Cardiovascular Care Innovation through Data-Driven Discoveries in the Electronic Health Record

Abstract

The electronic health record (EHR) represents a rich source of patient information, increasingly being leveraged for cardiovascular research. While its primary use remains the seamless delivery of healthcare, various longitudinally aggregated structured and unstructured data elements for each patient within the EHR can define the computational phenotypes of disease, and care signatures and their association with outcomes. While structured data elements, such as demographic characteristics, laboratory measurements, problem lists, and medications, are easily extracted, unstructured data is underutilized. The latter include free text in clinical narratives, documentation of procedures, and reports of imaging and pathology. Rapid scaling up of data storage, and rapid innovation in natural language processing and computer vision can power insights from unstructured data streams. However, despite an array of opportunities for research using the EHR, specific expertise is necessary to adequately address confidentiality, accuracy, completeness, and heterogeneity challenges in EHR-based research. These often require methodological innovation and best practices to design and conduct successful research studies. Our review discusses these challenges and proposed solutions for these challenges. Additionally, we highlight ongoing innovations in federated learning in the EHR through greater focus on common data models, and discuss ongoing work that defines such an approach to large-scale, multi-center, federated studies. Such parallel improvements in technology and research methodology enable innovative care and optimization of patient outcomes.

CURRENT STATE OF THE EHR IN CARDIOVASCULAR RESEARCH

Electronic health records (EHRs) represent electronic platforms that collect, store, and present clinical data linked longitudinally for each patient. The primary purpose of the EHR is to allow for the seamless delivery of healthcare.¹ However, EHR data can provide abundant sociodemographic and clinical data that can potentiate and galvanize clinical research, spanning investigations into the epidemiology of disease, healthcare resource utilization, disease phenotyping, outcome prediction, pragmatic trials, and assessments of comparative effectiveness and safety of treatments.^2–6

The history of EHRs dates to the 1970s, when they were originally designed to support health insurance and billing.^1,7 A series of investigations suggested additional benefits in patient safety with the use of EHRs, which led to the National Academy of Medicine (then known as the Institute of Medicine) releasing a report in 1997 promoting the adoption of EHRs in the healthcare systems across the US.⁸ As continued evidence emerged for improved quality and efficiency of healthcare delivery with the EHRs,⁹ the Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 incentivized EHR adoption and propelled the proliferation of EHR in healthcare systems.^10,11 Legislation continues to prioritize the ease of electronic health information exchange, as evidenced by the 21^st Century Cures Act of 2016, which addresses information blocking and mandates standards to further advance healthcare interoperability.¹² These challenges and their continued mitigation at both a federal and individual level are discussed in detail later in this review.

In addition to the clinical role of EHRs, the National Academy of Medicine identified research as their key secondary role.^8,13 EHRs present many opportunities to analyze, predict, and improve the care of patients with cardiovascular disorders and their outcomes. However, important scientific and technical considerations for optimizing EHR-based clinical research require careful and systematic data practices. This has prompted contemporary innovation in research strategy and data interoperability.^14–17

This review discusses the data infrastructure of EHRs that can help research, the challenges and opportunities for cardiovascular care innovation through data-driven discoveries in the EHR, and contemporary innovation in data interoperability and federated research strategies.

DATA ASSETS AND THE EHR

Data Streams within the EHR

The EHR aggregates a series of discrete data sources spanning structured and unstructured elements.^18,19 (Figure 1) Structured data exists within pre-defined fields of standardized data classes and elements which allow for easy extraction and analysis.¹⁹ Common structured data sources within the EHR include demographic characteristics (e.g. age, sex, race, and ethnicity), laboratory measurements (e.g. complete blood counts, chemical laboratory tests, biomarkers, etc.), vital signs (e.g. body mass index, heart rate, blood pressure, etc.), problem lists, diagnoses, procedures, medications, allergies, and healthcare utilization information (e.g. cost, hospitalization duration), among others.^1,20 These data sources have become relatively standardized through the Common Clinical Data Set (CCDS), which was developed by the Centers for Medicaid and Medicare Services (CMS) Meaningful Use Program in conjunction with the Office of the National Coordinator for Health Information Technology (ONC).^20,21

Figure 1.

Data Streams and Assets in the EHR

However, the majority of data contained within the EHR – approximately 80% – is unstructured.^22–24 Sources of unstructured data in EHRs include free text in the form of clinical narratives, procedure documentation, and diagnostic testing reports, such as those for imaging and pathology.^1,19,24,25 Picture Archiving and Communication Systems also store medical imaging and videographic data such as X-rays and echocardiograms.^19,23,26 Unstructured data captures the rich, multi-faceted, and complex nature of disease and provides insight into patient-provider interactions; these factors may be inadequately represented if only structured data is used. For example, a patient with a myocardial infarction requiring percutaneous coronary intervention may have corresponding diagnoses and procedural codes available in the chart that specify the territory of myocardial infarction and the culprit coronary artery that required intervention. However, signs and symptoms at presentation, intraprocedural challenges, post-procedural hospital course, and relevant complications are aspects of the disease that are often not captured within structured data. While unstructured data may be more useful in clinical concept recall than structured data,²⁷ the process of extracting and analyzing unstructured data can be time-consuming and expensive.^1,19 The emergence of specialized flexible applications of natural language processing (NLP), machine learning, and computer vision techniques enables health professionals and researchers to leverage these essential data sources for clinical and research.^4,19,28 Other forms of both structured and unstructured data streams such as continuous physiologic monitoring (e.g. vital signs, telemetry), data from mechanical ventilators and cardiopulmonary bypass machines, and raw videographic footage from endoscopic or laparoscopic procedures are not commonly stored due to very large file sizes and thus pose challenges in utilization and handling.^23,29,30 However, advances in data storage (e.g. data lakes) and computational methods, both discussed later in greater detail, allow researchers the potential to leverage these data more effectively.³¹

EHR-Adjacent Data Streams

Research using EHR is increasingly utilizing additional data streams such as genomics and molecular profiling, patient-generated health data (PGHD) obtained from wearable and nearable devices, patient-reported outcomes (PRO), and community and personal social determinants of health (SDOH).^20,23,24

Several initiatives have pioneered the integration of EHR and genomics data, including the National Human Genome Research Institute-funded eMERGE (Electronic Medical Records and Genomics) Network,^32–34 Vanderbilt University’s BioVU biobank,³⁵ and Kaiser Permanente’s Research Program on Genes, Environment and Health biobank,³⁶ with broad scientific guidance on merging genomics with EHRs.³⁷ This enables a new era of discovery arising from EHRs paired with genomics. An example is phenome-wide association studies (PheWAS) that map phenotypic information extracted from EHRs to single nucleotide polymorphisms (SNPs), allowing researchers to uncover novel associations between genetic variants and a variety of disease states. For example, PheWAS have revealed genetic variants implicated in the association between QRS duration in patients without cardiac disease and subsequent arrhythmia development, as well as the genetic component underlying the association of obesity with various disease phenotypes.^38,39

The growth of PGHD in recent years owes itself in large part to the increasing ubiquity of smartphones, wearable devices, remote monitoring devices, and mobile applications, which encompass a wide spectrum of functions including the monitoring of daily activity and fitness, heart rate and rhythm, and blood pressure.^20,40,41 Furthermore, EHR integration of PGHD is expected to accelerate following the 2020 ONC mandate for health information technology services to comply with interoperability standards and to implement technologies supporting application programming interfaces (API) that can better interface with EHRs.^25,42 Such integration will improve the monitoring and detection of patient risk factors through automated health messaging.⁴³ Additionally, these data can guide clinicians in decision-making for optimal patient treatment.⁴⁰ PGHD has revolutionized cardiovascular medicine via notable large-scale clinical trials that used wearables such as the Apple Heart Study, which enrolled over 400,000 ambulatory participants without atrial fibrillation and demonstrated an accurate irregular pulse notification algorithm utilizing the Apple Watch.⁴⁴ However, evidence supporting the use of wearables in improving health outcomes remains limited, and further research is warranted to ensure they are safely implemented and clinically meaningful.^45,46

Integrating PROs with EHR can similarly improve patient-centered health research. PROs provide significant indicators of patient health status post-intervention and are often primary or secondary outcomes within cardiovascular clinical trials as they encapsulate patient perspectives on the consequences of intervention on quality of life and symptoms.⁴⁷ The Patient-Reported Outcomes Measurement Information System (PROMIS) is a standardized series of PRO surveys that measure physical, mental, and social health developed by the NIH.⁴⁸ However, remote EHR-based collection of PROs presents several logistical and technical challenges, primarily in engaging patients to report their outcomes.^49,50 A recent study successfully implemented remote EHR-integrated data collection across seven orthopedic clinics to successfully administer standardized PROMIS measures across a heterogenous patient population.⁴⁹

Lastly, SDOH have long been recognized as important social, economic, psychosocial, and environmental factors mediating morbidity and mortality due to cardiovascular diseases.^51,52 Despite the significance of these determinants, the lack of systematic, standardized methods limits their capture within EHRs.^20,53,54 However, SDOH can better be captured into EHR through the development of systematic measurement of patients’ environment and interactions that influence their patient-provider relationship. For example, the IOM convened a committee of social scientists, clinicians, and informatics researchers who recommended 12 brief, standard SDOH screening measures in a panel that can be adopted through EHR.⁵³ This panel would reduce barriers to implementation, increase interoperability through widespread adoption, and reduce the need for redundant capture. Such a concise panel of standard measures increases clinical awareness of a patient’s health status and allows for integrating public health and community resources in patient care. A readily adopted standard set of measures would hopefully motivate EHR vendors to incorporate this panel into their products, health systems to adopt its use, and clinicians to incorporate this integral information. A variety of community-level SDOH data and indices are available to the public, including the American Community Survey published by the US Census Bureau, the SDOH database published by the Agency for Healthcare Research and Quality, and the Neighborhood Atlas Area Deprivation Index.^55,56 A study by Bhavsar and colleagues linked EHR and ACS data and found that lower neighborhood socioeconomic status was predictive of shorter time to use of emergency department visits and inpatient encounters, as well as shorter time to hospitalizations due to myocardial infarction and stroke.⁵⁴

Strategies and Infrastructures to Leverage the Data

Leveraging the heterogeneous and complex data within EHR and HER-adjacent data streams requires the development of novel infrastructures and strategies that can address data capture, maintenance, processing, and analysis challenges. The Coronavirus Disease 2019 (COVID-19) pandemic especially highlighted the need for platforms that could support rapidly updating datasets and agile analysis of real-time data to provide timely surveillance of incident cases to inform hospital operations.^57,58

Traditional data storage structures, such as enterprise data warehouses used in the analysis and reporting of structured data,²⁴ are not well-equipped to deal with heterogeneous data due to their structural inflexibility. Scalable data storage architectures, or data lakes, provide a centralized repository of data in its original form and enable flexible analytics through the concept of ‘late binding’, which enables researchers to perform customized analysis and bypass the rigid constraints of the schema in conventional data warehouses.^31,59 Data lakes are often linked to open-source computational frameworks such as Hadoop and Apache Spark, along with a suite of tools for data governance, data discovery, and extraction.^31,60–62 Other comprehensively integrated institutional data ecosystems like ‘data commons’ acquire and harmonize EHR data from multiple sources such that they can be readily used for research.^57,61

DISCOVERY SCIENCE IN EHR: OPPORTUNITIES, CHALLENGES, AND SOLUTIONS

Research Opportunities

Computational research in the EHR presents many opportunities for improving patient care and advancing medical knowledge. First, EHR data can improve the efficiency and efficacy of healthcare practice through real-world outcome assessments via retrospective record review to time-varying clinical risk prediction.^4,27,63–66 Second, techniques such as deep learning-based natural language processing (NLP) and computer vision can harness unstructured elements in the EHR for computational phenotyping of multifaceted and complex clinical disease scenarios, without resorting to abstractable structured elements alone.^{6,29,43,67,68} Third, data-driven investigations in the EHR can be utilized to achieve the 2007 FDA Amendment Act’s mission for post-marketing surveillance of medications, improving the safety and efficacy of these drugs.^69–72 Fourth, clinical trials can efficiently leverage the EHR to enrich screening using computable inclusion and exclusion criteria for potential trial subjects more likely to benefit from treatment.^69–74 Fifth, novel machine learning modalities can be used to derive distinct clusters of patients with clinical conditions (sub-phenotyping) by using the rich variety of data available within the EHR.^75,76 Such sub-phenotyping can be used to deliver precise therapy based on clinically relevant insights derived from other patients of the same phenotype.^77,78 Finally, patient-led research presents revolutionary opportunities if vendors’ EHR data is exchanged via write-back APIs, allowing patients to be directly involved in data-driven research.⁷⁹

Challenges and Solutions

While EHR represents a rich source of real-world data for evaluating clinical care, the data is often large, heterogenous, incomplete, and noisy.^2,80 This presents multiple challenges, and requires new methodologic approaches, significant data validation, and careful interpretation of results.⁵⁷

• Data Confidentiality

EHR data require rigorous handling practices due to its highly sensitive nature and are protected by institutional and governmental confidentiality guidelines, including the Health Insurance Portability and Accountability Act of 1996 (HIPAA) Privacy Rule.^81–84 Patient privacy can be protected by deidentifying and pseudonymizing data prior to analytical use, and ensuring user authentication and audit on devices and servers where the data is being used.^19,85–87

• Data Accuracy

EHR data accuracy can be compromised by errors and biases in manually recorded data due to physician overload and limited time for patient care, in addition to inter-institutional variations in encounter information practices.^88–90 A higher accuracy is associated with reimbursements linked to the diagnosis and procedure codes encoded in the billing information.⁵ Inconsistencies and differences in EHR accuracy data versus medical record abstraction can produce disparate results in performance analysis.⁹¹

When building reliable cohorts for data analysis, information should be triangulated based on multiple clinical domains such as diagnosis codes, imaging criteria, procedural intervention, medications administered, and clinical note documentation.^19,75 For example, to create a cohort of patients with hypertrophic cardiomyopathy (HCM), the use of diagnosis codes alone may not be accurate inclusion criteria. A more accurate cohort could be developed by including patients with diagnosis codes for HCM who are also undergoing treatment in a specialty HCM clinic, have hypertrophic features on their echocardiogram, and have documented symptoms and clinical manifestations of HCM in their notes. Regular data quality reporting standards can be institutionally implemented to ensure higher accuracy of EHR coding.⁹² Streamlined and physician-friendly EHR designs can also improve the accuracy of data capture.

• Data Completeness

As patients seek care at multiple sites, their health information is represented across multiple health systems. Clinical endpoints in studies cannot be reliably assessed unless all relevant patient encounters can be identified.⁹³ Moreover, the inability to capture clinical characteristics across these disparate EHR systems and data sources poses the risk of misclassifying patients. Randomly missing information can be ignored, but if the absence of captured information itself represents a clinical signal, it can be informative.^94,95

For longitudinal outcome assessments, only patients who consistently seek care at the same health system should be included to ensure completeness of follow-up.⁹⁶ Developing methodological standards to identify informative missingness in the EHR can improve predictive models for clinical disease outcomes.^94,95 For example, whether or not the physician ordered a laboratory measurement is often clinically relevant in determining if a patient’s prognosis is favorable, irrespective of the measured value,.⁹⁵ Regular and frequent data quality measurements across the linked data sources can lower encounters of random missing data.^97,98

• Data Heterogeneity

Institutional heterogeneity across multiple sites hinders the conduction of multi-site data-driven research and serves as a barrier for methodological reproducibility and consistency.^99,100 Emerging solutions to tackle data heterogeneity and promote data interoperability include harmonizing data to standard Detailed Clinical Models (DCMs) and mapping the databases to Common Data Models (CDMs). These are discussed further in the sections below.

• Data Capture

The challenges related to data capture in EHRs arise due to the large amounts of unstructured data representing clinical information, which are difficult to extract.^23,24

Methodological innovation within the realm of unstructured data, in the form of NLP on clinical notes and computer vision on medical images, can improve the amount of data captured during research.^28,101,102 NLP, a branch of computer science that combines computational linguistics with AI, can convert vast swaths of natural language data, such as unstructured free text contained within clinical notes, to structured, machine-interpretable data through sentence-splitting, tokenization, lemmatization, stemming, and terminology recognition via data dictionaries.¹⁶

Additionally, deep learning models that utilize structured and unstructured data performed well in predictions pertaining to key hospital metrics, including in-hospital mortality, 30-day hospital re-admission, and prolonged length of stay.^103,104 NLP can also be utilized to quantify stigmatizing language in provider notes, providing valuable insight into the impact of such language on patient care.¹⁰⁵ One such study found that stigmatizing language was more often used to describe non-Hispanic Black patients in hospital notes for treatment of patients with diabetes.¹⁰⁶

• Information Blocking

The intentional or unintentional obstruction of timely health information exchange poses a major barrier to the advancement of interoperability within the contemporary healthcare landscape, carrying downstream implications for healthcare delivery and clinical research. Informed by a substantial body of anecdotal evidence, the Office of the National Coordinator for Health Information Technology (ONC) released a report to Congress in 2015, proposing several targeted actions to mitigate this practice.¹⁰⁷ This eventually led to the 21^st Century Cures Act of 2016, as discussed earlier, which included provisions prohibiting information blocking, followed by the Cures Act Final Rule of 2020, which implemented these provisions and outlined eight exceptions to the rule.

Although information blocking remains prevalent,¹⁰⁸ efforts are underway to further mitigate this issue. These include implementing concrete enforcement guidelines by the Department of Health and Human Services (HHS), which may result in penalties spanning from disincentives for healthcare providers to civil monetary penalties for health IT developers and health information exchanges and networks. Currently, any party can anonymously submit information blocking complaints through the information blocking portal,¹⁰⁹ which may lead to a formalized investigation by the Office of Inspector General (OIG) or the ONC. Moreover, federal initiatives that incentivize and reward interoperability, such as the Medicare Promoting Interoperability Program, may also ameliorate information blocking.¹¹⁰

• Physician Burnout

Although EHRs were intended to help clinicians improve productivity, efficiency, and work-life balance,¹¹¹ they have also significantly contributed to clinician burnout.^112–116 Documentation tasks, inbox workload, after-hours charting, and poor EHR design and workflow, are commonly cited factors contributing to this issue, among many others.¹¹⁵ EHR-associated clinician burnout may thus lead to healthcare providers’ resistance to accepting ever-evolving EHR systems and limit the utilization of innovative EHR-based research modalities.

EHR-associated burnout is increasingly recognized as a major issue and has received widespread attention from professional medical societies (e.g. American College of Physicians, American Medical Informatics Association),¹¹⁷ standards-development organization Health Level Seven (HL7),¹¹⁸ as well as federal bodies including the ONC.¹¹⁹ Although the process of redesigning EHRs to be more clinician-friendly can be a costly and logistically challenging endeavor, continuing EHR improvements are feasible and have been successfully implemented at a multi-institutional level.^117,120 Additionally, numerous major EHR vendors have spearheaded efforts focused on improving the clinician-EHR experience, ranging from Allscripts’ Human-Centered Design initiative,¹²¹ to Epic and Cerner’s continued development of artificial intelligence (AI) solutions such as smart voice assistants.¹²² The continued refinement of AI technologies, advancements in data storage and data models supporting interoperability, and evolution of EHR-based research methodology are poised to further enhance the clinician-EHR experience.

• Implementing AI Models in the EHR

Some of the limitations of AI augmented EHR-based research lie within the fact that input data are generated within a non-stationary environment with shifting patient populations, and clinical care practices shift over time. Thus, introducing and implementing novel prediction algorithms could potentially cause a change in practice, which influences the training data for algorithms. Therefore, developing methods to identify drift and update models in response to performance deterioration is critical. Careful performance quantification and evaluation with periodical retraining could potentially mitigate these challenges. Data-driven testing procedures have the potential to appropriately update models to maintain a similar level of performance over time.¹²³

Moreover, the economic impacts of AI-based research and developing technologies that must be considered. The U.S. healthcare payment system continues to evolve to account for the added value of these services, as well as the constant support they require from technical experts. For example, as radiology AI algorithms progress, one Category III code for AI use in CT and two for non-elastographic quantitative analysis with ultrasound became active in 2022.¹²⁴ The transition to value-based care may additionally lower some of the barriers to the adoption of AI tools in clinical care related to lack of reimbursement through fee-for-service paradigms. Additionally, AI models used in healthcare have enabled significant cost savings when utilized in diagnosis and treatment.¹²⁵ Furthermore, the economics of AI can be improved through the incorporation of pruning, reduction in bias, explainability, and regulatory approval.

SCALING INNOVATION FROM THE EHR: DATA MODELS AND FEDERATED STRATEGIES

Data Interoperability – Standardized Vocabularies, Detailed Clinical Models (DCMs), and Common Data Models (CDMs)

Although operational interoperability is rare in EHR data infrastructures, recent policies and tools aim to enhance EHR access and exchange.^126–129 Lack of interoperability is a major challenge due to multiple types of EHR used across and within health systems producing disparate data infrastructure and coding terminologies that lead to inter-research inconsistency. The United States Core Data for Interoperability (USCDI) is a standardized set of health data classes and elements intended to enable nationwide, interoperable EHR exchange.¹²⁸ The use of standardized terminologies and ontologies, such as International Classification of Diseases (ICD),¹³⁰ Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT),¹³¹ Current Procedural Terminology (CPT),¹³² RxNorm,¹³³ and Logical Observation Identifiers Names and Codes (LOINC),¹³⁴ for encoding of structured data elements in the EHR aids in accurate outcome comparisons across diverse health systems.⁸⁶ Moreover, due to a lack of consistency in EHR vendor data storage, outcomes in patients seeking care at multiple sites cannot be tracked conveniently.²⁰ Consistent data infrastructure for storing EHR data across different health systems or mapping the data to common data models can enable patient-level linking of databases, allowing better capture of information for each patient.^135,136

Detailed Clinical Models (DCMs) represent an effort to harmonize clinical knowledge modeling and preserving computable meaning in EHR data during an exchange between heterogenous data infrastructures.^137–139 DCMs are a small, standalone information models designed to express clinical concepts in a standardized and reusable manner, such that they can be independently maintained during the transformation of data across formats.^138,140 The Clinical Informatics Modeling Initiative (CIMI) and Health Level Seven’s (HL7) International Fast Healthcare Interoperability Resources (FHIR) standard are examples of DCM initiatives that improve semantic interoperability in healthcare information systems.^141,142

While the models like CIMI and FHIR serve as a layer of standardization for clinical data elements, integration across broad data architectures and formats requires mapping of EHR data to common data models (CDMs).^99,143 CDMs attempt to achieve true interoperability – by defining the same tables, same column names, and each element defined a consistent way to ensure the process of creating programs for one data stream in the EHR is immediately usable across all other EHR data streams in the same CDM format. Therefore, CDMs represent standard data schemas that can integrate data from multiple clinical areas and incorporate standard vocabularies. While various CDMs have been proposed, the Observational Medical Outcomes Partnership (OMOP) CDM and the National Patient-Centered Clinical Research Network (PCORnet) CDM are commonly used.^143–145

The process of data harmonization and transforming raw source EHR data to a CDM is referred to as ‘Extract, Transform, Load’ (ETL).^143,146,147 ETL involves extracting the EHR data from the source, transforming it by de-duplicating, linking, cleaning, conducting quality checks, and loading it into the target CDM database. While transforming data from the EHR system to the CDM, potentially large amounts of information can be misrepresented or lost in translation. Thus, systematic care must be taken to ensure data quality in terms of conformance, completeness, plausibility, and precision:^148,149

• Conformance

  Conformance refers to the representation of source EHR data against the formatting, relational, and computational definitions of the CDM. Syntactical conformance relates to the formatting and structure of datasets. This often involves matching data types or transforming the format of datasets to meet CDM standards, such as reshaping data from wide-to-long or vice versa. For example, in the case of mapping to the OMOP CDM, syntactical conformance means that the EHR data should be transformed to columns in the OMOP CDM tables. Semantic conformance measures how accurately the clinical values map to the corresponding concepts in the standard CDM vocabulary.^86,150,151 This can be done via automated ontology mapping and rigorous manual review. For example, in the case of OMOP, semantic conformance refers to how accurately the diagnosis codes in the EHR (often present as ICD-10 codes) map to the standard concepts derived from the SNOMED CT vocabulary.⁸⁶

• Completeness

  Completeness refers to the frequencies of source EHR data attributes present in the CDM dataset. An adequate ETL process ensures minimal loss of EHR data in translation from the source data structure to the CDM.^97,152

• Plausibility

  Plausibility refers to the believability and accuracy of data values. Unlike conformance and completeness, which focus only on the structure and presence of values, plausibility focuses on the actual values as representations of real-world source data. To evaluate the plausibility of the mapped data, different data fields can be checked for erroneous values that are physiologically implausible by performing quality assessments on the mapped data.^148,153 Assessing the distributions for numerical variables and the frequencies for categorical variables can help screen for such implausible mapping. This necessitates edits to the ETL process design, often aided by graphical and dynamic tools,^144,154 with inputs from clinical experts guiding the mapping process.

• Precision

  Due to differences in granularity between EHR data coding systems and the standard vocabularies, data transformations can combine lists of multiple elements from the source data into a manageable aggregated category. While this may be convenient and essential for mapping, the aggregation should be guided by clinical and methodologic expertise to preserve the necessary precision for assessing study variables.¹⁵³ Moreover, many CDMs, including the OMOP CDM, ensure that the EHR source values are preserved in the transformed tables in different columns. Thus, research investigators who need to modify the mapping using different precision and granularity criteria can perform these processes without re-extracting the EHR data.¹⁴⁶

Close collaboration between computational and clinical experts may be necessary to define the logic of the ETL design and the debugging process after thorough data quality assessments of the mapped datasets. Multiple graphical and dynamic tools can aid in the design and modification of the ETL approach.^144,154

Federated Strategies for Multi-institutional Research

Interoperable CDM databases enable research studies to be conducted in a large-scale, multi-center network setting. EHR data from multiple health systems are mapped to a common data model, enabling analytical strategies to be deployed without the need for sharing patient-level data across institutes. This enables multi-center studies while ensuring data confidentiality and patient privacy.⁸⁷ Federated strategies make multi-center studies feasible, ensure consistency in the implementation of methods, and improve the reproducibility of results.¹⁵⁵ (Figure 2)

Figure 2.

Scaling Innovation in EHR Research with Federated Network Studies

These strategies are being adopted widely for large-scale pharmacoepidemiologic and comparative assessments of medical therapies. For utilization trends, effectiveness and safety comparisons of medication use in hypertension and diabetes, ‘Large-scale Evidence Generation and Evaluation across a Network of Databases (LEGEND) – Hypertension’ and LEGEND – Type 2 Diabetes Mellitus (T2DM), are being conducted respectively.^155–159 Figure 3 represents the specific study approach of LEGEND-T2DM and its participating health system databases.¹⁵⁸ Similarly, a large data repository comprising EHR data from patients across 8 diverse healthcare systems, The Guideline Advantage™ (TGA) repository, has been used to evaluate whether cardiovascular health trends across large populations are progressing towards the American Heart Association’s (AHA) Impact Goals.¹⁶⁰ Such federated strategies were also adopted during the COVID-19 pandemic to create a large repository of deidentified patient information, National COVID Cohort Collaborative (N3C), which could be used by clinical researchers in the open-source community.^151,161

Figure 3.

The ‘Large-scale Evidence Generation and Evaluation across a Network of Databases for Type 2 Diabetes Mellitus’ (LEGEND-T2DM) Study

All data sources participating in such federated studies need IRB approval to contribute to a multi-site study. While federated studies enable the collaborative deployment of analytical methods without exchanging patient-level information, there are specific regulatory considerations for these studies. The IRBs need to ensure investigator compliance with the proposed research protocol. Automating this IRB monitoring of the data resources that are extracted and accessed can improve compliance with the regulations for human subject protection and increase operational efficiency while minimizing the logistical burden on the IRB.¹⁶² However, although the sharing of code for analysis is more secure than sharing of patient-level data, there is a limited risk for data leakage in federated research.¹⁶³ Malicious actors can perform property or membership inference attacks which can be used to determine limited information about the patients.^163,164 Such attacks can be mitigated by regulatory interventions of maintaining stringent data access across participating institutes and evaluating the code before deployment. Methodological interventions, including differential privacy, the introduction of stochastic noise, and intentional rounding of values in the data, can also be used across sites.^165,166

Beyond these conventional federated strategies, novel approaches such as swarm learning are being developed for decentralized machine learning on peer-to-peer networking technology and optimization of confidential study designs.^87,167 While the disease classifier models developed using this approach have been shown to outperform models made at individual sites, the adoption of this approach in clinical research is lacking.¹⁶⁸

CONCLUSIONS

The EHR provides a valuable source of clinical structured and unstructured data for research that can predict cardiovascular clinical outcomes, enhance cardiology practice and diagnosis, address healthcare disparities, and advance precision medicine. Several challenges currently exist in utilizing EHR for data-driven research — however, recent innovations in data storage, synthesis, and analytics can ameliorate current challenges. Moreover, emerging interoperable data infrastructure, enhanced regulation of data exchange, and patient-integrated research can enable large-scale, multi-center evidence generation with improved methodological consistency. With continued development and innovation, HER-based research can bring novel technology to patient care and improve health outcomes.