#26 · ML / AI methods

Multimodal TCGA survival prediction targets calibration alongside predictive performance

The dominant thread is multimodal fusion for prediction from heterogeneous TCGA cancer data. A locally deployable LLM framework fuses clinical text, covariates and genomics for survival prediction and explicitly targets calibration alongside performance, reporting gains over baselines. MUST, a modality-specific transformer, uses diffusion-based generation to handle missing pathology or genomics modalities in survival prediction. SynLeaF applies dual-stage multimodal fusion with VAE cross-encoders to synthetic lethality prediction across pan- and single-cancer contexts rather than survival outcomes. Only the first item names calibration as a target; the other two report predictive modeling without stated uncertainty evaluation. What is new is calibration appearing as an explicit objective alongside predictive performance in multimodal survival modeling.