MultiSurv: Long-term cancer survival prediction using multimodal deep learning

The age of precision medicine demands powerful computational techniques to handle high-dimensional patient data. We present MultiSurv, a multimodal deep learning method for long-term pan-cancer survival prediction. MultiSurv is composed of three main modules. A feature representation module includes a dedicated submodel for each input data modality. A data fusion layer aggregates the multimodal representations. Finally, a prediction submodel yields conditional survival probabilities for a predefined set of follow-up time intervals. We trained MultiSurv on clinical, imaging, and four different high-dimensional omics data modalities from patients diagnosed with one of 33 different cancer types. We evaluated unimodal input configurations against several previous methods and different multimodal data combinations. MultiSurv achieved the best results according to different time-dependent metrics and delivered highly accurate long-term patient survival curves. The best performance was obtained when combining clinical information with either gene expression or DNA methylation data, depending on the evaluation metric. Additionally, MultiSurv can handle missing data, including missing values and complete data modalitites. Interestingly, for unimodal data we found that simpler modeling approaches, including the classical Cox proportional hazards method, can achieve results rivaling those of more complex methods for certain data modalities. We also show how the learned feature representations of MultiSurv can be used to visualize relationships between cancer types and individual patients, after embedding into a low-dimensional space.