SHIFT transformer predicts cancer survival from incomplete genomic data
Handles missing genes across hospitals without imputation or patient exclusion
Researchers propose SHIFT, a missingness-aware survival model that predicts survival directly from incomplete genomic inputs. Using masked self-attention and variable-rate feature masking, SHIFT handles heterogeneous sequencing panels without test-time imputation. Validated on glioblastoma and lung squamous cell carcinoma across multiple cohorts, it compares favorably with standard survival baselines and imputation-based approaches, and shows that incorporating patients from incomplete cohorts can improve performance on external data.
- SHIFT uses masked self-attention and feature-availability masks to predict survival from incomplete genomic data without test-time imputation
- Validated on glioblastoma and lung squamous cell carcinoma across multiple cohorts, including severe cross-cohort panel mismatch
- Variable-rate feature masking during training improves robustness to heterogeneous missingness patterns
Why It Matters
Removes a key barrier to multi-institutional genomic collaboration, enabling broader and more robust survival models in precision oncology.