Multimodality Stacking with Blockwise missing values and application to the PIONeeR biomarkers study for prediction of resistance to immunotherapy

2025-05-01 · Source: stat.ML updates on arXiv.org · Field: Health & Wellbeing — Health & Medical Research, Medical Specialties & Subspecialties, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

The Multimodality Stacking with Blockwise missing values (MSB) framework is introduced for survival analysis, specifically to predict progression-free survival in advanced non-small cell lung cancer patients undergoing immunotherapy. Developed to address high dimensionality and blockwise missingness in clinical oncology datasets, MSB is a late-fusion approach that independently models modality-specific features before aggregating predictions using a cross-validated stacking meta-learner. Validated on the PIONeeR study (443 patients, 378 biomarkers from eight sources), MSB achieved superior predictive performance (C-index) compared to baseline algorithms. It demonstrated a 15.9% increase for linear models ($p<0.001$), 5.4% for random survival forests ($p=0.002$), and 2.1% for gradient boosting methods ($p=0.030$). MSB also significantly reduced the generalization gap, for instance, from 0.380 to 0.055 for linear models. Permutation importance analysis identified routine laboratory markers, clinical features, and PD-L1 expression as primary predictive drivers, with missing block indicators showing negligible importance. The framework's implementation is available under an Inria license.

Key takeaway

For AI or Research Scientists building prognostic models with multimodal, incomplete clinical data, you should consider the MSB framework. It significantly improves predictive performance and reduces overfitting compared to traditional methods, particularly when dealing with blockwise missingness and high dimensionality. By leveraging modality-specific risk scores, MSB allows for robust biomarker evaluation without requiring complete data. Prioritize integrating readily available clinical features, routine lab markers, and PD-L1 expression, as these were identified as primary predictive drivers.

Key insights

MSB effectively handles high-dimensional, blockwise missing multimodal clinical data for survival prediction via a late-fusion stacking approach.

Principles

• Late-fusion stacking mitigates high dimensionality and blockwise missingness.
• Decomposing global feature space into modality-specific sub-problems enhances robustness.
• Cross-validated aggregation reduces overfitting in small cohorts.

Method

Train independent base-learners on available data for each modality. Aggregate out-of-fold cross-validated predictions (risk scores) from base learners using a meta-learner. Handle residual intra-block missingness with source-specific kNN imputation or MIA.

In practice

• Apply MSB to integrate diverse clinical data sources for survival prediction.
• Prioritize clinical features, routine lab markers, and PD-L1 expression in oncology models.
• Consider late-fusion for small clinical cohorts with high-dimensional, sparse data.

Topics

• Multimodality Stacking
• Survival Analysis
• Blockwise Missing Data
• Clinical Oncology
• Immunotherapy Resistance
• Biomarker Discovery

Code references

• MohamedBoussena/MSB

Best for: AI Scientist, Research Scientist, Data Scientist

Related on AIssential

• Predicting Immune Biomarkers with MultiModal Mixture-of-Expert Pathology Foundation Models Empowers Precision Oncology — MixTIME integrates multimodal pathology foundation models via a Mixture-of-Experts architecture to predict immune biomarkers for precision oncology.
• Multi-FRuGaL: Multimodal Flexible Redundancy-aware Decomposed Gated Learning for Cancer Diagnosis and Prognosis — Adaptive, redundancy-aware gated fusion improves cancer prognosis with incomplete multimodal patient data.
• MamaBench: Benchmarking LLM Robustness in Maternal and Child Health Diagnosis through Counterfactual Clinical Perturbation — Current medical LLM benchmarks lack counterfactual evaluation, overstating diagnostic robustness in clinically similar cases.
• Systematically decoding pathological morphologies and molecular profiles with unified multimodal embedding — Multi-Embed unifies morphological and molecular data for superior disease understanding and prediction.
• 🔬 Training Transformers to solve 95% failure rate of Cancer Trials — Ron Alfa & Daniel Bear, Noetik — Improving cancer drug success requires advanced AI models for precise patient selection based on multimodal biological data.
• Generalizable AI predicts immunotherapy outcomes across cancers and treatments — COMPASS uses a concept bottleneck transformer to predict ICI response from transcriptomes, offering interpretable mechanistic insights.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.