Research map / Vision-language and multimodal / Cross-modal learning
Distilled Prompt Learning for Incomplete Multimodal Survival Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DisPro, a two-stage prompt-learning framework that uses an LLM (BioBERT) with unimodal and multimodal prompts to compensate both modality-common and modality-specific knowledge for missing pathology or genomic data in survival prediction.
Problem
Multimodal survival prediction from pathology WSIs and gene profiles degrades when one modality is missing, and existing imputation methods only recover modality-common information.
Approach
Two-stage prompting: UniPro distills per-modality knowledge into learnable MIL-style prompts (CoOp-inspired); MultiPro uses the available modality as LLM prompts to infer the missing one, supervised by UniPro distillation loss, with UniPro scoring to select discriminative tokens.
Key results
- At 60% training missing rate, DisPro averages 0.6796 C-Index on pathology-only test, outperforming the best incomplete baseline (COM/HGCN ~0.66) by about 1.6-2 points.
- With complete test data, DisPro at 60% missing training reaches 0.7021 average C-Index, surpassing SurvPath trained on complete data (0.6982).
- Ablation on 60% missing rate: UniPro distillation (UD) adds about +1 point and UniPro scoring (US) adds about +1 point on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DisPro instead of HGCN | pathology-only inference with 60% training missing rate · TCGA five-dataset average | C-Index: 0.6620 -> 0.6796 | Helped |
| Use DisPro instead of COM | genomics-only inference with 60% training missing rate · TCGA five-dataset average | C-Index: 0.6645 -> 0.6802 | Helped |
| Use DisPro instead of M3Care | complete-modality inference with 60% training missing rate · TCGA five-dataset average | C-Index: 0.6788 -> 0.7021 | Helped |
| Add UniPro Distillation (ablation) | DisPro on pathology-only inference · TCGA five-dataset average, 60% training missing rate | C-Index: 0.6534 -> 0.6652 | Helped |
| Add UniPro Scoring (ablation) | DisPro on pathology-only inference · TCGA five-dataset average, 60% training missing rate | C-Index: 0.6534 -> 0.6714 | Helped |
| Add UniPro Distillation (ablation) | DisPro on genomics-only inference · TCGA five-dataset average, 60% training missing rate | C-Index: 0.6647 -> 0.6732 | Helped |
| Add UniPro Scoring (ablation) | DisPro on genomics-only inference · TCGA five-dataset average, 60% training missing rate | C-Index: 0.6647 -> 0.6701 | Helped |
| Add UniPro Distillation (ablation) | DisPro on complete-modality inference · TCGA five-dataset average, 60% training missing rate | C-Index: 0.6816 -> 0.6926 | Helped |
| Add UniPro Scoring (ablation) | DisPro on complete-modality inference · TCGA five-dataset average, 60% training missing rate | C-Index: 0.6816 -> 0.6943 | Helped |
Surprising findings
- Trained only on incomplete data with 60% missing rate, DisPro's complete-test average (0.7021) slightly exceeds SurvPath trained on complete data (0.6982).
Limitations
- Evaluated only on TCGA survival cohorts with simulated missingness rather than real clinical missing data.
- Relies on a BERT-style LLM with a 512-token input limit, requiring token selection.
Experimental setup
Datasets: TCGA BLCA TCGA BRCA TCGA COADREAD TCGA LUAD TCGA UCEC
Benchmarks: TCGA BLCA TCGA BRCA TCGA COADREAD TCGA LUAD TCGA UCEC
Metrics: C-Index standard deviation
Baselines: TransMIL CoOp SNN MOTCat SurvPath COM M3Care HGCN MUSE MAP
Builds on: CoOp MCAT SurvPath MAP
Closest papers by meaning
- Knowledge Bridger: Towards Training-Free Missing Modality Completion (CVPR 2025 · similarity 0.88)
- Flex-MoE: Modeling Arbitrary Modality Combination via the Flexible Mixture-of-Experts (NeurIPS 2024 · similarity 0.87)
- Prototypical Information Bottlenecking and Disentangling for Multimodal Cancer Survival Prediction (ICLR 2024 · similarity 0.87)
- CA-MLIF: Cross-Attention and Multimodal Low-Rank Interaction Fusion Framework for Tumor Prognostic Prediction (AAAI 2025 · similarity 0.86)
- HEALNet: Multimodal Fusion for Heterogeneous Biomedical Data (NeurIPS 2024 · similarity 0.86)
- Modalities Contribute Unequally: Enhancing Medical Multi-modal Learning through Adaptive Modality Token Re-balancing (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.