Ebook: Intelligent Health Systems – From Technology to Data and Knowledge

The 35th Medical Informatics Europe Conference, MIE 2025, was held in Glasgow, Scotland, UK, from the 19–21 May 2025. The Conference was co-hosted by the British Computer Society (BCS) and the European Federation for Medical Informatics (EFMI). The Scientific Programme Committee was co-chaired by Dr. Elisavet Andrikopoulou and Assistant Professor Parisis Gallos.

The theme of MIE 2025 was ‘Intelligent Health Systems – From Technology to Data and Knowledge’. In an era in which artificial intelligence (AI) is once again being over-hyped as a panacea, the purpose of this theme serves to emphasise that technology is a necessary but insufficient dependency for intelligent health systems. Technology must not only provide functionality for healthcare processes, but also enable high quality data capture and analysis. Whether structured or unstructured, data does not of itself provide unmediated insights, but requires the input of skilled humans who understand context and purpose and can use analytics to draw meaningful conclusions to effect improvement. Iteratively, we can derive knowledge from the synthesis of operational data and research evidence that will become the basis for better practice as part of a learning health system.

These proceedings present the current trends in health informatics. Contributions cover topics like learning health systems, education, data science and visualization, information systems, human factors, social and organisational issues, computable knowledge and decision support, AI, automation and robotics in healthcare, natural language processing and generation, personal health records and mHealth, population health, precision care, one health, one digital health and climate changes, and other topics related to health informatics.

The proceedings are published by IOS Press as an e-book in the open access series Studies in Health Technology and Informatics. Volumes in the Studies in Health Technology and Informatics series are submitted (for evaluation) for indexing by MEDLINE/PubMed; Web of Science: Conference Proceedings Citation Index – Science (CPCI-S) and Book Citation Index – Science (BKCI-S); Google Scholar; Scopus, and EMCare.

The Editors,

Elisavet Andrikopoulou, Parisis Gallos, Theodoros N. Arvanitis, Rosalynn Austin, Arriel Benis, Ronald Cornet, Panagiotis Chatzistergos, Alexander Dejaco, Linda Dusseljee-Peute, Alaa Mohasseb, Pantelis Natsiavas, Haythem Nakkas, and Philip Scott

Glasgow, 1 April 2025

In the global healthcare landscape, the accessibility and standardization of medical terminologies across different languages are crucial. HealthTermFinder is a tool designed to harness the Unified Medical Language System (UMLS) for the multilingual discovery of standardized healthcare terms. Our tool explores the relationships between atomic units of UMLS terminology sourced from various databases. We have mapped up to 24 languages, including Arabic, French, German, and Spanish, to SNOMED CT and HPO. Furthermore, we mapped clinical features for three German case studies using HealthTermFinder and achieved a total coverage of 82 % with high accuracy. HealthTermFinder aims to overcome language barriers in accessing standardized healthcare terminologies, facilitating global interoperability and data sharing in the healthcare sector.

One goal of the Fast Healthcare Interoperability Resources (FHIR) standard is to prevent semantic ambiguities when patient data are electronically exchanged. To assess whether the FHIR specifications live up to this expectation, we examined FHIR’s Condition resource thereby focusing on the resource elements ‘clinical status’ and ‘verification status’ when used in combination. We found that the definitions for these elements, as well as of several of their allowed values, suffer from the following semantic difficulties: the use of disjunctive descriptions, the presence of pseudo-synonyms lacking clear explanation of what distinctions FHIR has in mind, and insufficient discrimination between evidence and what such evidence would be about.

Certified registered nurse anesthetists (CRNAs) rely on various information sources to make clinical decisions. Clinical decision support systems (CDSS) can enhance decision-making, but few are tailored specifically for CRNAs. This rapid review examines the current evidence on CDSS use in anesthesia care. A rapid review was conducted following PRISMA guidelines. Databases Medline CINAHL and Embase were searched for studies focusing on CDSS for CRNAs. Six studies were included. No CDSS was designed specifically for CRNAs. The identified systems, both rule-based and AI-based, improved decision-making, particularly in risk prediction and intraoperative management, but were not adapted to the specific needs of CRNAs. While CDSS can enhance anesthesia care, systems designed for other professionals may not fully meet CRNA requirements. Tailored CDSS are needed to address their unique decision-making processes. Developing CRNA-specific CDSS could optimize decision-making and improve patient outcomes in anesthesia practice.

SNOMED CT is a comprehensive controlled biomedical ontology widely used as an information exchange standard among various healthcare institutions. To ensure the unambiguous expression of health data and effective linguistic computation of word meanings, the hierarchical relation of a partially antonymous biomedical concept pair, which shares a common context but has antonymous modifiers such as in magnetic resonance imaging without contrast — magnetic resonance imaging with contrast, must be validated. This study examined the hierarchical matchings of partially antonymous concepts by the prepositional phrase without in SNOMED CT’s hierarchy. Among 132 and 477 partially antonymous pairs, 10 (7.6%) and 35 (7.3%) pairs were undesirably located under the Procedure and Disease hierarchies, respectively. Auditing efforts need to address partially antonymous concept pairs in SNOMED CT to provide a more reliable representation of semantic relations.

We analyzed the integration of differential privacy into data synthesis for survival analyses, focusing on the trade-off between privacy protection and model accuracy. The dataset of lung cancer patients from Germany was synthesized using CTAB-GAN+. For survival analyses, the CoxPH and DeepSurv models were applied. Missing values were imputed with Miss Forest or treated as a category; in case of CoxPH categorical variables were label and one-hot encoded. Our findings show that privacy budgets significantly affect accuracy, but model choice and data preprocessing also lead to improvements of up to 4.5%. With differential privacy, the CoxPH model using Miss Forest imputation and one-hot encoding achieved a concordance index of over 0.68.

The role played by physical activity in slowing down the progression of type-2 diabetes is well recognized. However, except for general clinical guidelines, quantitative real-time estimates of the recommended amount of physical activity, based on the evolving individual conditions, are still missing in the literature. The aim of this work is to provide a control-theoretical formulation of the exercise encoding all the exercise-related features (intensity, duration, period). Specifically, we design a feedback law in terms of recommended physical activity, following a model predictive control approach, based on a widespread compact diabetes progression model, suitably modified to account for the long-term effects of regular exercise. Preliminary simulations show promising results, well aligned with clinical evidence. These findings can be the basis for further validation of the control law on high-dimensional diabetes progression models to ultimately translate the predictions of the controller into meaningful recommendations.

Rare diseases are challenging to diagnose and collectively affect a large fraction of the population. This work sought to develop an approach to generate models for probabilistic reasoning focused on the presence of a specified phenotypic abnormality. The approach generates a Bayesian network, a graphical AI model that uses probability to reason under uncertainty, that includes all diseases that can cause the specified abnormality as well as all phenotypic abnormalities caused by those diseases. The approach efficiently computes the probabilities of the possible diagnoses and evaluates the impact of additional evidence. One can use the model to identify the observations that yield the greatest information to reduce uncertainty. An example model for diagnosis of a finding of enlarged kidney is presented to demonstrate the feasibility and advantages of the approach. Further work includes incorporation of age of onset and inheritance pattern of the diseases, hierarchical relationships among diseases and phenotypic abnormalities to allow diagnosis based on information at varying levels of granularity, and user interfaces to simplify interaction with the models.

Multimorbidity is increasingly prevalent as the population ages and individuals with multiple long-term conditions (MLTCs) live longer. Often each condition is treated by a separate clinician, which can lead to harmful drug-drug and drug-disease interactions. Artificial Intelligence (AI) can help to identify those at risk of poor outcomes, which is particularly difficult when managing frail patients with MLTCs. We aim to analyze MLTC trajectories, but there is limited work on clustering using Electronic Health Records (EHR) and they largely ignore the time elapsed between clinical events. In this work we adapted three machine learning methods (Word2Vec, Autoencoder, TG-CNN) to allow us to cluster the entire trajectory. The proposed methods are tested on the ACT-MOOC dataset as a proof of concept, using timelines of interaction with an online course instead of an EHR. We find that clustering using the novel TG-CNN approach, which accounts for the time between events, shows a clear separation among different clusters and is better able to represent patient/user trajectories. In future work, we will apply this methodology within CPRD as part of the NIHR-funded DynAIRx project.

Prostate cancer is a leading cause of cancer-related deaths, with Gleason grading being key for assessing tumor aggressiveness. We propose a Graph Neural Network-based approach to automate Gleason grading using the Automated Gleason Grading Challenge 2022 dataset. Patch-level graphs constructed from Hematoxylin and Eosin-stained Whole-Slide Images were classified into Gleason grades. Our results show that Graph Neural Networks, specifically Graph Attention Networks and Graph Convolutional Networks, effectively distinguish between grades despite class imbalance. Focal Loss improves the classification of the minority Gleason Grade 5, which is crucial for detecting aggressive prostate cancer. Our models outperform state-of-the-art methods, achieving higher F1-scores without scanner generalization techniques.

During the 68th annual conference of the German Association for Medical Informatics, Biometry and Informatics, the Working Group Medical Terminologies and Classifications held a terminology server challenge, investigating the boundary between general-purpose FHIR servers and purpose-built FHIR terminology servers. While direct comparisons between the implementations were not the goal of this challenge, it showed that such a boundary exists: General-purpose FHIR servers need to consider many different domains of the FHIR standard, and generally aren’t optimized for the very different terminology use cases.

Digital pathology has made significant advances in tumor diagnosis and segmentation; however, image variability resulting from tissue preparation and digitization - referred to as domain shift - remains a significant challenge. Variations caused by heterogeneous scanners introduce color inconsistencies that negatively affect the performance of segmentation algorithms. To address this issue, we have developed a joint multitask U-net architecture trained for both segmentation and stain separation. This model isolates the stain matrix and stain density, allowing it to handle color variations and improve generalization across different scanners. On 180 stain images from three different scanners, our model achieved a Dice score of 0.898 and an Intersection Over Union (IoU) score of 0.816, outperforming conventional supervised learning methods by +1.5% and +2.5%, respectively. On external datasets containing images from six different scanners, our model averaged a Dice score and IoU of 0.792. By leveraging our novel approach to stain separation, we improved segmentation accuracy and generalization across diverse histopathological samples. These advances may pave the way for more reliable and consistent diagnostic tools for breast adenocarcinoma.

The cut-off on an ordinal test is often determined by its ‘diagnostic accuracy’, measured solely by its Sensitivity and Specificity in relation to a reference standard. This involves selecting the cut-off that maximises their combination, as calculated by measures, such as Youden’s J statistic, that treat False Positives and False Negatives as equally important. Prevalence-adjusted Predictive Values Positive and Negative at each possible cut-off are often calculated, but the set of complementary False Alarm Rates and False Reassurance Rates are left implicit. The False Alarms per False Reassurance Number, which we refer to as the FARN, represents the error rate trade-off embedded at each possible cut-off. The routine display of the full set of FARNs in a dataset would make transparent the preference-sensitivity of cut-off selection and virtually mandate the exploration of the quantitative relative disutility of a FA and a FR essential to establishing the cut-off appropriate in the given decision making context. We link to a prototype generic online calculator that instantly reveals the FARNs implicit in a research dataset. The GAD-7 test for the detection of Generalised Anxiety Disorder provides our empirical illustration. Focusing on the ‘accuracy’ of ordinal diagnostic or screening tests threatens the pursuit of therapeutic optimality.

In Spain, traumatic brain injury (TBI) affects 200 new cases per 100,000 inhabitants, making it a significant cause of morbidity and mortality globally and a frequent reason for emergency department attention. This study addresses the lack of consensus on the use of diagnostic tests to exclude intracranial lesions, leading to overuse of Computed Tomography (CT) scans. To this end, we propose to improve the structuring and standardisation of information on Traumatic Brain Injury (TBI) in the Electronic Health Record (EHR) using terminologies and standards. To this end, a unification of the TBI diagnostic code, the creation of a formulary for the structured collection of clinical observations and the development of a dashboard in Power Business Inteligence (PowerBI) are carried out. Results revealed a cohort of 442 TBI cases and reduced data search time, improving management efficiency and real-time indicator monitoring. This approach enhances clinical care, promotes data interoperability, and supports TBI research.

A significant risk following a kidney transplantation is graft loss. The Screen Reject Project has developed a Clinical Data Warehouse (CDWH) as a foundation for a clinical decision support system designed to improve the diagnosis of graft rejections. The CDWH integrates patient data and event records of n = 141 kidney transplant patients. These data are not directly comparable within the cohort as they consist of irregular time series, particularly of laboratory values. Therefore, a pre-processing routine was developed which divides a relative time window before the last biopsy (the relevant end event of the reference period for subsequent machine learning procedures) into equal time intervals for each patient. For each of these intervals a representative value is calculated from the contained laboratory values. These representative values are used to train models for predicting kidney rejection. The comparison with an existing study from the project, in which a classification model was developed without considering the temporal dependencies, shows an improved sensitivity and specificity in predicting kidney rejection for the harmonised data using the same random forest model.

This study presents the development of a customizable rule-based engine, integrated into the LETHE Clinical Trial Management System (CTMS), designed to enhance clinical decision-making in a multidomain intervention for dementia risk reduction. The rule engine allows health professionals to create and personalize rules based on patient-specific profiles and evolving clinical knowledge, enabling more precise interventions. The system was developed through a collaborative approach with one health professional and employs a JSON based rules engine for flexible rule creation and implementation. Its application in the LETHE project, a randomized control trial involving 156 participants across multiple centers, demonstrates the potential of tailored digital interventions in improving clinical outcomes. Future enhancements will focus on broadening the system’s usability and automating patient-specific tasks to streamline care.

We present a system to aid humane endpoint decisions in laboratory rodents. The system derives the future probability distribution of any quantified parameter suitable as endpoint criterion. This provides experimenters with knowledge about the likelihood of a criterion exceeding a fixed threshold. Researchers thus gain valuable information about the individual progression of the recorded parameters, to reduce distress and avoidable loss of animals. We examine the system on the data of a xenograft model using body weight as an endpoint criterion.

Migraine is a common chronic headache disorder characterizsed by episodes of moderate to severe headaches, resulting in a large personal- and societal burden. To address this, we implemented a mobile app solution aimed at enhancing continuous data collection and increasing data coverage for the Empatica E4 biometric sensor device. Our ultimate goal is to use this system in a future migraine event prediction system. In our initial study, three participants wore the E4 device for eight days and were interviewed about their experience. Main user-experience feedback included the need for more frequent reminders, more detailed information on the collected data, and improvements with respects to connectivity issues. Regarding the app interface, participants recommended adding more diagnostic tools and statistics, enhancing the “look and feel” design, clarifying explanations of variables, providing information on data usage, and supporting additional sensor devices.

Manual segmentation of histopathological images is both resource-intensive and prone to human error, particularly when dealing with challenging tumor types like Glioblastoma (GBM), an aggressive and highly heterogeneous brain tumor. The fuzzy borders of GBM make it especially difficult to segment, requiring models with strong generalization capabilities to achieve reliable results. In this study, we leverage the Medical Open Network for Artificial Intelligence (MONAI) framework to segment GBM tissue from hematoxylin and eosin-stained Whole-Slide Images. MONAI performed comparably well to state-of-the-art AutoML tools on our in-house dataset, achieving a Dice score of 79%. These promising results highlight the potential for future research on public datasets.

Hospital-acquired pressure injuries (HAPIs) are common complications that impact patient outcomes and strain healthcare resources. The Braden Scale is the standard tool for assessing HAPI risk, but it has limitations, including a high false-positive rate, potential oversight of subtle symptoms, and added workload for nurses. To address these issues, a fully automated AI clinical decision support system (CDSS) achieving 0.90 AUROC on retrospective data has been deployed.

Molecular tumor boards present special challenges when it comes to information collection for case preparation. It is one of the most time-consuming tasks participating pathologists and oncologists face, limiting the number of cases that can be discussed in these specialized tumor boards and in turn can profit from a potential highly personalized therapy. Digital support is a necessity to enable medical professionals to efficiently make use of the vast amount of data available for each patient and their genomic and clinical profile. This includes historically recommended therapies for patients with molecularly similar tumors. To combat this issue, we developed an extension for the MTB-cBioPortal in collaboration with clinicians, enabling users to access previously documented therapy recommendations combined with corresponding follow-up data based on HL7 FHIR profiles and modules established in the Medical Informatics Initiative (MII). The information is made available through an additional annotation in the MTB-cBioPortal patient view. In doing so we intend to improve the efficiency of the case preparation process for molecular tumor boards and lay the groundwork for a potential multicentric exchange of therapy recommendations and follow-up data.

This study explores the potential for estimating heartbeats and heart rate variability (HRV) parameters using multiple multi-axis ballistocardiographic (BCG) sensors in recording disturbed by speech interference. The results demonstrate that an adaptive approach, which detects and interpolates J-peaks in disturbed signal parts, provides more accurate heartbeat evaluations than relying on any single BCG channel in the same recording.

Clinical decision support systems (CDSSs) are designed to enhance patient safety by providing alerts to prescribers about potential medication issues. However, a significant proportion of these alerts are ignored, which can compromise patient safety. This study explores the feasibility of using subgroup discovery, a machine learning method, to identify determinants influencing physicians’ medication-related CDSS alert handling. By analyzing CDSS log data from the electronic health record, this research shows the feasibility of the use of subgroup discovery on this data, and its potential to uncover behavioral patterns and factors that affect how alerts are managed. This can ultimately contribute to the design of more effective CDSS alerts and improving patient safety.

Due to demographic change, health economics is increasingly focused on the quality of life in advanced age and the associated cost aspects. Dementia is one of the key issues in this area and its efficient treatment will become increasingly relevant in the near future. Our system aims to automatically create treatment plans for dementia patients in a digital platform. To this end, three algorithms were implemented: a rule-based approach, an approach based on artificial neural networks, and an approach using large language models. A comprehensive synthetic dataset with fictitious patients, medical problems, and treatments was created for training and evaluation of the algorithms.