Knowledge-guided meta learning for disease prediction

Annotated data samples in real-world biomedical applications are often limited. However, many machine learning approaches rely primarily on training systems based on big labeled data corpora. For instance, deep learning systems must be shown many examples before they can classify things accurately. While deep learning has achieved great success in many fields, it can easily lead to overfitting issues when dealing with a small number of samples with high-dimensional features in a cohort. The Cancer Genome Atlas (TCGA)[1] is such an example, which characterizes the molecular profiles of over 20,000 cancer and normal samples with clinical outcomes spanning 33 cancer types. However, the number of patients for each cancer type is small (varying from 51 to 1098), which shows the scenario of many different domains (cancer types) with each domain having few samples. Transfer-learning has been proposed to re-use the trained model parameters in similar applications. However, transfer learning cannot be effectively applied from one domain to a different domain. More recently, meta learning, which utilizes prior knowledge learned from related tasks and generalizes to new tasks of limited supervised experience, has shown to be an effective approach for few-shot learning. In this chapter, we will first show the advantage of meta learning over traditional classification in prediction problems and use TCGA data to demonstrate that meta learning can outperform conventional transfer learning in prediction problems. We will then present a knowledge-guided meta learning strategy which integrates biological knowledge with meta learning for improved classification performance on few-shot learning problems.