[Paper Review] Semi-supervised Rare Disease Detection Using Generative Adversarial Network

Wenyuan Li, Yunlong Wang+4 more|arXiv (Cornell University)|Dec 3, 2018

Artificial Intelligence in Healthcare12 references17 citations

TL;DR

This paper proposes a semi-supervised generative adversarial network (GAN) framework for rare disease detection using limited labeled data and abundant unlabeled electronic health records. By leveraging unlabeled data through a modified GAN loss with feature matching and pull-together regularization, the model achieves a 34.18% PR-AUC, significantly outperforming baseline methods like logistic regression and random forest.

ABSTRACT

Rare diseases affect a relatively small number of people, which limits investment in research for treatments and cures. Developing an efficient method for rare disease detection is a crucial first step towards subsequent clinical research. In this paper, we present a semi-supervised learning framework for rare disease detection using generative adversarial networks. Our method takes advantage of the large amount of unlabeled data for disease detection and achieves the best results in terms of precision-recall score compared to baseline techniques.

Motivation & Objective

• Address the challenge of rare disease detection in highly imbalanced medical datasets with limited labeled cases.
• Leverage large volumes of unlabeled electronic health record (EHR) data to improve model performance despite scarce labeled examples.
• Develop a GAN-based semi-supervised learning framework tailored for rare disease detection to enhance precision and recall.
• Design a novel loss function that improves generator quality by encouraging generation of samples in the complement of the data manifold.
• Conduct an ablation study to evaluate the impact of individual components on model performance in rare disease detection.

Proposed method

• Use a GAN architecture with a discriminator D and generator G, where D classifies patients as having or not having the rare disease.
• Train the GAN using a semi-supervised loss that combines supervised cross-entropy on labeled data and a consistency loss on unlabeled data.
• Introduce feature matching (FM) to align generated samples with real data distributions in latent space.
• Apply a pull-together (PT) regularization term to encourage the generator to produce diverse samples near low-density regions of the data manifold.
• Optimize the generator loss as $ L_G = L_{fm} + L_{pt} $, combining feature matching and diversity-promoting regularization.
• Normalize all 797 input features (symptom counts, frequencies, time differences, age, gender) to [-1, 1] to stabilize GAN training.

Experimental results

Research questions

• RQ1 Can a semi-supervised GAN framework effectively improve rare disease detection performance when only a small number of labeled cases are available?
• RQ2 How do different components of the GAN loss function—specifically feature matching and pull-together regularization—affect model performance?
• RQ3 What is the contribution of unlabeled data in improving precision and recall in highly imbalanced rare disease detection tasks?
• RQ4 Does the inclusion of a conditional entropy term improve model generalization, or does it degrade performance due to exploration-exploitation trade-offs?
• RQ5 How does the proposed GAN-based method compare to traditional supervised models like logistic regression, neural networks, and random forests in rare disease detection?

Key findings

• The proposed semi-supervised GAN achieved a PR-AUC of 34.18%, representing a 5% improvement over the next-best baseline model.
• The model outperformed logistic regression (29.04% PR-AUC), a neural network with the same architecture as the discriminator (28.95% PR-AUC), and random forest (10.51% PR-AUC).
• The ablation study showed that the combination of feature matching and pull-together regularization (FM + PT) yielded the highest performance (34.18% PR-AUC).
• Models without unlabeled data (e.g., NN(D) and original GAN) achieved significantly lower PR-AUC scores, confirming the value of unlabeled data in semi-supervised learning.
• The inclusion of the conditional entropy term (Ent) degraded performance, suggesting potential negative effects due to over-exploration or instability in the training process.
• The results demonstrate that generating samples in the complement of the data manifold improves decision boundary generalization, especially in low-data regimes.