SHIFT Predicts Survival from Incomplete Genomic Data

Genomic prediction models often face significant challenges when transferring across different institutions due to variations in sequencing panels, which lead to structural missingness in feature data during deployment. Traditional solutions, such as restricting analysis to shared genes, excluding incomplete patient profiles, or relying on test-time imputation, can compromise model robustness and limit the utility of multi-center datasets. To address this, researchers have developed SHIFT (Survival prediction Handling Incomplete Features using Transformer). SHIFT is a novel missingness-aware survival model designed to make predictions directly from incomplete genomic inputs, eliminating the need for test-time imputation. It achieves this by representing each genomic feature separately and employing masked self-attention, along with a feature-availability mask, ensuring that predictions are based solely on observed data. Furthermore, SHIFT incorporates variable-rate feature masking during training, which significantly enhances its robustness to diverse missingness patterns. The model was evaluated on glioblastoma and lung squamous cell carcinoma datasets, demonstrating strong generalization and favorable performance compared to standard survival baselines and imputation-based methods, even in challenging scenarios with severe cross-cohort panel mismatches. The study also found that including patients from partially observed cohorts during development can improve external data performance, suggesting that such data should not be discarded.