SHIFT: Survival Prediction from Incomplete and Heterogeneous Genomic Data

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Science & Research — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Health & Medical Research · Depth: Expert, extended

Summary

SHIFT, a novel Transformer-based model, addresses the challenge of survival prediction from incomplete and heterogeneous genomic data in precision oncology. Traditional genomic prediction models struggle with cross-institutional transfer due to structural missingness caused by varying sequencing panels. SHIFT employs masked self-attention and a feature-availability mask to directly predict from observed inputs, eliminating the need for test-time imputation. It incorporates variable-rate feature masking (VRM) during training to enhance robustness against diverse missingness patterns. Evaluated on glioblastoma and lung squamous cell carcinoma, SHIFT demonstrated strong generalization, outperforming standard survival baselines and imputation methods, particularly in a challenging LUSC setting where 88.8% of features were structurally absent. The model also showed that integrating incomplete cohorts, such as 102 US LUSC patients with only 22 of 197 features, significantly improves performance on external data, suggesting a more inclusive approach to multi-center model development.

Key takeaway

For AI Scientists and Machine Learning Engineers developing survival prediction models with multi-center genomic data, consider adopting missingness-aware architectures like SHIFT. This approach allows you to train a single robust model across heterogeneous genomic panels, avoiding complex test-time imputation strategies. You can also incorporate partially observed patient cohorts into your training data, improving overall model performance and diversity without discarding valuable information.

Key insights

Missingness-aware Transformer models can robustly predict survival from incomplete genomic data without imputation.

Principles

• A single model can predict survival across institutions despite data variability.
• Training with simulated missing data improves robustness and performance.
• Incomplete cohorts provide useful signals for model development.

Method

SHIFT uses a feature embedding block, a masked Transformer encoder with a feature-availability mask, and a survival prediction block. Variable-rate feature masking during training enhances robustness to heterogeneous missingness.

In practice

• Use masked self-attention for structural missingness in tabular data.
• Implement variable-rate masking during training for improved generalization.
• Include partially observed patient cohorts in model development.

Topics

• Survival Prediction
• Genomic Data Analysis
• Missing Data Handling
• Transformer Models
• Precision Oncology
• Multi-center Data

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Tabular Foundation Models for Clinical Survival Analysis via Survival-Aware Adaptation — Adapting tabular foundation models with a survival-aware head significantly improves clinical time-to-event prediction.
• Evaluating Reliability in Machine Learning Models for Early Chronic Kidney Disease Prediction: A Systematic Review of Data Leakage and Predictor Stability — Data leakage significantly inflates reported machine learning model performance in early CKD prediction, with most predictors lacking stability.
• 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.
• I Gave a Model Eyes and a Genome. Here’s What It Learned About Brain Cancer. — Cross-attention fusion of pathology and genomics significantly improves glioblastoma survival prediction over unimodal approaches.
• Progression as Latent Drift: Generative Forecasting of Slow-Evolving Pathologies — Latent Drift forecasts disease progression by learning subtle changes in compressed semantic representations, avoiding common generative model failures.
• Multimodality Stacking with Blockwise missing values and application to the PIONeeR biomarkers study for prediction of resistance to immunotherapy — MSB effectively handles high-dimensional, blockwise missing multimodal clinical data for survival prediction via a late-fusion stacking approach.

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 cs.AI updates on arXiv.org.