This work focusses on our rst integration steps of complex and partly unstructured medical data into a clinical research database. Our main application is an integrated facetted search tool in nephrology based on automatic information extraction results from textual documents. We describe the details of our technical architecture which is based on open-source tools|to be replicated at other universities, research institutes, or hospitals.
As medical records may cover a very long history of diseases (up to 30 years) and include a vast number of diagnoses, symptoms, results, medications, and laboratory values, we could highly bene t from advanced search capabilities in clinical information systems to allow for the retrieval of relevant data. However, medical information systems often su er from good search capabilities for data which has many unstructured text parts. Therefore, concepts to implement knowledge based systems, based on textual information extraction in medicine, are in focus of many recent research initiatives [11].
In this paper, we propose a three stage process: (
The facetted search application is based on the nephrology database TBase R . The web-based electronic patient record TBase R has been implemented in a German kidney transplantation programme as a cooperation between the Nephrology of Charite Universitatsmedizin Berlin and the AI Lab of the Institute of Computer Sciences of the Humboldt University of Berlin [3,10]. Currently, TBase R automatically integrates essential laboratory data (9.9 million values), clinical pharmacology (237.000 prescribed medications), diagnostic ndings from radiology, pathology and virology (146.000 ndings), and administrative data from the SAP-system of the Charite (70.000 diagnoses, 25.000 hospitalisations). Two groups of use cases for the application of facetted search in the medical eld, and nephrology, can be identi ed: rst, the use in clinical research, and second, the implementation in the individual treatment as a decision support system in the clinical routine.
Sacco [7] describes an approach of a guided interactive diagnostic system based an dynamic taxonomies. Biron et al. [2] describe an information retrieval system for computerised patient records We extend these approaches by a special multi-facet functionality. Our approach shows the following main advantages: { In our facetted search application, the user may remove any restriction he or she may have made in previous steps. This allows for a much better navigation through the search space where related systems only allow the subsequent thinning [8, 9]. { The ranking of facet values by cardinality supports the survey of remaining subsets. { We base automatically generated facets (e.g., disease/symptom relationships and negations) on multi-term extraction and relation extraction, by employing state-of-the-art, high-precision textual information extraction modules.
Only recently, new text mining approaches on Web-based medical literature have been proposed. For extracting adverse drug events from text [6] or automatic symptom extraction from texts on rare diseases [4], for example. However, clinical information extraction from patient records is still underrepresented and underdeveloped in clinical settings. Earlier work includes evaluating context features for medical relation mining on medical abstracts; the identi cation of semantic relations, such as substance A treats disease B, remains a non-trivial task [13]. Recent work and comparative baseline experiments include temporal information extraction [5]. A special trend becomes apparent, the need for ontology modelling of medical terminology and corresponding information extraction results [12]. Because of enormous annotation costs, mainly unsupervised methods are being used [1]. In industry and in the context of reliable clinical relevance, however, very detailed (and labor-intensive) supervised rule-based approaches represent the state-of-the-art.3 3 Here, we use our research project partner's solution (Averbis), which is based on shallow text parsing, see https://averbis.com/en/research/
The annotated texts are transferred in XMI format4 and stored in a local database at DFKI (see gure 1). Important components are the Solr search platform, the information extraction module, and the facetted search and presentation user interface modules. Solr5 is an open source enterprise search platform used in many large websites and applications and is one of the most popular enterprise search engines.6 Solr runs as a standalone full-text search server and uses the Lucene Java search library at its core for full-text indexing and (facetted) search. We chose the Solr system mainly because of some interesting features like facetted navigation, a query language that supports structured and textual search, the possibility for automatic result clustering based on Carrot27, its scalability and extensibility through plug-ins, and its various APIs for input (text, xml, JSON, etc.) and output (JSON, XML, PHP, python, etc.).
The rst step in our process is o ine informative extraction. The text data for our system originate from the TBase R database of Charite Berlin containing medical information about nephrology patients. In the rst phase we only 4 http://www.omg.org/spec/XMI/ 5 http://lucene.apache.org/solr/ 6 http://db-engines.com/en/ranking/search+engine 7 http://project.carrot2.org/ used about 5000 unstructured, free texts (no meta data or structured data of patients) of four types: 'Befunde" ( ndings), "Untersuchungen" (visits), "Entlassungsbriefe" (clinical reports), and "Verlaufe" (progress reports). These free texts are processed by the project partner Averbis, which anonymises the texts and adds annotations based on several medical reference systems and dictionaries (LOINC8, ICD109, ABDAMED10).
A software module extracts the relevant medical tags and features and stores
these in a database structure similar to the i2b2 star structure (in order to
simplify the updates of the target system i2b211). The user interface to search
and explore the annotated text database by using facets is built as a web
service based on the Solr extension "solarium" for PHP systems.12 This extension
provides for an API to specify all parameters necessary to create complex Solr
requests. The presentation/validation user interface of the system (see gure 2)
consists of two parts: the upper part shows the original text with highlighted
annotations, the lower part contains tabs listing the di erent relevant annotations.
Clicking on an item in the lower part scrolls the text above to the corresponding
8 https://loinc.org/
9 http://www.icd-code.de/
10 http://www.wuv-gmbh.de/abdata-pharma-daten-service/datenangebot/abdamed/
11 https://www.i2b2.org/about/intro.html
12 http://www.solarium-project.org/
position. The original XMI contents representing the complete original
annotation information is shown in a pop-up window when a highlighted annotation in
the text is clicked. Accordingly, this page serves two di erent purposes: (
We demonstrated that new facetted search applications in the use case of
transplant medicine in nephrology, based on open-source software tools and
exchangeable information extraction modules, are feasible and a very suitable
decisionsupport tool for the doctor: this type of a knowledge based system provides
physicians with a practicable tool for the analysis of medical data and decision
support for cohort selection. We developed a user interface for facetted search
which is based on the Solr Engine. In the next project phase, we will extend
the capabilities of the facetted search application, mainly including the
following aspects: (
This research is part of the project "clinical data intelligence" (KDI) which is founded by the Federal Ministry for Economic A airs and Energy (BMWi). 13 https://carrotsearch.com/foamtree-overview 5. Mkrtchyan, T., Sonntag, D.: Deep parsing at the CLEF2014 IE task. In: Working Notes for CLEF 2014 Conference, She eld, UK, September 15-18, 2014. pp. 138{ 146 (2014) 6. Odom, P., Bangera, V., Khot, T., Page, D., Natarajan, S.: Extracting adverse drug events from text using human advice. In: Arti cial Intelligence in Medicine - 15th Conference on Arti cial Intelligence in Medicine, AIME 2015, Pavia, Italy, June 17-20, 2015. Proceedings. pp. 195{204 (2015) 7. Sacco, G.: Guided interactive diagnostic systems. In: Computer-Based Medical
Systems. pp. 117{122 (2005) 8. Sacco, G.: Dynamic taxonomies and guided searches. Journal of the American
Society for Information Science and Technology 57(6), 792{796 (2006)
9. Sacco, G.: Dynamic taxonomies for intelligent information access. In:
KhosrowPour, M. (ed.) Encyclopedia of Information Science and Technology, pp. 3883{
3892. 3 edn. (2014)
10. Schroter, K.: Tbase2, a web-based electronic patient record. Fundamenta
Informaticae 43(