A Contrastive Variational AutoEncoder for NSCLC Survival Prediction with Missing Modalities

Technical Deep Dive: A Contrastive Variational AutoEncoder for NSCLC Survival Prediction with Missing Modalities

Overview

This paper proposes a novel multimodal learning framework called Multimodal Contrastive Variational AutoEncoder (MCVAE) to address the challenge of missing modalities in predicting survival outcomes for non-small cell lung cancer (NSCLC) patients. MCVAE learns robust, modality-agnostic representations by explicitly modeling the uncertainty in each data source and using an adaptive fusion mechanism to weigh available modalities based on their learned importance and availability.

Problem & Context

• Predicting survival outcomes for NSCLC patients is difficult due to the diverse prognostic features.
• Integrating histopathology, transcriptomics, and DNA methylation data can provide complementary views, but real-world datasets often have missing modalities.
• Existing methods either discard information through modality dropout or introduce bias through imputation, limiting their robustness to severe missingness.

Methodology

• MCVAE uses modality-specific variational encoders to capture the uncertainty in each data source.
• An adaptive fusion module with learned gating weights normalizes the contributions from present modalities.
• The training objective combines survival prediction, reconstruction regularization, KL divergence, and contrastive cross-modal alignment.
• Stochastic modality masking during training improves robustness to arbitrary missingness patterns.

Data & Experimental Setup

• Evaluated on TCGA-LUAD (n=475) and TCGA-LUSC (n=446) datasets, using disease-specific survival (DSS) as the clinical endpoint.
• Extracted features from histopathology, transcriptomics, and DNA methylation data using state-of-the-art methods.
• Compared to two representative methods: MUSE (representation-based) and M3Care (imputation-based).
• Conducted 5-fold cross-validation with 3 random seeds, reporting mean and standard deviation of Harrell's C-index.

Results

• On LUAD, MCVAE achieved a mean C-index of 0.651 ± 0.026, outperforming MUSE (0.607 ± 0.056) and M3Care (0.609 ± 0.034).
• On LUSC, MCVAE obtained 0.575 ± 0.033, still surpassing MUSE (0.542 ± 0.070) and M3Care (0.558 ± 0.037).
• MCVAE demonstrated more stable performance across folds compared to the baselines.

Interpretation

• Modality combination analysis revealed that integrating multiple modalities does not universally improve performance, with some unimodal or bimodal models outperforming those using all modalities.
• Increasing modality dropout rates during training improved MCVAE's performance on LUAD, suggesting that consistent masking enforces learning of modality-agnostic features.
• MCVAE maintained robust performance even under extreme missingness (up to 90% of modalities missing), outperforming the baselines.

Limitations & Uncertainties

• Evaluation was limited to NSCLC cohorts from TCGA, a retrospective research database. Validation on diverse cancer types and prospective clinical datasets is needed.
• Broader comparison against additional multimodal methods and systematic unimodal baselines could provide a more comprehensive evaluation.
• The reconstruction component of MCVAE does not impute missing modalities, which could be valuable for understanding disease patterns.

What Comes Next

• Investigate the interpretability of MCVAE's learned representations and cross-modal relationships using probing strategies like cross-attention transformers.
• Explore controlled imputation strategies that quantify uncertainty in generated modalities.
• Assess the clinical applicability of MCVAE's uncertainty estimates and validate the prospective predictive performance.