MEDICAL KNOWLEDGE-ENABLED MULTI-TASK LEARNING FOR GASTRIC CANCER SURVIVAL PREDICTION

Accurate cancer survival prediction can inform patient prognostication and treatment planning. Multi-source clinical data, including radiological images and pathological markers, demand effective machine learning strategies to make smart use of these data to enhance patient outcome. In this work, we proposed a novel strategy of translating multi-task learning to clinical applications for gastric cancer survival prediction, by leveraging the established medical knowledge of abdominal CT images and pathological staging markers to construct meaningful auxiliary tasks for joint learning with the target task. In implementing our method for survival prediction, we also customized a convolutional neural network to account for censored data. We evaluated our method with extensive experiments on a large cohort of 1,061 gastric cancer patients for both overall and progression-free survival prediction, showing superior performance when compared to multi-input learning and two state-of-the-art prediction methods. As an additional benefit, our method also enables survival prediction even before surgery. Source code is available at https://github.com/usernameoliver/MTL.git.