[The Medical Informatics Initiative at a glance-establishing a health research data infrastructure in Germany]. / Die Medizininformatik-Initiative im Überblick ­ Aufbau einer Gesundheitsforschungsdateninfrastruktur in Deutschland.

The Medical Informatics Initiative (MII) funded by the Federal Ministry of Education and Research (BMBF) 2016-2027 is successfully laying the foundations for data-based medicine in Germany. As part of this funding, 51 new professorships, 21 junior research groups, and various new degree programs have been established to strengthen teaching, training, and continuing education in the field of medical informatics and to improve expertise in medical data sciences. A joint decentralized federated research data infrastructure encompassing the entire university medical center and its partners was created in the form of data integration centers (DIC) at all locations and the German Portal for Medical Research Data (FDPG) as a central access point. A modular core dataset (KDS) was defined and implemented for the secondary use of patient treatment data with consistent use of international standards (e.g., FHIR, SNOMED CT, and LOINC). An officially approved nationwide broad consent was introduced as the legal basis. The first data exports and data use projects have been carried out, embedded in an overarching usage policy and standardized contractual regulations. The further development of the MII health research data infrastructures within the cooperative framework of the Network of University Medicine (NUM) offers an excellent starting point for a German contribution to the upcoming European Health Data Space (EHDS), which opens opportunities for Germany as a medical research location.

[Development of competencies in the Medical Informatics Initiative (MII)-educational offers for a competent and secure handling of medical data]. / Kompetenzentwicklung in der Medizininformatik-Initiative (MII) ­ Lehrangebote für einen souveränen und sicheren Umgang mit medizinischen Daten.

In order to achieve the goals of the Medical Informatics Initiative (MII), staff with skills in the field of medical informatics and data science are required. Each consortium has established training activities. Further, cross-consortium activities have emerged. This article describes the concepts, implemented programs, and experiences in the consortia. Fifty-one new professorships have been established and 10 new study programs have been created: 1 bachelor's degree and 6 consecutive and 3 part-time master's degree programs. Further, learning and training opportunities can be used by all MII partners. Certification and recognition opportunities have been created.The educational offers are aimed at target groups with a background in computer science, medicine, nursing, bioinformatics, biology, natural science, and data science. Additional qualifications for physicians in computer science and computer scientists in medicine seem to be particularly important. They can lead to higher quality in software development and better support for treatment processes by application systems.Digital learning methods were important in all consortia. They offer flexibility for cross-location and interprofessional training. This enables learning at an individual pace and an exchange between professional groups.The success of the MII depends largely on society's acceptance of the multiple use of medical data in both healthcare and research. The information required for this is provided by the MII's public relations work. There is also an enormous need in society for medical and digital literacy.

[FAIR health data in the national and international data space]. / FAIRe Gesundheitsdaten im nationalen und internationalen Datenraum.

Health data are extremely important in today's data-driven world. Through automation, healthcare processes can be optimized, and clinical decisions can be supported. For any reuse of data, the quality, validity, and trustworthiness of data are essential, and it is the only way to guarantee that data can be reused sensibly. Specific requirements for the description and coding of reusable data are defined in the FAIR guiding principles for data stewardship. Various national research associations and infrastructure projects in the German healthcare sector have already clearly positioned themselves on the FAIR principles: both the infrastructures of the Medical Informatics Initiative and the University Medicine Network operate explicitly on the basis of the FAIR principles, as do the National Research Data Infrastructure for Personal Health Data and the German Center for Diabetes Research.To ensure that a resource complies with the FAIR principles, the degree of FAIRness should first be determined (so-called FAIR assessment), followed by the prioritization for improvement steps (so-called FAIRification). Since 2016, a set of tools and guidelines have been developed for both steps, based on the different, domain-specific interpretations of the FAIR principles.Neighboring European countries have also invested in the development of a national framework for semantic interoperability in the context of the FAIR (Findable, Accessible, Interoperable, Reusable) principles. Concepts for comprehensive data enrichment were developed to simplify data analysis, for example, in the European Health Data Space or via the Observational Health Data Sciences and Informatics network. With the support of the European Open Science Cloud, among others, structured FAIRification measures have already been taken for German health datasets.

[Digital health progress hubs-data integration beyond university hospitals]. / Die digitalen Fortschrittshubs Gesundheit ­ Gemeinsame Datennutzung über die Universitätsmedizin hinaus.

The digital health progress hubs pilot the extensibility of the concepts and solutions of the Medical Informatics Initiative to improve regional healthcare and research. The six funded projects address different diseases, areas in regional healthcare, and methods of cross-institutional data linking and use. Despite the diversity of the scenarios and regional conditions, the technical, regulatory, and organizational challenges and barriers that the progress hubs encounter in the actual implementation of the solutions are often similar. This results in some common approaches to solutions, but also in political demands that go beyond the Health Data Utilization Act, which is considered a welcome improvement by the progress hubs.In this article, we present the digital progress hubs and discuss achievements, challenges, and approaches to solutions that enable the shared use of data from university hospitals and non-academic institutions in the healthcare system and can make a sustainable contribution to improving medical care and research.

Quantifying the Suitability of Biosignals Acquired During Surgery for Multimodal Analysis.

Goal: Recently, large datasets of biosignals acquired during surgery became available. As they offer multiple physiological signals measured in parallel, multimodal analysis - which involves their joint analysis - can be conducted and could provide deeper insights than unimodal analysis based on a single signal. However, it is unclear what percentage of intraoperatively acquired data is suitable for multimodal analysis. Due to the large amount of data, manual inspection and labelling into suitable and unsuitable segments are not feasible. Nevertheless, multimodal analysis is performed successfully in sleep studies since many years as their signals have proven suitable. Hence, this study evaluates the suitability to perform multimodal analysis on a surgery dataset (VitalDB) using a multi-center sleep dataset (SIESTA) as reference. Methods: We applied widely known algorithms entitled "signal quality indicators" to the common biosignals in both datasets, namely electrocardiography, electroencephalography, and respiratory signals split in segments of 10 s duration. As there are no multimodal methods available, we used only unimodal signal quality indicators. In case, all three signals were determined as being adequate by the indicators, we assumed that the whole signal segment was suitable for multimodal analysis. Results: 82% of SIESTA and 72% of VitalDB are suitable for multimodal analysis. Unsuitable signal segments exhibit constant or physiologically unreasonable values. Histogram examination indicated similar signal quality distributions between the datasets, albeit with potential statistical biases due to different measurement setups. Conclusions: The majority of data within VitalDB is suitable for multimodal analysis.

Treatment switches of disease-modifying therapies in people with multiple sclerosis: long-term experience from the German MS Registry.

Background: The spectrum of disease-modifying therapies (DMTs) for people with multiple sclerosis (PwMS) has expanded over years, but data on treatment strategies is largely lacking. DMT switches are common clinical practice. Objective: To compare switchers and non-switchers, characterize the first DMT switch and identify reasons and predictors for switching the first DMT. Methods: Data on 2722 PwMS from the German MS Registry were retrospectively analyzed regarding sociodemographic/clinical differences between 1361 switchers (PwMS discontinuing the first DMT) and non-switchers matched according to age, sex, and observation period. Frequencies of first and second DMTs were calculated and switch reasons identified. Predictors for DMT switches were revealed using univariable and multivariable regression models. Results: Switchers and non-switchers differed significantly regarding time to first DMT, education, calendar period of the first DMT start (2014-2017 versus 2018-2021), first DMT class used [mild-to-moderate efficacy (MME) versus high-efficacy (HE) DMT], time on first DMT, and disease activity at first DMT start or cessation/last follow-up. The majority of PwMS started with MME DMTs (77.1%), with the most common being glatiramer acetate, dimethyl/diroximel fumarate, and beta-interferon variants. Switchers changed treatment more often to HE DMTs (39.6%), most commonly sphingosine-1-phosphate receptor modulators, anti-CD20 monoclonal antibodies, and natalizumab. Fewer PwMS switched to MME DMTs (35.9%), with the most common being dimethyl/diroximel fumarate, teriflunomide, or beta-interferon. Among 1045 PwMS with sufficient data (76.8% of 1361 switchers), the most frequent reasons for discontinuing the first DMT were disease activity despite DMT (63.1%), adverse events (17.1%), and patient request (8.3%). Predictors for the first DMT switch were MME DMT as initial treatment [odds ratio (OR) = 2.83 (1.76-4.61), p < 0.001; reference: HE DMT], first DMT initiation between 2014 and 2017 [OR = 11.55 (6.93-19.94), p < 0.001; reference: 2018-2021], and shorter time on first DMT [OR = 0.22 (0.18-0.27), p < 0.001]. Conclusion: The initial use of MME DMTs was among the strongest predictors of DMT discontinuation in a large German retrospective MS cohort, arguing for the need for prospective treatment strategy trials, not only but also on the initial broad use of HE DMTs in PwMS.

Herding Cats in Pandemic Times - Towards Technological and Organizational Convergence of Heterogeneous Solutions for Investigating and Mastering the Pandemic in University Medical Centers.

To understand and handle the COVID-19 pandemic, digital tools and infrastructures were built in very short timeframes, resulting in stand-alone and non-interoperable solutions. To shape an interoperable, sustainable, and extensible ecosystem to advance biomedical research and healthcare during the pandemic and beyond, a short-term project called "Collaborative Data Exchange and Usage" (CODEX+) was initiated to integrate and connect multiple COVID-19 projects into a common organizational and technical framework. In this paper, we present the conceptual design, provide an overview of the results, and discuss the impact of such a project for the trade-off between innovation and sustainable infrastructures.

Ethical and coordinative challenges in setting up a national cohort study during the COVID-19 pandemic in Germany.

With the outbreak of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), global researchers were confronted with major challenges. The German National Pandemic Cohort Network (NAPKON) was launched in fall 2020 to effectively leverage resources and bundle research activities in the fight against the coronavirus disease 2019 (COVID-19) pandemic. We analyzed the setup phase of NAPKON as an example for multicenter studies in Germany, highlighting challenges and optimization potential in connecting 59 university and nonuniversity study sites. We examined the ethics application process of 121 ethics submissions considering durations, annotations, and outcomes. Study site activation and recruitment processes were investigated and related to the incidence of SARS-CoV-2 infections. For all initial ethics applications, the median time to a positive ethics vote was less than two weeks and 30 of these study sites (65%) joined NAPKON within less than three weeks each. Electronic instead of postal ethics submission (9.5 days (Q1: 5.75, Q3: 17) vs. 14 days (Q1: 11, Q3: 26), p value = 0.01) and adoption of the primary ethics vote significantly accelerated the ethics application process. Each study center enrolled a median of 37 patients during the 14-month observation period, with large differences depending on the health sector. We found a positive correlation between recruitment performance and COVID-19 incidence as well as hospitalization incidence. Our analysis highlighted the challenges and opportunities of the federated system in Germany. Digital ethics application tools, adoption of a primary ethics vote and standardized formal requirements lead to harmonized and thus faster study initiation processes during a pandemic.

Completing the Cabrera Circle: deriving adaptable leads from ECG limb leads by combining constraints with a correction factor.

Objective.We present a concept for processing 6-lead electrocardiography (ECG) signals which can be applied to various use cases in quantitative electrocardiography.Approach.Our work builds upon the mathematics of the well-known Cabrera sequence which is a re-sorting of the six limb leads (I,II,III,aVR,aVL,aVF) into a clockwise and physiologically-interpretable order. By deriving correction factors for harmonizing lead strengths and choosing an appropriate basis for the leads, we extend this concept towards what we call the 'Cabrera Circle' based on a mathematically sound foundation.Main results.To demonstrate the practical effectiveness and relevance of this concept, we analyze its suitability for deriving interpolated leads between the six limb leads and a 'radial' lead which both can be useful for specific use cases. We focus on the use cases of i) determination of the electrical heart axis by proposing a novel interactive tool for reconstructing the heart's vector loop and ii) improving accuracy in time of automatic R-wave detection and T-wave delineation in 6-lead ECG. For the first use case, we derive an equation which allows projections of the 2-dimensional vector loops to arbitrary angles of the Cabrera Circle. For the second use case, we apply several state-of-the-art algorithms to a freely-available 12-lead dataset (Lobachevsky University Database). Out-of-the-box results show that the derived radial lead outperforms the other limb leads (I,II,III,aVR,aVL,aVF) by improving F1 scores of R-peak and T-peak detection by 0.61 and 2.12, respectively. Results of on- and offset computations are also improved but on a smaller scale.Significance.In summary, the Cabrera Circle offers a methodology that might be useful for quantitative electrocardiography of the 6-lead subsystem-especially in the digital age.

How to Make Outpatient Healthcare Data in Germany Available for Research in the Dynamic Course of Digital Transformation.

INTRODUCTION: There is increasing interest on re-use of outpatient healthcare data for research, as most medical diagnosis and treatment is provided in the ambulatory sector. One of the early projects to bring primary data from German ambulatory care into clinical research technically, organizationally and in compliance with legal demands has been the RADAR project, that is based on a broad consent and has used the then available practice information system's interfaces to extract and transfer data to a research repository. In course of the digital transformation of the German healthcare system, former standards are abandoned and new interoperability standards, interfaces and regulations on secondary use of patient data are defined, however with slow adoption by Health-IT systems. Therefore, it is of importance for all initiatives that aim at using ambulatory healthcare data for research, how to access this data in an efficient and effective way. METHODS: Currently defined healthcare standards are compared regarding coverage of data relevant for research as defined by the RADAR project. We compare four architectural options to access ambulatory health data through different components of healthcare and health research data infrastructures along the technical, organizational and regulatory conditions, the timetable of dissemination and the researcher's perspective. RESULTS: A high-level comparison showed a high degree of semantic overlap in the information models used. Electronic patient records and practice information systems are alternative data sources for ambulatory health data - but differ strongly in data richness and accessibility. CONCLUSION: Considering the compared dimensions of architectural routes to access health data for secondary research use we conclude that data extraction from practice information systems is currently the most promising way due to data availability on a mid-term perspective. Integration of routine data into the national research data infrastructures might be enforced by convergence of to date different information models.

Reusing Biomedical Data as Agreed - Towards Structured Metadata for Data Use Agreements.

INTRODUCTION: With increasing availability of reusable biomedical data - from cohort studies to clinical routine data, data re-users face the problem to manage transferred data according to the heterogeneous data use agreements. While structured metadata is addressed in many contexts including informed consent, contracts are to date still unstructured text documents. In particular within collaborative and active working groups the actual usage agreement's regulations are highly relevant for the daily practice - can I share the data with colleagues from the same university or the same research network, can they be stored on a PHD student's laptop, can I store the data for further approved data usage requests? METHODS: In this article, we inspect and review seven different data usage agreements. We focus on digital data that is copied and transferred to the requester's environment. RESULTS: We identified 24 metadata items in the four main categories data usage, storage, and sharing, as well as publication of results. DISCUSSION: While the topics are largely overlap in the data use agreements, the actual regulations of the topics are diverse. Although we do not explicitly investigate trusted research environments, where data is offered within an analytics platform, we consider them a as subgroup, where most of the practical questions from the data scientist's perspective also arise. CONCLUSION: With a limited set of structured metadata items, data scientists could have information about the data use agreement at hand along with the transferred data in an easily accessible way.

Adaptability of Existing Feasibility Tools for Clinical Study Research Data Platforms.

INTRODUCTION: The increasing need for secondary use of clinical study data requires FAIR infrastructures, i.e. provide findable, accessible, interoperable and reusable data. It is crucial for data scientists to assess the number and distribution of cohorts that meet complex combinations of criteria defined by the research question. This so-called feasibility test is increasingly offered as a self-service, where scientists can filter the available data according to specific parameters. Early feasibility tools have been developed for biosamples or image collections. They are of high interest for clinical study platforms that federate multiple studies and data types, but they pose specific requirements on the integration of data sources and data protection. METHODS: Mandatory and desired requirements for such tools were acquired from two user groups - primary users and staff managing a platform's transfer office. Open Source feasibility tools were sought by different literature search strategies and evaluated on their adaptability to the requirements. RESULTS: We identified seven feasibility tools that we evaluated based on six mandatory properties. DISCUSSION: We determined five feasibility tools to be most promising candidates for adaption to a clinical study research data platform, the Clinical Communication Platform, the German Portal for Medical Research Data, the Feasibility Explorer, Medical Controlling, and the Sample Locator.

Impact of Clinical Study Implementation on Data Quality Assessments - Using Contradictions within Interdependent Health Data Items as a Pilot Indicator.

INTRODUCTION: Contradiction is a relevant data quality indicator to evaluate the plausibility of interdependent health data items. However, while contradiction assessment is achieved using domain-established contradictory dependencies, recent studies have shown the necessity for additional requirements to reach conclusive contradiction findings. For example, the oral or rectal methods used in measuring the body temperature will influence the thresholds of fever definition. The availability of this required information as explicit data items must be guaranteed during study design. In this work, we investigate the impact of activities related to study database implementation on contradiction assessment from two perspectives including: 1) additionally required metadata and 2) implementation of checks within electronic case report forms to prevent contradictory data entries. METHODS: Relevant information (timestamps, measurement methods, units, and interdependency rules) required for contradiction checks are identified. Scores are assigned to these parameters and two different studies are evaluated based on the fulfillment of the requirements by two selected interdependent data item sets. RESULTS: None of the studies have fulfilled all requirements. While timestamps and measurement units are found, missing information about measurement methods may impede conclusive contradiction assessment. Implemented checks are only found if data are directly entered. DISCUSSION: Conclusive contradiction assessment typically requires metadata in the context of captured data items. Consideration during study design and implementation of data capture systems may support better data quality in studies and could be further adopted in primary health information systems to enhance clinical anamnestic documentation.

Ten Topics to Get Started in Medical Informatics Research.

The vast and heterogeneous data being constantly generated in clinics can provide great wealth for patients and research alike. The quickly evolving field of medical informatics research has contributed numerous concepts, algorithms, and standards to facilitate this development. However, these difficult relationships, complex terminologies, and multiple implementations can present obstacles for people who want to get active in the field. With a particular focus on medical informatics research conducted in Germany, we present in our Viewpoint a set of 10 important topics to improve the overall interdisciplinary communication between different stakeholders (eg, physicians, computational experts, experimentalists, students, patient representatives). This may lower the barriers to entry and offer a starting point for collaborations at different levels. The suggested topics are briefly introduced, then general best practice guidance is given, and further resources for in-depth reading or hands-on tutorials are recommended. In addition, the topics are set to cover current aspects and open research gaps of the medical informatics domain, including data regulations and concepts; data harmonization and processing; and data evaluation, visualization, and dissemination. In addition, we give an example on how these topics can be integrated in a medical informatics curriculum for higher education. By recognizing these topics, readers will be able to (1) set clinical and research data into the context of medical informatics, understanding what is possible to achieve with data or how data should be handled in terms of data privacy and storage; (2) distinguish current interoperability standards and obtain first insights into the processes leading to effective data transfer and analysis; and (3) value the use of newly developed technical approaches to utilize the full potential of clinical data.

Insights into the FAIRness of the German Network University Medicine: A Survey.

The need to harness large amounts of data, possibly within a short period of time, became apparent during the Covid-19 pandemic outbreak. In 2022, the Corona Data Exchange Platform (CODEX), which had been developed within the German Network University Medicine (NUM), was extended by a number of common components, including a section on FAIR science. The FAIR principles enable research networks to evaluate how well they comply with current standards in open and reproducible science. To be more transparent, but also to guide scientists on how to improve data and software reusability, we disseminated an online survey within the NUM. Here we present the outcomes and lessons learnt.

Benchmarking the Impact of Noise on Deep Learning-Based Classification of Atrial Fibrillation in 12-Lead ECG.

Electrocardiography analysis is widely used in various clinical applications and Deep Learning models for classification tasks are currently in the focus of research. Due to their data-driven character, they bear the potential to handle signal noise efficiently, but its influence on the accuracy of these methods is still unclear. Therefore, we benchmark the influence of four types of noise on the accuracy of a Deep Learning-based method for atrial fibrillation detection in 12-lead electrocardiograms. We use a subset of a publicly available dataset (PTB-XL) and use the metadata provided by human experts regarding noise for assigning a signal quality to each electrocardiogram. Furthermore, we compute a quantitative signal-to-noise ratio for each electrocardiogram. We analyze the accuracy of the Deep Learning model with respect to both metrics and observe that the method can robustly identify atrial fibrillation, even in cases signals are labelled by human experts as being noisy on multiple leads. False positive and false negative rates are slightly worse for data being labelled as noisy. Interestingly, data annotated as showing baseline drift noise results in an accuracy very similar to data without. We conclude that the issue of processing noisy electrocardiography data can be addressed successfully by Deep Learning methods that might not need preprocessing as many conventional methods do.

The Importance of Being FAIR and FAST - The Clinical Epidemiology and Study Platform of the German Network University Medicine (NUKLEUS).

The COVID-19 pandemic has urged the need to set up, conduct and analyze high-quality epidemiological studies within a very short time-scale to provide timely evidence on influential factors on the pandemic, e.g. COVID-19 severity and disease course. The comprehensive research infrastructure developed to run the German National Pandemic Cohort Network within the Network University Medicine is now maintained within a generic clinical epidemiology and study platform NUKLEUS. It is operated and subsequently extended to allow efficient joint planning, execution and evaluation of clinical and clinical-epidemiological studies. We aim to provide high-quality biomedical data and biospecimens and make its results widely available to the scientific community by implementing findability, accessibility, interoperability and reusability - i.e. following the FAIR guiding principles. Thus, NUKLEUS might serve as role model for FAIR and fast implementation of clinical epidemiological studies within the setting of University Medical Centers and beyond.

Local Data Quality Assessments on EHR-Based Real-World Data for Rare Diseases.

The project "Collaboration on Rare Diseases" CORD-MI connects various university hospitals in Germany to collect sufficient harmonized electronic health record (EHR) data for supporting clinical research in the field of rare diseases (RDs). However, the integration and transformation of heterogeneous data into an interoperable standard through Extract-Transform-Load (ETL) processes is a complex task that may influence the data quality (DQ). Local DQ assessments and control processes are needed to ensure and improve the quality of RD data. We therefore aim to investigate the impact of ETL processes on the quality of transformed RD data. Seven DQ indicators for three independent DQ dimensions were evaluated. The resulting reports show the correctness of calculated DQ metrics and detected DQ issues. Our study provides the first comparison results between the DQ of RD data before and after ETL processes. We found that ETL processes are challenging tasks that influence the quality of RD data. We have demonstrated that our methodology is useful and capable of evaluating the quality of real-world data stored in different formats and structures. Our methodology can therefore be used to improve the quality of RD documentation and to support clinical research.

Towards a Consistent Representation of Contradictions Within Health Data for Efficient Implementation of Data Quality Assessments.

Contradictions as a data quality indicator are typically understood as impossible combinations of values in interdependent data items. While the handling of a single dependency between two data items is well established, for more complex interdependencies, there is not yet a common notation or structured evaluation method established to our knowledge. For the definition of such contradictions, specific biomedical domain knowledge is required, while informatics domain knowledge is responsible for the efficient implementation in assessment tools. We propose a notation of contradiction patterns that reflects the provided and required information by the different domains. We consider three parameters (α, ß, θ): the number of interdependent items as α, the number of contradictory dependencies defined by domain experts as ß, and the minimal number of required Boolean rules to assess these contradictions as θ. Inspection of the contradiction patterns in existing R packages for data quality assessments shows that all six examined packages implement the (2,1,1) class. We investigate more complex contradiction patterns in the biobank and COVID-19 domains showing that the minimum number of Boolean rules might be significantly lower than the number of described contradictions. While there might be a different number of contradictions formulated by the domain experts, we are confident that such a notation and structured analysis of the contradiction patterns helps to handle the complexity of multidimensional interdependencies within health data sets. A structured classification of contradiction checks will allow scoping of different contradiction patterns across multiple domains and effectively support the implementation of a generalized contradiction assessment framework.

Analysis of a Deep Learning Model for 12-Lead ECG Classification Reveals Learned Features Similar to Diagnostic Criteria.

Despite their remarkable performance, deep neural networks remain unadopted in clinical practice, which is considered to be partially due to their lack of explainability. In this work, we apply explainable attribution methods to a pre-trained deep neural network for abnormality classification in 12-lead electrocardiography to open this "black box" and understand the relationship between model prediction and learned features. We classify data from two public databases (CPSC 2018, PTB-XL) and the attribution methods assign a "relevance score" to each sample of the classified signals. This allows analyzing what the network learned during training, for which we propose quantitative methods: average relevance scores over a) classes, b) leads, and c) average beats. The analyses of relevance scores for atrial fibrillation and left bundle branch block compared to healthy controls show that their mean values a) increase with higher classification probability and correspond to false classifications when around zero, and b) correspond to clinical recommendations regarding which lead to consider. Furthermore, c) visible P-waves and concordant T-waves result in clearly negative relevance scores in atrial fibrillation and left bundle branch block classification, respectively. Results are similar across both databases despite differences in study population and hardware. In summary, our analysis suggests that the DNN learned features similar to cardiology textbook knowledge.