Pan-Cancer Outcome Prediction Using Multi-Classifier Systems

Predicting the outcome of cancer progression from high-throughput expression profiling data is a daunting task in practice due to tumor heterogeneity and lack of observations compared to high dimensionality. Multi-task learning methods have been applied to leverage the similarity between cancers of different tissue origins and improve the performance of survival analysis models. Nonetheless, the existence of both inter-cancer similarity and intra-cancer heterogeneity renders accurate predictive modeling difficult. In this work, we propose a novel strategy by considering many different cancer types together and leveraging patient similarity within individual cancers and across different cancer types. Experimental results suggest that our proposed multi-classifier approach outperforms baselines and state-of-the-art multi-task learning methods in a pan-cancer setting. Furthermore, we demonstrate that the proposed method is particularly effective in improving the predictive performance for samples in rare subgroups.