EIRsurvival: Deep Learning-based time-to-event analysis on high-dimensional genotype and multi-omics data

Time-to-event data in disease occurrence is often right-censored, requiring survival models for accurate predictions. While deep learning advancements have extended traditional Cox models, current approaches do not allow modeling on individual-level, large-scale genotype data. Scalable models integrating genetic and clinical data could enhance precision medicine and disease prediction. We introduce EIRsurvival, a deep learning-based tool designed to perform time-to-event analyses on millions of SNPs from large, individual-level cohorts. We applied EIRsurvival to predict time-to-diagnosis for eight diseases in the UK Biobank, utilizing 1.13M genetic variants from 487,027 individuals. The model achieved C-indices between 0.55 and 0.68 using genetic data alone, with performance improving to 0.67–0.96 when integrating multi-omics data. Automated feature attribution analysis confirmed biologically relevant feature associations. EIRsurvival is available through the eir-dl framework at github.com/arnor-sigurdsson/EIR, with documentation accessible at eir.readthedocs.io/en/latest/other/01_survival_genotypes.html. The tool is optimized for GPU acceleration, facilitating efficient training on high-dimensional data. Simon Rasmussen ([email protected]) Supplementary information is available online.