The objective of GRALENIA project is to develop a multidisciplinary, comprehensive and interoperable platform incorporating artificial intelligence algorithms and natural language processing techniques to improve the management of antimicrobial resistance (AMR) and reduce the impact of antimicrobial- or antibiotic-resistant microorganisms (aka superbugs) in hospitals. GRALENIA is supported through a Red.es grant for research and development projects in artificial intelligence and other digital technologies and their integration into value chains. Natural language processing, artificial intelligence, machine learning, deep learning, cloud services, healthcare, antimicrobial resistance, superbug, industrial research project
Over the last two decades, antimicrobial resistance
(AMR) has become a threat to public health systems
worldwide. World Health Organization (WHO)
considers this problem as one of its priorities [
Furthermore, the uncontrolled transmission of such superbugs1 between patients and healthcare workers poses a serious risk to the healthcare system.
0000-0003-0411-3056 (C. Bernardo-Castiñeira); 0000-00018837-0062 (G. Bou); 0000-0002-5233-3769 (M. Campos); 00000002-0777-0441 (B. Cánovas-Segura); 0009-0003-1680-9269 (S. Figueiras); 0000-0003-0752-8812 (C. Gómez-Rodríguez); 00000003-2941-1834 (J. Vilares) © 2023 Copyright for this paper by its authors. The use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
1 Superbugs is a name given to harmful bacteria that have acquired
resistance to one or more of the antibiotics used to treat them.
Appropriate use of antimicrobials is very complicated
because of the complexity of infectious diseases and
the spread of antibiotic resistance. Because of this, the
Spanish National Plan for Antibiotic Resistance [
There is a wealth of information that can be exploited to manage this serious problem, but the analysis of all this information is a very complex process. In many cases, such analysis is done manually, which results in a work overload for professionals, thus increasing hospital costs or, failing that, worsening health care work.
Artificial intelligence (AI) algorithms can improve interventions and decision making by ASP and healthcare-associated infection (HCAI) surveillance teams. The information held by the different services of a hospital is susceptible to be exploitable by AI algorithms for the development of intelligent AMR and HCAI control strategies, and for improving evidencebased decision making. However, the exploitation of such data by AI is very complex due to the heterogeneity of the data and its frequent lack of standardization and structuring. This fact has hindered the development of AI models for dealing with AMR and superbug infections, which are generally still in a laboratory/research phase.
In this context, GRALENIA constitutes an ambitious industrial research project that seeks to develop solutions based on the use of AI and natural language processing (NLP) techniques to exploit this information and provide innovative solutions to the AMR challenge.
The rest of the paper is organized as follows. Section 2 describes the project, its objectives, the lines of work involved in its development, the participant teams and their contribution. Section 3 presents the general architecture of the GRALENIA platform and describes the modules that comprise it, with the exception of the IA module, to be described in Section 4. Finally, Section 5 summarizes and closes the paper.
As a whole, the GRALENIA project2,3 aims to drive and accelerate digital transformation in healthcare, specifically in the area of AMR and superbug management. Hospital information systems contain a lot of information from various specialties directly or indirectly related to AMR occurrences, such as the possible inappropriate use of antibiotics (dosage information can be provided by the Pharmacy service) or diagnostic and characterization tests for superbugs (provided by the Microbiology service). Nevertheless, much of this information is not interrelated or standardized, making it difficult to exploit it automatically and to make decisions considering all the available information.
The GRALENIA project pursues the development of a multidisciplinary, comprehensive and interoperable platform that incorporates AI algorithms and NLP techniques to improve the digital management of AMR, and to reduce the impact of superbugs in hospitals. Starting from this general objective, the project has the following specific objectives: 1. The integration and standardization of relevant clinical data held by different hospital services. This allows its ulterior automatic processing. 2. The development of a base infrastructure for automated annotation of unstructured clinical documents using NLP techniques. 3. The prediction of the risk of superbug emergence in the hospital, and the identification of groups of patients with high susceptibility to superbug infection. 4. To improve the management and visualization of integrated data for AMR surveillance.
GRALENIA platform is being developed to be compatible with the data infrastructure of the A Coruña University Hospital Complex (CHUAC). This allows AI models to be designed using real-world information, thus providing the hospital with a digital solution to improve decision making regarding an existing problem.
To this end, GRALENIA not only focuses on R&D tasks for developing predictive models, but also 2 Spanish acronym for Gestión de Resistencias a Antimicrobianos basada en Lenguaje Natural e Inteligencia Artificial (Antimicrobial Resistance Management based on Natural Language and Artificial Intelligence). provide tools for structuring and standardizing both structured and unstructured information from various information silos. This makes it possible to integrate it as a whole, in a structured way, into a common data model on the platform. In turn, GRALENIA enables easy data visualization and exploitation by AI algorithms.
To achieve these ambitious objectives, considerable R&D effort is needed. Some of the main research lines involved in this project are: 1. To determinate which type of information, directly or indirectly related to AMR, is potentially exploitable. 2. To design an NLP-based system for the automatic extraction of AMR-related information from unstructured clinical documents. 3. Definition and development of advanced AI models for AMR-related risk prediction. 4. Validation of model results in a real-world environment.
In addition, it is also necessary to provide the system with the necessary mechanisms so that it can access information from the different hospital services relevant for AMR management.
These work lines are developed by a consortium of four participants.
Bahía Software.4 Coordinating team and industrial partner of the consortium. Founded by professionals with experience in the IT and Health & Healthcare sectors, the commercial activity of the company is mainly focused on e-Health, with important presence on e-Administration, industry and banking too. The company is recognized as an Innovative SME and as a national reference in medical image, tumor committees and surgical block management.
LYS Group (CITIC-UDC). The second organization involved in the project is the Language and Information Society Group (LYS)5 of the ICT Research Center (CITIC)6 of the University of A Coruña (UDC). LYS is an interdisciplinary research group formed by professors and researchers in AI and Linguistics, with extensive experience in the fields of NLP and Computational Linguistics. CITIC has an outstanding activity in the transfer of research results to society and industry. Based on its experience and expertise, the contribution of the LYS research team to the project will focus on the development of a base infrastructure for the automatic annotation of unstructured clinical information: • The NLP-based clinical Information
Extraction (IE) system for processing the 3 gralenia.es (visited on February 2024). 4 bahiasoftware.es (visited on February 2024). 5 www.grupolys.org (visited on February 2024). 6 citic.udc.es (visited on February 2024). 7 www.um.es/aike/ (visited on February 2024). 8 www.inibic.es (visited on February 2024). •
Co-design of the end-user tools (healthcare professionals). This guarantees the usability of tools and their algorithms, as well as the highest possible added value.
In addition, INIBIC coordinates the management of access to data and infrastructures with the Information Systems Service of CHUAC.
The user interface consists of a web interface, allowing a simpler and more accessible use. This reduces the learning curve of its handling and, therefore, makes it more pleasant to use. Additionally, the front-end also acts as a display window and information control panel for the Dashboard module (see Section 3.4). The front-end works as another service running within the GRALENIA cluster.
The data ingestion process of the GRALENIA platform includes several (sub)processes: 1. Extracting the data from the various information silos of the hospital. 2. Transforming the data to adapt it to the required format and structure. 9 aws.amazon.com (visited on February 2024). Microservice that connects to the Amazon S3 and allows both accessing to the information stored there and uploading new documents.
Due to CHUAC internal regulations, privacy and information security issues, and project deadlines, a data integration strategy through CSV data downloads has been chosen. Given the heterogeneous nature of both the information systems of the hospital and the variables required, specific extraction, anonymization and transformation processes have been developed to fit a common data model. The resulting data are dumped into CSV files to be ingested into the platform storage system.
However, our platform also contemplates the possibility of integration with hospital information systems using HL7 standards.10
GRALENIA platform implements a data lake. A data lake is a data repository used to store large amounts of both structured and unstructured data in its original format. Unlike traditional database systems, that require data to conform to a predefined schema, a data lake allows the ingestion of data from various raw sources and in various formats (text, images, sensor data, logs, etc.). The use of this data lake allows data from different information silos to be stored centrally and securely.
Despite being a centralized repository, a threelayer design has been used: 1. Landing zone. This is the entry point of the platform. In this first layer the data from the different information silos is stored in raw form, without any type of transformation. 2. Staging zone. An intermediate layer where validations and transformations are performed to adapt the data to the data model of the platform. 3. Refined zone. The last layer of the data lake. This is where the already structured and curated data is stored so that it can be consumed by the systems and tools that require it.
In the case of the first two layers, the landing and staging zones, an Amazon Simple Storage Service (S3) bucket11 is used. This is a versatile cloud object storage service that stands out for its scalability, security and accessibility. On the other hand, the last layer, the refined zone, uses the Amazon RDS service12 to host two relational databases: 1. An operational database to feed the various functionalities of the platform. The operational database is organized around hospital admissions, which uniquely identify each patient stay in a hospital. 2. An OMOP-CDM database to perform observational data analysis. The Observational Medical Outcomes Partnership (OMOP)13 is an open community standard, designed to provide a standardized way to represent data structure (i.e. a Common Data Model or CDM) and content (terminologies, vocabularies, coding scheme, etc.), and to enable efficient analyses that can produce reliable evidence.
The GRALENIA platform has a catalog of components oriented to the visualization of the information in an intuitive way so that it can be exploited by any user, from generalist profiles to specialists in data analysis. This catalog facilitates the correct assimilation of the data and, when integrated with the AI module, it also allows showing the results obtained by the algorithms.
The dashboard designed in GRALENIA makes it possible to quickly and easily analyze various Key Performance Indicators (KPI) potentially related to the emergence of superbugs in a hospital. The dashboard integrates a set of filters that can be configured by the user to run customized analyses such as, for example, studying the behavior of a specific superbug in a given time period and/or area of the hospital.
The back-end is the part of the platform that runs on the GRALENIA servers. We have decided to divide the functionalities of the system into several microservices. Thus, for each task of the platform, there is a single microservice with that single functionality. These microservices are described in Table 1. 10 www.hl7.org/implement/standards/ (visited on February 2024). 11 aws.amazon.com/s3/storage-classes/ (visited on February 2024). 12 aws.amazon.com/rds/ (visited on February 2024). 13 www.ohdsi.org/data-standardization/ (visited on February 2024).
In order to guarantee the security of a system that handles such sensitive data as this one, it is essential to use secure user and role control tools. For this purpose, the system uses Keycloak,14 an open source and widely used identity and access management (IAM) tool. Keycloak is deployed as another service running on the GRALENIA cluster.
The purpose of the AI module is to identify the triggers for the occurrence of multi-resistances. To this end, it integrates three systems: (1) an NLP-based IE system for clinical texts; (2) a system for predicting the risk of superbug outbreaks in different areas of the hospital; and (3) a system for identifying patient profiles with a high susceptibility to superbug infection.
The NLP subsystem of the IA module consists of an IE
system to identify and extract relevant clinical
information on free text sources such as electronic
health records (EHR), nursing notes, etc. [
The general architecture of the indication extractor
system is based on a pipeline, which gives it great
flexibility. Two different pipeline configurations are
provided:
1. A low-resource rule-based approach [
With the international market in mind, both approaches have been implemented for Spanish and English languages.
The IE system has been dockerized. for its final deployment to facilitate its integration with the rest of components of the platform.
The goal of this second component focuses on
predicting the risk of outbreaks in different areas of
the hospital at different times in the future [
The scope of the project has limited the analysis to the most prevalent superbugs (MRSA, Klebsiella ESBL and Klebsiella carbapenemase), as well as to the most vulnerable areas of the hospital, those where an outbreak of these characteristics could have catastrophic consequences: the intensive care unit and the recovery room. For each superbug, two approaches have been, although a possible third approach (based on time series) is also currently under consideration.
The first approach consists on predicting the risk of occurrence of at least one superbug. Classification algorithms are used for this purpose. A battery of alternatives including logistic regression, decision trees, random forest, support vector machines (SVM), boosting and artificial neural network (ANN) based algorithms are being evaluated.
The second approach corresponds to the prediction of the number of cases of superbug contagion. In this case, due to the discrete nature of the target variable, Poisson regression models are used.
The objective of this third and final system is to identify patient profiles with high susceptibility to trigger a superbug outbreak.
For the phenotyping algorithms, an approach
based on subgroup discovery has been selected [
As a whole, GRALENIA constitutes an industrial research project with high added value for the healthcare professionals of the A Coruña University Hospital Complex (CHUAC), who are facing such complex problems as antimicrobial resistance (AMR) and superbug management. GRALENIA project lays the groundwork for extending the technologies developed within it not only to other hospital services, but also to other hospitals and health services.
It is also worth noting that many of the tools and techniques based on artificial intelligence developed within the project (e.g. predictive and spatial models, natural language processing tools, etc.) lay the foundations for their adaptation to other clinical areas, further boosting the digital transformation of the healthcare sector.
This work is partially funded by the GRALENIA project (reference 2021/C005/00150055) supported by the Spanish Ministry of Economic Affairs and Digital Transformation, the Spanish Secretariat of State for Digitization and Artificial Intelligence, Red.es and by the NextGenerationEU funding.
Research on CITIC is also financed by the Xunta de Galicia through the collaboration agreement between the Ministry of Culture, Education, Vocational Training and Universities and the Galician universities to reinforce the System's research centers University of Galicia (CIGUS)
We would also like to thank Dr. José María Gutiérrez Urbón (CHUAC Pharmacy service), Dr. Laura Gutiérrez Fernández (CHUAC Infectious Diseases unit), Daniel Llamas Gómez (current Chief Information Officer-CIO, A Coruña-Cee Health District, SERGAS) and Guillermo Vázquez González (former CIO) for their valuable contributions to the project.