Convex space learning for synthetic data generation on clinical tabular datasets

合成数据生成在生物医学研究中解决实际问题日益突出：个性化；临床试验中群体代表性不足；数据私隐阻碍了机构间的数据共享等。基于深度生成网络的医学图像合成数据生成是一个新兴的研究领域。对于图像数据集，有一个感知上的优势，从某种意义上说，人们可以通过观察来直观地判断合成图像的逼真程度。然而，在生物医学科学中，表格数据集是存储患者数据的一种非常常见的方式，对于这样的数据，视觉感知的优势是有限的。自2017年以来，研究人员一直专注于为表格数据集开发深度生成模型。在过去的三年中，我们在表格合成数据生成方面发展了专业知识，以解决不平衡分类问题。我们在过采样驱动的不平衡分类领域开发了多种算法，并测试了它们在生物学问题（如单细胞转录组学数据的稀有细胞注释）中的适用性。在我们的研究中出现了凸空间学习的概念，并对凸空间学习的理论基础进行了探讨。使用我们最新的凸空间学习模型ConvGeN，与为表格数据集设计的最先进的深度生成算法相比，我们能够使用合成样本生成来改进表格不平衡数据集的分类。与现有的表格数据集深度生成算法相比，使用ConvGeN生成的合成样本可以更好地近似特征统计分布，因为来自ConvGeN的合成样本在表格数据中固定了特征智能方法，同时以非线性迭代的方式学习适当的特征智能高阶矩。我们认为，凸空间学习在我们迄今为止探索的不平衡分类领域之外具有广泛的潜力。我们建议扩展我们的模型ConvGeN，使其能够在数据不平衡的情况下生成合成表格数据。此外，我们建议研究使用凸空间学习生成的合成数据在临床领域的几种机器学习应用的潜在用途，如患者分层、分类、回归问题等。目标是确定涉及合成数据生成的给定机器学习工作流是否可以产生与使用真实数据相似的足够性能，例如在患者分层中。最后，我们建议将开发的算法用于实际临床问题的合成样本生成，以解决与临床合作伙伴相关的隐私保护等问题。

Synthetic data generation is gaining prominence in biomedical research in solving practical problems: personalization; underrepresentation of groups in clinical trials; data privacy hindering sharing of data among institutions etc. Synthetic data generation using deep generative networks for medical images is a booming research field. For image datasets, there is a perceptional advantage, in the sense, that one can visually judge how realistic the synthetic image is, just by looking at it. However, in biomedical science, tabular datasets are a very common way of storing patient data, and for such data the advantage of visual perception is limited. Since 2017, researchers have focused on developing deep generative models for tabular datasets. Over the last three years, we have developed expertise in tabular synthetic data generation to solve the problem of imbalanced classification. We developed multiple algorithms in the domain of oversampling-driven imbalanced classification and tested their applicability to biological problems such as rare-cell annotation from single-cell transcriptomics data. From our studies emerged the idea of convex space learning, whose theoretical foundations were also explored in our studies. With our newest convex space learning model ConvGeN, we were able to improve classification on tabular imbalanced datasets using synthetic sample generation, compared to the state-of-the-art deep generative algorithms designed for tabular datasets. Synthetic samples generated using ConvGeN can approximate feature-wise statistical distributions better compared to existing deep generative algorithms for tabular datasets since the synthetic samples from ConvGeN fix feature-wise means in tabular data while learning appropriate feature-wise higher-order moments in a non-linear iterative fashion. We argue that convex space learning has extensive potential outside the domain of imbalanced classification that we have explored so far. We propose to extend our model ConvGeN, enabling it to generate synthetic tabular data outside the context of data imbalance. Furthermore, we propose to investigate the potential use of the synthetic data generated using convex space learning for several applications of machine learning in the clinical domain such as patient stratification, classification, regression problems, etc. The goal is to establish whether a given machine learning workflow involving synthetic data generation can produce similar enough performance as using real data, e.g. in patient stratification. Finally, we propose to use the developed algorithm for synthetic sample generation in real-life clinical problems to solve issues like privacy preservation in association with our clinical partners.