MIEdeep - Medical Information Extraction for German Medical Texts using Deep Learning Methods

德国非结构化医学文本的稳定增长需要一个强大的，数据驱动的和模块化的软件框架来进行医学信息提取，这得益于自然语言处理（NLP）领域深度学习的最新成就。目前的研究还没有充分解决这一需求，特别是在德国医学文本数据的背景下。我们建议MIEdeep（使用深度学习的医学信息提取），以（i）整合最先进的深度学习方法进行数据准备，计算机辅助快速训练数据创建和医学信息提取，以及（ii）通过将这些技术插入现代化，易于访问的图形用户界面，使临床研究人员可以使用这些技术。特别是，我们使用数据编程，这使得用户能够快速创建大量的标记训练数据，而不需要耗时且昂贵的手动注释。我们在这些标记的训练数据的基础上，整合了最先进的监督深度学习模型，用于医学信息提取。MIEdeep将这些步骤包装成一个直观易用的图形用户界面，使深度学习对临床领域专家变得平易近人。我们的建议旨在利用来自临床常规的大量非结构化医疗数据，用于临床研究等次要用例。使用数据编程来创建训练数据最终可以将医疗数据和深度学习模型结合在一起。在后续的项目中，这些研究成果可以作为医务人员的决策支持，从而改善患者的治疗。该项目有可能对临床NLP研究产生影响，通过提供一个强大的软件框架，使用最先进的方法快速创建训练数据，以解决当前医学信息提取中的障碍。然后，提取的信息可以反馈到临床常规和程序中。由于MIEdeep框架可以应用于不同的临床环境，我们的目标是在知识共享许可下提供框架，并将其发布在公共存储库中。这使我们能够传播正在进行的工作，并与潜在的利益相关者和最终用户进行互动。

A steadily growing pool of German unstructured medical texts created a need for a robust, data-driven and modular software framework for medical information extraction, which benefits from most recent achievements in deep learning in the domain of natural language processing (NLP). This need has not been sufficiently addressed by current research especially in the context of German medical textual data. We propose MIEdeep (Medical Information Extraction using deep learning) to (i) integrate state-of-the-art deep learning approaches for data preparation, computer-assisted rapid training data creation and medical information extraction and (ii) to make these technologies available for clinical researchers by plugging them into a modern, easy-to-access graphical user interface.In particular, we use data programming, which enables users to rapidly create large amounts of labeled training data without the need for time consuming and costly manual annotation. We build on these labeled training data to integrate state-of-the-art supervised deep learning models for medical information extraction. MIEdeep wraps these steps up into one intuitive and easy-to-use graphical user interface to make deep learning approachable for clinical domain experts.Our proposal aims to leverage the vast pool of unstructured medical data from clinical routine for secondary use cases like clinical research. The use of data programming to create training data could finally bring together healthcare data and deep learning models. In a following project these research outcomes can be used as decision support of the medical staff, thus to improve patient treatment. This project has the potential to create impact on research in clinical NLP, by providing a robust software framework using a state-of-the-art method for rapid training data creation to address current obstacles in medical information extraction. The extracted information can then be fed back into clinical routines and procedures. As the MIEdeep framework can be applied in different clinical settings, we aim to provide the framework under a creative commons license and to publish it on a public repository. This enables us, to disseminate the ongoing work and to interact with potential stakeholders and end-users.