OMNI-SURV-X: Ontology-guided Multimodal Weakly-supervised Contrastive Survival Modeling with Communication-efficient Federated Distillation

This research presents OMNI-SURV-X, a clinically practical survival-prediction framework that learns from routine chest X-rays and radiology reports while respecting privacy and coping with scarce, censored labels.The novelty lies in a four-part design tailored to real clinical constraints.First, ontology-guided fusion anchors Vision-Transformer image tokens and BioBERT text features to shared UMLS concepts, producing explanations clinicians can audit and reproducing the concept extraction and filtering pipeline in full.Second, a survival-aware supervised contrastive objective forms positives by risk proximity and concept overlap, strengthening representations when events are rare or censored.Third, communication-efficient federated knowledge distillation (Fed-KD) aggregates teacher signals across sites using lightweight logits and embeddings-without sharing raw data or full weights-making collaboration feasible under strict privacy and bandwidth limits and grounding this setup in an explicit threat model with a privacy-risk assessment.Fourth, a stability curriculum (CyclicAugment + SurvMix) alternates weak→strong→weak augmentations and mixes survival-aware features to reduce overfitting.Across three public cohorts with patient counts, event counts, censoring rates, and follow-up summaries reported, and using patient-level, leakage-checked splits with external validation, OMNI-SURV-X improves concordance index by +0.024 to +0.041, lowers integrated Brier score by 1.7-3.5 points, and raises time-dependent AUC at 6/12 months versus strong unimodal and recent state-of-the-art baselines, with bootstrap 95% confidence intervals and sensitivity checks confirming stability of gains.The model is better calibrated (ECE ↓ 18-26%), robust to missing modality, and achieves comparable accuracy with ~40× less communication than weight-sharing federation, offering a clear path to trustworthy, deployable multimodal prognosis and enabling replication through released split manifests, fixed seeds, and configuration details.