Semantic Text Analytics for Quality Controlled Extraction of Clinical Phenotype Information in Healthcare Integrated Bioban­king STACI2B2

越来越多的高质量生物材料的可用性是转化生物医学中可持续和可重复性结果的先决条件。这一观察结果不仅适用于探索性研究，而且越来越多地适用于验证研究结果。然而，一些人已经表达了对大量临床前研究结果在转化为临床实践时是否能兑现承诺的怀疑。以正在进行的生物标志物研究为例。对新型生物标志物的大量研究与临床验证应用的数量之间存在明显的差异。一个主要问题是缺乏对辅助样本质量差异的关注，以及基于与定义明确的疾病和与目标疾病不同的合并症的比较集体的潜在标记物的验证不足。虽然在过去，通过建立和维护专业的生物库，在许多临床地点建立了高质量收集和储存这些生物样本的基础设施，但仍然缺乏常规的工作流程来采集有效的表型数据，确定有效的比较组，并为高质量的生物库正确选择样本。在我们的项目中，我们建议使用自动自然语言处理的方法从临床文档中提取这些信息。我们计划使用半监督机器学习技术建立一个文本分析管道，从非结构化的临床文档（如出院摘要、放射学或病理报告等）中获取医学相关的命名实体（如疾病、药物、诊断）以及这些实体之间的关系（如药物相对于疾病和患者的有效性或剂量、诊断的实验室和测试数据等）。因此，自动文本分析将构成从耶拿大学医院临床信息系统中包含的文件中计算医学相关上下文数据的基础，并将立即提供结构化评估程序，以便在进入常规实验室时实时选择定义明确的患者群体的样本。同时，不需要进一步诊断的剩余材料可用于建立比较样本库。这种用于德语临床文件的信息提取系统及其与常规临床工作流程的集成目前在任何德国医院都是不可用的。此外，我们认为，这样的联合努力将对转化医学的未来进展产生深远的影响，这将超越范例应用，以确定有效的表型数据，选择明确定义的患者群体，并生产高质量的生物材料。

The growing availability of high-quality biomaterials is a prerequisite for sustainable and reproducible results in translational biomedicine. This observation not only holds for exploratory but also, increasingly, for validating research outcomes. However, some skepticism has already been expressed whether the plethora of results from preclinical studies hold their promises when transferred into clinical practice.Consider, as an example, ongoing research on biomarkers. There is an apparent discrepancy between the multitude of studies on novel biomarkers and the number of clinically validated applications. One major problem is the lacking concern for quality differences in ancillary samples and the insufficient validation of potential markers based on comparison collectives with well-defined diseases and comorbidities differing from the target disease.Whereas in the past the infrastructure for high-quality collection and warehousing of such bio samples has been established at many clinical sites by building and maintaining professional biobanks, what is still lacking are routine workflows to sample valid phenotype data, to determine valid comparison collectives and to properly select samples for high-quality biobanking. In our project, we propose to extract such information from clinical documents using methods from automatic natural language processing. We plan to build a text analytics pipeline using semi-supervised machine learning techniques to harvest medically relevant named entities (such as diseases, drugs, diagnoses) and relations among these entities (such as the effectiveness or dosage of medications relative to a disease and a patient, lab and test data for diagnosis, etc.) from unstructured clinical documents (such as discharge summaries, radiology or pathology reports, etc.). Automatic text analysis will thus form the basis for computing medically relevant context data from the documents contained in the clinical information system of the university hospital in Jena and will instantaneously feed structured evaluation procedures for the real-time selection of samples of well-defined collectives of patients when they enter routine laboratories. At the same time, residual material not needed for further diagnosis can be utilized for building up a repository of comparison samples.Such an information extraction system for German-language clinical documents and its integration into routine clinical workflows is currently not available in any German hospital. Moreover, we stipulate that such a combined effort will have far-reaching implications for future progress in translational medicine which goes beyond the exemplary application to determine validated phenotype data, to select well-defined collectives of patients and to produce high-quality bio materials.