overfit to high-dimensional imaging features and fail to generalize across clinical settings. To address these limitations, we developed FA-DeepMSM (Few-shot Adapted Deep Multimodal Survival Model), which integrates self-supervised MRI features with clinical and molecular variables within a few-shot learning framework. High-dimensional MRI embeddings were extracted from 1,359 adult-type diffuse glioma patients using large multi-institutional datasets combining in-house and publicly available cohorts with survival data through a pretrained DINOv2 vision transformer and fused with structured tabular data in a transformer-based survival architecture. The model was fine-tuned under 0-, 5-, 10-, 20-, and 40-shot settings using an external glioblastoma cohort from UPenn (n = 452). Few-shot adaptation significantly improved generalization, increasing the average time-dependent C-index from 0.643 (0-shot) to 0.680 (40-shot). To enhance interpretability, FA-DeepMSM incorporates a time-resolved explainability module based on permutation analysis, enabling variable-level risk attribution across survival timepoints (3, 6, 12, 18, and 24 months). Early prognosis was primarily driven by extent of resection, whereas later survival phases were more influenced by MGMT promoter methylation and image-derived features. By addressing data scarcity, cross-modal misalignment, and limited interpretability, FA-DeepMSM establishes a clinically scalable and explainable paradigm for outcome prediction in neuro-oncology.