Data Resource Profile: The Cardiac Analytics and Innovation (CardiacAI) Data Repository

The CardiacAI Data Repository is a prospective, linked electronic medical record (EMR) data resource that was established to provide a digital health care and research ecosystem for cardiovascular care. The project was established in 2021 and includes cardiovascular patient EMR data from 1 January 2017 onwards, receiving a prospective nightly extract from two Local Health Districts in New South Wales, Australia, with integration with more Local Health Districts currently underway. EMR data are currently linked with other datasets including state-wide hospital and emergency department admissions and death data, and new datasets are actively being integrated including remote patient monitoring and electrocardiogram waveform data. The repository currently receives an average of 12,560 cardiovascular hospital admissions each year for an average of 9,215 patients with an average age of 68 years, of whom 60.5% are male, 59.2% are Australian born and 1.8% are Aboriginal or Torres Strait Islander. The CardiacAI project is not-for-profit, offering data access to ethically approved researchers on a cost-recovery basis. The repository will enable the rapid examination of contemporary cardiovascular health-related activity, patient characteristics, unmet clinical needs, pragmatic analyses of therapeutic interventions, development of patient risk stratification tools and personalized care innovations. Cardiovascular diseases (CVD) are the leading cause of death and hospital expenditure in Australia1,2 They significantly impact on the population, being the leading cause of disease burden by daily adjusted life-years.3 Managing patients with CVD is increasingly complex due to rising multimorbidity, long-term care requirements and a lack of coordinated care across health care providers.4 These challenges could be addressed more effectively by creating repositories of linked electronic medical record (EMR) data that provide streamlined access for clinical, research and quality improvement activities with close collaborations between clinicians, academia and policy makers. Not only would this have the potential to initiate a cycle of continuous evaluation and improvement, but it would also serve as an incubator for data-driven innovations that address critical needs, ultimately improving patient outcomes and overall efficiency of care.5,6 Realizing this goal is a significant challenge. Australia has a complex environment of public and private health care providers which lack a unified EMR system and are rarely linked.4,7 There are considerable privacy, security and governance requirements to satisfy when performing such large-scale data linkages.6,8 Data custodians and ethics committees are understandably cautious about approving such data linkages since they are responsible for ensuring that patient privacy is protected, research and quality improvement activities have appropriate oversight and data are used for the benefit of patients.6,9 Furthermore, there are significant technical challenges involved with linking EMR data sources, and establishing a secure environment for the data requires significant resources and ongoing funding that are not always available to small research teams. The Cardiac Analytics and Innovation (CardiacAI) Data Repository (Figure 1) is an independent, not-for-profit platform that provides a secure and appropriately governed environment for research groups to access and analyse prospective cardiovascular EMR data. The project, initiated by a group of senior cardiologists from South Eastern Sydney Local Health District (SESLHD) and Illawarra Shoalhaven Local Health District (ISLHD) and data scientists from the University of New South Wales (UNSW) Centre for Big Data Research in Health in 2021, has the objective of building a digital health care and research ecosystem for cardiovascular care. The Cardiac Analytics and Innovation (CardiacAI) Data Repository. CardiacAI Data Governance Committee (CardiacAI-DGC); machine learning tools (ML tools); artificial intelligence models (AI models); E-Research Institutional Cloud Architecture (ERICA) The project was established through in-kind funding from both SESLHD and UNSW and research grants. Ethics approval for the project has been granted by the SESLHD Human Research Ethics Committee and Population Health Services Research Ethics Committee. The project operates under an opt-out model whereby patients are informed about the project via information sheets and posters hung in cardiovascular wards and emergency departments. These have been translated into 20 languages and interpreter services are available on request. No patients have opted out of the study to date. Access to data is provided to researchers on a cost-recovery basis, thereby avoiding duplication of the lengthy data extraction and linkage processes already conducted by the CardiacAI research team, while allowing for the sustainability and improvement of the data repository. By providing the necessary research governance and data security infrastructure, and by fostering strong and mutually beneficial partnerships with data contributors, the project aims to expand on its success to become a comprehensive state-wide or national cardiovascular data asset which supports innovative and translational cardiovascular research. The project currently collects data prospectively on a nightly basis from hospitals in SESLHD and ISLHD, two of the 15 Local Health Districts (LHD) in New South Wales (NSW), Australia. The CardiacAI cohort comprises all individuals admitted to an SESLHD or ISLHD hospital on or after 1 January 2017, who fit at least one of the following criteria: admitted under a cardiovascular specialist (cardiologist, cardiothoracic surgeon, stroke physician, interventional neurologist or vascular surgeon); admitted to a cardiovascular ward (cardiac medical or surgical ward, stroke unit, coronary care unit, cardiothoracic intensive care); admitted, and died in the emergency department; discharged with a cardiovascular International Statistical Classification of Diseases and Related Health Problems, 10th Revision, Australian Modification (ICD-10-AM) or Australian Classification of Health Interventions code. To extract hospital EMR data, the team have developed custom database extraction scripts that can be easily shared with other Australian hospitals that have a Cerner Millennium10 EMR system, the predominant system used in Australian public hospitals.11 These EMR records are linked with state-wide population health datasets by the NSW Centre for Health Record Linkage including the Admitted Patient Data Collection, Emergency Department Data Collection, Registry of Birth, Deaths and Marriages and the Cause of Death Unit Record File.12 To provide an overview of the CardiacAI cohort characteristics, a snapshot of cardiac admissions between 1 January 2017 and 31 December 2020 are shown in Table 1. Three major cardiovascular subgroups are also provided, identified by their primary discharge ICD-10-AM code including acute coronary syndrome I20-I25, heart failure I50 and atrial fibrillation I48. The snapshot includes 50 241 admissions by 36 861 patients with an average of 12 560 admissions added to the repository each year. A list of the top five diagnosis ICD-10-AM and procedure Australian Classification of Health Interventions codes for the snapshot admissions are provided in Table 2. More details about the methods used can be found on pages 4–7 of the Supplementary Material, available as Supplementary data at IJE online. Cardiac Analytics and Innovation (CardiacAI) cohort characteristics ACS, acute coronary syndrome; AF, atrial fibrillation; CV, cardiovascular; HF, heart failure; ICU, intensive care unit; SD, standard deviation; Dx, diagnosis; AI, artificial intelligence; min, minimum; max, maximum; Dx, diagnosis. The denominator for categorical characteristics is total number of admissions for that cohort group. The denominator for the CardiacAI cohort admission year characteristic is the total number of admissions in the snapshot. The denominator for the sub-cohort (ACS, HF, AF) admission year characteristic is the total number of admissions in the CardiacAI cohort for that year. Cardiac Analytics and Innovation (CardiacAI) cohort characteristics ACS, acute coronary syndrome; AF, atrial fibrillation; CV, cardiovascular; HF, heart failure; ICU, intensive care unit; SD, standard deviation; Dx, diagnosis; AI, artificial intelligence; min, minimum; max, maximum; Dx, diagnosis. The denominator for categorical characteristics is total number of admissions for that cohort group. The denominator for the CardiacAI cohort admission year characteristic is the total number of admissions in the snapshot. The denominator for the sub-cohort (ACS, HF, AF) admission year characteristic is the total number of admissions in the CardiacAI cohort for that year. Top five primary diagnosis and procedure codes for Cardiac Analytics and Innovation (CardiacAI) admissions ACHI, Australian Classification of Health Interventions; ICD-10-AM, International Statistical Classifications of Diseases and Related Health Problems, 10th Revision, Australian Modification. The denominator is the total number of admissions in the snapshot for the CardiacAI cohort. Top five primary diagnosis and procedure codes for Cardiac Analytics and Innovation (CardiacAI) admissions ACHI, Australian Classification of Health Interventions; ICD-10-AM, International Statistical Classifications of Diseases and Related Health Problems, 10th Revision, Australian Modification. The denominator is the total number of admissions in the snapshot for the CardiacAI cohort. The CardiacAI team aims to incorporate new data sources iteratively by sharing the existing EMR extraction scripts or adapting the extraction scripts to new hospital EMR systems. The CardiacAI project is currently undergoing expansion to include electrocardiogram wave-form data and incorporating two additional LHDs in regional and metropolitan NSW, serving diverse indigenous and socioeconomic populations, which will effectively double the size of the repository. The repository is also being expanded to include cardiovascular remote patient monitoring data from the TeleClinical Care (TCC)13 application. This innovative application will provide regular biometric observations, medication schedules and patient-reported measures from CardiacAI acute coronary syndrome, heart failure and stroke patients in the community. The team is also actively exploring the inclusion of cardiac catheter procedure data, cardiac imaging data and community-based data sources such as pathology, radiology, cardiac rehabilitation, heart failure outreach and primary care data. A wide range of de-identified EMR data are available in the CardiacAI data repository. These include diagnoses and surgical procedures, de-identified, free-text clinical documents, pathology and medical imaging investigations, physiological observations, medication and blood product administrations and dates and times of admission and ward movements. The linked population health data sources provide administrative coded hospitalization, emergency visit and death data for all patients in the CardiacAI Data Repository and allow ascertainment of longitudinal outcomes. The types of data collected are shown in Table 3 and a full data dictionary can be found at [www.CardiacAI.org/data-dictionary]. More information describing the data can be found on page 3 of the Supplementary Material, available as Supplementary data at IJE online. Cardiac Analytics and Innovation (CardiacAI) data elements ACHI, Australian Classification of Health Interventions; NSW, New South Wales; SNOMED, Systematized Nomenclature of Medicine Clinical Terms. Currently does not include cardiac catheterization procedures. Currently does not include images or report text. Dataset linkage in progress at time of publication. Cardiac Analytics and Innovation (CardiacAI) data elements ACHI, Australian Classification of Health Interventions; NSW, New South Wales; SNOMED, Systematized Nomenclature of Medicine Clinical Terms. Currently does not include cardiac catheterization procedures. Currently does not include images or report text. Dataset linkage in progress at time of publication. To ensure accuracy and reliability of the EMR data, extensive validation is performed against the source systems. This involves rigorous testing of the data extracts to verify that what is being extracted accurately represents the data as they exist in the originating databases. The complex, encoded EMR data are automatically transformed into a research-ready data model composed of standard tables (such as the patient, encounter, diagnosis and medication administration tables), with well-defined primary and foreign keys, and with data elements using standard ontologies (e.g. SNOMED CT) when required. The EMR data are linked by the Centre for Health Record Linkage through probabilistic linkage techniques. Once the linked data are received by the CardiacAI Data Repository, they are harmonized and cleaned. This includes removing erroneous EMR records, removing patients with admissions occurring after their date of death, linking EMR admissions with their associated Admitted Patient Data Collection and Emergency Department Data Collection record, grouping transfer admissions into ‘episode sets’ where the patient has been transferred between facilities as a continuation of care for the same event and harmonizing date of death across the various datasets. All structured personal identifiers are removed from the data and kept in a secure location within the contributing organization to facilitate data linkage. However, personal identifiers may still be embedded within the free-text clinical documents including patients’ and doctors’ names and other personal identifiers. To address this issue, the CardiacAI project has established a process that automatically removes personal identifiers including person names, identification numbers, dates of birth, addresses and phone numbers from the text using a custom-built, de-identification algorithm that performs with high accuracy, removing person names with a recall score (positive predictive value) of 95.05.14 In addition, the de-identified clinical text is stored separately from the structured EMR data and its access is strictly controlled. Researchers must seek permission from the CardiacAI Data Governance Committee to access the clinical text and provide a compelling rationale for doing so. This approach reduces the number of people with access to the raw text, thereby reducing the risk of re-identification while still facilitating the conduct of research. Recent advances in cloud-computing and machine learning techniques are unlocking the potential of EMR data, paving the way for enhanced understanding and care of individuals with cardiovascular disease.15 The CardiacAI Data Repository, an enriched research-ready repository and data analytics platform, presents an opportunity for rapidly generating insights and data-driven tools for cardiovascular health care. Figure 2 outlines the five pillars of research activity envisioned for the CardiacAI Data Repository, including examples for each pillar. Cardiac Analytics and Innovation (CardiacAI) pillars of research activity As this is a new data resource, there have been no published studies that use the data to date. However, several projects are currently underway which include: development of a machine learning readmission and mortality predictive algorithm for heart failure patients using structured and unstructured EMR data; development of a heart failure clinical dashboard which delivers real-time insights for the management of heart failure patients in the hospital, which will be rolled out in SESLHD and ISLHD hospitals; an analysis of end-of-life care for heart failure patients in the hospital setting; the application of natural language processing techniques to the clinical text to extract cardiovascular concepts such as diagnoses, comorbidities and cardiovascular risk factors. The CardiacAI Data Repository incorporates data from LHDs servicing diverse populations in NSW, including those from culturally and linguistically diverse communities, indigenous peoples and individuals from wide-ranging socioeconomic backgrounds.16,17 Nonetheless, the repository currently only covers a portion of cardiovascular hospitalizations in NSW and lacks data from private hospitals, which represent nearly half of all cardiovascular hospitalisations.18 Furthermore, the repository currently does not include data from cardiac catheter procedure and diagnostic imaging data, or primary and community care datasets such as general practice and cardiac rehabilitation, which are important datasets to include for a holistic understanding of CVD management, but are more challenging to link due to being held in siloed or privatized databases. To address these limitations, the CardiacAI project team is iteratively expanding the repository to other datasets as funding becomes available. NSW currently lacks a state-wide EMR system, which makes implementing a state-wide cardiovascular data repository more challenging, as separate ethics and governance approvals and database extracts need to be established at each LHD. However, the CardiacAI project is making good progress to expand to more LHDs in NSW, with two LHDs already committed to joining the project. The CardiacAI team are also actively engaging with LHDs who will serve as the pilot sites for the new state-wide EMR system, commencing development in the coming years, future-proofing the CardiacAI project. The CardiacAI team are also looking to enrich the dataset of the existing cohort by linking with additional datasets including: primary care data through linkage with administrative datasets including the Australian Medical Benefits Scheme and Pharmaceutical Benefits Scheme; cardiovascular community health data through LHD-controlled cardiac rehabilitation and heart failure community outreach databases; cardiac imaging including computerized tomography and echocardiogram; cardiac catheter procedure data. Undoubtedly, EMR data are a valuable resource for the rapid generation of real-world evidence and the creation of tools to support clinical decision making. However, it is known to suffer from data quality issues.19,20 The CardiacAI dataset has undergone extensive testing to ensure it contains the correct information as recorded in the medical record, and a detailed and systematic assessment of its quality in terms of correctness, completeness, concordance, and plausibility is underway and will be reported in the near future. To facilitate the participation of CardiacAI data users in international studies, we also plan to extend our existing Australian-based data standardization procedure to the Observational Medical Outcomes Partnership Common Data Model,21 following previous work in this area.22,23 This is the lead common data model used internationally to support real-world evidence and it is maintained and used by the Observational Health Data Science and Informatics community.24 The team will also explore the validation of structured information against concepts extracted from clinical text, as well as applying other established tools that offer more detailed EMR data quality assessment features.23,25 By taking these steps, we aim to maximize the value of the CardiacAI dataset, ensuring the reliability and accuracy of its clinical information. The CardiacAI Data Repository is securely stored within the E-Research Institutional Cloud Architecture (ERICA),26, a highly secure cloud platform, which was designed for storing and using sensitive big health data and is managed by UNSW. ERICA has been accredited for this purpose by NSW under their and The platform has been to the requirements of the with access and oversight of the of data in and out of the More information about ERICA can be found on page 2 of the Supplementary Material, available as Supplementary data at IJE online. Researchers to use the CardiacAI Data Repository are to a project to the CardiacAI who will the of the request. it is the application will be to the CardiacAI Data Governance Committee for The comprises health data scientists and cardiovascular and data custodians from and a in data use projects that will cardiovascular understanding and and research and projects that will provide the building for innovations. on the projects may be under the existing CardiacAI ethics approval or be to separate ethics evaluation by the Human Research Ethics Committee. Researchers will only be provided with access to the data through are provided with an with and machine learning available for The is a with only sites of raw data is not and all research are by senior CardiacAI researchers for with the ethics approval and to from researchers to access the data within the CardiacAI platform may be to CardiacAI at More information about data sharing in the CardiacAI project can be found at This project was approved by South Eastern Sydney Local Health District Human Research Ethics Committee and by the New South Wales Population Health Services Research Ethics Committee The project to the Australian on in Human The data for this are available from the CardiacAI Data Repository. Data Supplementary data are available at IJE online. conducted the data and analysis for the CardiacAI cohort characteristics snapshot and the the CardiacAI project. the technical of the CardiacAI project, including the electronic medical record data extraction and CardiacAI data model and the of the analysis and of this and the data extracts and conducted the testing of the and the data extraction from their Local Health and provided and on the ethics data data and and and the The Cardiac Analytics and Innovation project is a between the South Eastern Sydney Local Health District Illawarra Shoalhaven Local Health District (ISLHD) and the University of New South Wales (UNSW) Centre for Big Data Research in who have provided in-kind funding for the project. support has been provided by the 2020 Medical Research Cardiovascular Health and the 2021 UNSW Cardiac and Medicine Big Analytics and Innovation data A big data resource for machine research and health quality The in this represent those of the and not the funding extend our to the South Eastern Sydney (SESLHD) and Illawarra Shoalhaven (ISLHD) Local Health Districts for their participation in and support for the CardiacAI project. also the of the Centre for Health Record Linkage in facilitating linkage of the EMR data with population health datasets.