Learning Clinical Outcomes from Heterogeneous Genomic Data Sources

Abstract

Translating the vast data generated by genomic platforms into reliable predictions of clinical outcomes remains a critical challenge in realizing the promise of genomic medicine largely due to small number of independent samples. In this paper, we show that neural networks can be trained to predict clinical outcomes using heterogeneous genomic data sources via multi-task learning and adversarial representation learning, allowing one to combine multiple cohorts and outcomes in training. We compare our proposed method to two baselines and demonstrate that it can be used to help mitigate the data scarcity and clinical outcome censorship in cancer genomics learning problems.