Search

Introduction: The German Medical Text Project (GeMTeX) is one of the largest infrastructure efforts targeting German-language clinical documents. We here introduce the architecture of the de-identification pipeline of GeMTeX., Methods: This pipeline comprises the export of raw clinical documents from the local hospital information system, the import into the annotation platform INCEpTION, fully automatic pre-tagging with protected health information (PHI) items by the Averbis Health Discovery pipeline, a manual curation step of these pre-annotated data, and, finally, the automatic replacement of PHI items with type-conformant substitutes. This design was implemented in a pilot study involving six annotators and two curators each at the Data Integration Centers of the University Hospitals Leipzig and Erlangen., Results: As a proof of concept, the publicly available Graz Synthetic Text Clinical Corpus (GRASSCO) was enhanced with PHI annotations in an annotation campaign for which reasonable inter-annotator agreement values of Krippendorff's α ≈ 0.97 can be reported., Conclusion: These curated 1.4 K PHI annotations are released as open-source data constituting the first publicly available German clinical language text corpus with PHI metadata.

The extraction of medication information from unstructured clinical documents has been a major application of clinical NLP in the past decade as evidenced by the conduct of two shared tasks under the I2B2 and N2C2 umbrella. We here propose a new methodological approach which has already shown a tremendous potential for increasing system performance for general NLP tasks, but has so far not been applied to medication extraction from EHR data, namely deep learning based on transformer models. We ran experiments on established clinical data sets for English (exploiting I2B2 and N2C2 corpora) and German (based on the 3000PA corpus, a German reference data set). Our results reveal that transformer models are on a par with current state-of-the-art results for English, but yield new ones for German data. We further address the influence of context on the overall performance of transformer-based medication relation extraction.

We here report on one of the outcomes of a large-scale German research program, the Medical Informatics Initiative (MII), aiming at the development of a solid data and software infrastructure for German-language clinical natural language processing. Within this framework, we have developed 3000PA, a national clinical reference corpus composed of patient records from three clinical university sites and annotated with a multitude of semantic annotation layers (including medical named entities, semantic and temporal relations between entities, as well as certainty and negation information related to entities and relations). This non-sharable corpus has been complemented by three sharable ones (JSYNCC, GGPONC, and GRASCCO). Overall, 3000PA, JSYNCC and GRASCCO feature about 2.1 million metadata points.

We describe the creation of GRASCCO, a novel German-language corpus composed of some 60 clinical documents with more than.43,000 tokens. GRASCCO is a synthetic corpus resulting from a series of alienation steps to obfuscate privacy-sensitive information contained in real clinical documents, the true origin of all GRASCCO texts. Therefore, it is publicly shareable without any legal restrictions We also explore whether this corpus still represents common clinical language use by comparison with a real (non-shareable) clinical corpus we developed as a contribution to the Medical Informatics Initiative in Germany (MII) within the SMITH consortium. We find evidence that such a claim can indeed be made.

We describe the adaptation of a non-clinical pseudonymization system, originally developed for a German email corpus, for clinical use. This tool replaces previously identified Protected Health Information (PHI) items as carriers of privacy-sensitive information (original names for people, organizations, places, etc.) with semantic type-conformant, yet, fictitious surrogates. We evaluate the generated substitutes for grammatical correctness, semantic and medical plausibility and find particularly low numbers of error instances (less than 1%) on all of these dimensions.

We here describe the evolution of annotation guidelines for major clinical named entities, namely Diagnosis, Findings and Symptoms, on a corpus of approximately 1,000 German discharge letters. Due to their intrinsic opaqueness and complexity, clinical annotation tasks require continuous guideline tuning, beginning from the initial definition of crucial entities and the subsequent iterative evolution of guidelines based on empirical evidence. We describe rationales for adaptation, with focus on several metrical criteria and task-centered clinical constraints.

We devised annotation guidelines for the de-identification of German clinical documents and assembled a corpus of 1,106 discharge summaries and transfer letters with 44K annotated protected health information (PHI) items. After three iteration rounds, our annotation team finally reached an inter-annotator agreement of 0.96 on the instance level and 0.97 on the token level of annotation (averaged pair-wise F1 score). To establish a baseline for automatic de-identification on our corpus, we trained a recurrent neural network (RNN) and achieved F1 scores greater than 0.9 on most major PHI categories.

We introduce 3000PA, a clinical document corpus composed of 3,000 EPRs from three different clinical sites, which will serve as the backbone of a national reference language resource for German clinical NLP. We outline its design principles, results from a medication annotation campaign and the evaluation of a first medication information extraction prototype using a subset of 3000PA.

With the increasing availability of complete full texts (journal articles), rather than their surrogates (titles, abstracts), as resources for text analytics, entirely new opportunities arise for information extraction and text mining from scholarly publications. Yet, we gathered evidence that a range of problems are encountered for full-text processing when biomedical text analytics simply reuse existing NLP pipelines which were developed on the basis of abstracts (rather than full texts). We conducted experiments with four different relation extraction engines all of which were top performers in previous BioNLP Event Extraction Challenges. We found that abstract-trained engines loose up to 6.6% F-score points when run on full-text data. Hence, the reuse of existing abstract-based NLP software in a full-text scenario is considered harmful because of heavy performance losses. Given the current lack of annotated full-text resources to train on, our study quantifies the price paid for this short cut.

The automatic processing of non-English clinical documents is massively hampered by the lack of publicly available medical language resources for training, testing and evaluating NLP components. We suggest sharing statistical models derived from access-protected clinical documents as a reasonable substitute and provide solutions for sentence splitting, tokenization and POS tagging of German clinical texts. These three components were trained on the confidential FRAMED corpus, a non-sharable collection of various German-language clinical document types. The models derived therefrom outperform alternative components from OPENNLP and the Stanford POS tagger, also trained on FRAMED.

Creating and maintaining biomedical terminologies for multiple natural languages is a resource-expensive task, typically carried out by human domain experts. We here report on efforts to computationally support this process by treating term acquisition as a machine translation-guided classification problem capitalizing on parallel corpora. We report on experiments for French, German, Spanish and Dutch parts of a UMLS-derived terminology for which we generated 18 k, 23 k, 19 k and 12 k new terms and synonyms, respectively. Based on expert assessments of a novel German terminology segment about 80% of the newly acquired terms were judged as bio-medically reasonable and terminologically valid.

Terminologies which lack semantic connectivity hamper the effective search in biomedical fact databases and document retrieval systems. We here focus on the integration of two such isolated resources, the term lists from the protein fact database UNIPROT and the indexing vocabulary MESH from the bibliographic database MEDLINE. The generated semantic ties result from string matching and term set inclusion. We investigated the implicit terminological overlap between both resources in the domain of human proteins and evaluated our approach on a sample of 550 randomly selected UNIPROT entries that were manually mapped to their corresponding MESH headings. We achieved 90% precision and 79% recall (applying taxonomy-sensitive metrics). Fortunately, those proteins we were able to map to the MESH are ten times as frequently discussed in the literature as those on which we failed.

The Gene Regulation Ontology (GRO) is designed as a novel approach to model complex events that are part of the gene regulatory processes. We introduce the design requirements for such a conceptual model and discuss terminological resources suitable to base its construction on. The ontology defines gene regulation events in terms of ontological classes and imposes constraints on them by specifying the participants involved. The logical structure of the ontology is intended to meet the needs of advanced information extraction and text mining systems which target the identification of event representations in scientific literature. The GRO has just been submitted to the OBO library and is currently under review. It is available at http://www.ebi.ac.uk/Rebholz-srv/GRO/GRO.html.

Natural language processing of real-world documents requires several low-level tasks such as splitting a piece of text into its constituent sentences, and splitting each sentence into its constituent tokens to be performed by some preprocessor (prior to linguistic analysis). While this task is often considered as unsophisticated clerical work, in the life sciences domain it poses enormous problems due to complex naming conventions. In this paper, we first introduce an annotation framework for sentence and token splitting underlying a newly constructed sentence- and token-tagged biomedical text corpus. This corpus serves as a training environment and test bed for machine-learning based sentence and token splitters using Conditional Random Fields (CRFs). Our evaluation experiments reveal that CRFs with a rich feature set substantially increase sentence and token detection performance.

We propose an approach to multilingual medical document retrieval in which complex word forms are segmented according to medically relevant morpho-semantic criteria. At its core lies a multilingual dictionary, in which entries are equivalence classes of subwords, i.e. semantically minimal units. Using two different standard test collections for the medical domain, we evaluate our approach for six languages covered by our system.

We propose a method that aligns biomedical acronyms and their definitions across different languages. The approach is based upon a freely available tool for the extraction of abbreviations together with their expansions, and the subsequent normalization of language-specific variants, synonyms, and translations of the extracted acronym definitions. In this step, acronym expansions are mapped onto a language-independent concept-layer on which intra- as well as interlingual comparisons are drawn.

We propose a pipelined system for the automatic classification of medical documents according to their language (English, Spanish and German) and their target user group (medical experts vs. health care consumers). We use a simple n-gram based categorization model and present experimental results for both classification tasks. We also demonstrate how this methodology can be integrated into a health care document retrieval system.

SNOMED((R)) CT is emerging as a reference terminology for the entire health care process. It claims to be founded on logic-based modeling principles. In this work we analyze a special encoding scheme for SNOMED disease and procedure entities, the so-called relationship groups which had been devised in order to avoid ambiguities in entity definitions. We show that these artifacts may represent hidden mereological relations. We also report discrepancies encountered between the defined semantics of many SNOMED((R)) CT entity terms and their intuitive meaning, and inconsistencies detected between the definition of some complex composed entities and the definition of their top-level parents. As a result we formulate recommendations for improvements of SNOMED((R)) CT.

We present a method for the automated acquisition of a multilingual medical lexicon (for Spanish and Swedish) to be used within the framework of a medical cross-language text retrieval system. We incorporate seed lexicons and parallel corpora derived from the UMLS Metathesaurus. The seed lexicons for Spanish and Swedish are automatically generated from (previously manually constructed) Portuguese, German and English sources. Lexical and semantic hypotheses are then validated making iterative use of co-occurrence patterns of hypothesized translation synonyms in the parallel corpora.

We develop a formal framework by which clinical guidelines and protocols (CGPs) can be partially represented as a set of terminological concept definitions using standard description logics. There are two benefits in pursuing such an approach. First, it provides a foundation for logic-based CGP fusion and collision detection. Second, it allows for the checking of clinical treatment episodes from the EPR against CGPs.

We compare the performance of two part-of-speech taggers trained on a German newspaper corpus for mixed types of medical documents. TnT, a tagger based on a statistical language model, outperforms Brill's rule-based tagger, and supplied with additional lexicon resources matches state-of-the-art performance figures (close to 97% accuracy) on the medical corpus. We explain this unexpected result by focusing on the statistically significant part-of-speech type overlap between the newspaper training set and the medical test set. At least at that level, sublanguage differences seem to vanish. Thus, statistical off-the-shelf part-of-speech taggers can immediately be reused for medical language processing

In biomedical documents, there is ample evidence for complex morphological structures in specialized terms. While inflection is relatively easy to deal with, productive morphological processes such as derivation and single-word composition constitute a major challenge. Considering the problem from an information retrieval perspective, we split morphologically complex words into biomedically significant, morpheme-like subwords and match subwords the query terms and document terms are composed of. This way, morphologically motivated word form alterations can be eliminated from the retrieval procedure. Based on a series of retrieval experiments, we have gathered evidence that subword-based indexing and retrieval for the German biomedical sublanguage, at least--outperforms conventional string matching approaches.

Assisting users to search medical information on the WWW is here considered from two perspectives--the linguistic complexity of medical terms, nominal compounds in particular, and cross-lingual relationships between monolingual medical terminologies. In order to solve the first problem, we present an approach to automatic indexing in which medically plausible morphological units are used. Semantic value is added to these index terms by a compact domain-specific thesaurus. We further discuss tools for morphological segmentation and morphosemantic normalization of HTML documents, as well as an adaptation of a standard WWW search engine for morpheme-based retrieval. The second problem, cross-lingual medical document retrieval, is dealt with by defining cross-lingual equivalence relations on the emerging morpheme sets.

In order to transform a major portion of the Unified Medical Language System (UMLS) into a formally sound description logics system we have developed a four-step knowledge engineering approach. We report on experiments with a Loom knowledge base which consists of 164,000 concepts and 76,000 relations covering human anatomy and pathology. We discuss the fully automatized as well as interactive semiautomatic steps. The latter ones focus on the manual work necessary to render the formal target representation structures complete and adequate.

The semantic interpretation of natural language utterances is usually based on a large number of transformation rules which map syntactic structures (parse trees) onto some kind of meaning representation. However, those interpretation rules exhibit an insufficient degree of abstraction so that the scalability and portability of such natural language processing systems is hard to maintain. In this paper, we introduce an approach that is able to cope with a wide variety of semantic interpretation patterns in medical free texts by applying a small inventory of abstract semantic interpretation schemata. These schemata address generalized graph configurations within syntactic dependency parse trees, which abstract away from specific syntactic constructions.

We introduce a methodology for the segmentation of complex compounds into medically plausible morphemes. A tool for thesaurus compilation and management is presented, and design principles for a multilingual morpheme thesaurus are outlined. Our goal is to enhance the quality of medical free-text retrieval by replacing lexically based through morpheme-based search procedures.

The automatic analysis of medical full-texts currently suffers from neglecting text coherence phenomena such as reference relations between discourse units. This has unwarranted effects on the description adequacy of medical knowledge bases automatically generated from texts. The resulting representation bias can be characterized in terms of artificially fragmented, incomplete and invalid knowledge structures. We discuss three types of textual phenomena (pronominal and nominal anaphora, as well as textual ellipsis) and outline basic methodologies how to deal with them.