MLGAN: a Meta-Learning based Generative Adversarial Network adapter for rare disease differentiation tasks

Abstract

Rare disease diagnosis is very challenging due to the rarity and lack of scientific knowledge. Many patients with rare diseases take years to get diagnosed and many stay misdiagnosed or are not diagnosed. Comparing with traditional diagnosis prediction task, rare disease detection has the unique challenges of low prevalence and label noise. In this paper, we propose Meta-Learning based Generative Adversarial Network module MLGAN, a rare disease detection enhancement module that can adapt any existing diagnosis prediction methods to rare disease detection task. We use generative adversarial network to generate synthetic positive em-beddings and we use Meta-Weight-Net to automatically assign weight to real data and synthetic data. MLGAN helps us to leverage the time-aware sequential modeling ability in diagnosis prediction methods, and also mitigate the low prevalence and label noise of rare disease dataset. We empirically show that MLGAN can greatly boost the prediction performance and have good robustness on four real-world rare disease datasets. We release our code at https://github.com/ruilialice/MLGAN.