By Sarah Handzel, BSN, RN
Worldwide, in both acute care and ambulatory settings, more than 300 million ECGs are performed annually.1 The enormous amount of ECG data yielded provides a wealth of information regarding cardiac diagnoses, but clinicians may balk at the work required to sift through such massive data sets. This is especially true if data comes from multiple healthcare organizations.
Fortunately, a combination of comprehensive cardiology data management systems and artificial intelligence (AI) can help streamline this process and provide pertinent insights, allowing cardiologists to better identify risk factors for significant cardiovascular events.
Invaluable Insight into Patient Health Data
New technologies provide data management platforms that streamline communication, keep patient data secure, and integrate with other systems (such as electronic medical records). As a result, clinicians have access to pertinent patient data in one centralized location that is capable of being amended and shared among the entire healthcare team.2 Many of these systems also seamlessly connect to diagnostic testing devices such as ECG. This allows healthcare providers to retrieve near real-time data, proving invaluable for providing a correct diagnosis.
In addition, a growing number of healthcare facilities are adopting data management technologies and merging them with electronic health records, financial data, and other forms of clinical data.3 With the assistance of these systems, cardiologists can view patient information and clinical results from any location, providing greater workflow flexibility and leading to better patient care.
In the case of ECGs specifically, collecting and examining data retrospectively helps reduce future diagnostic inaccuracies to better guide clinical decision-making as physicians can more easily identify diagnostic trends within specific data sets.
Data Mining and ECG Retrospective Analysis
Several studies have illustrated how large-scale ECG data analysis can uncover useful insights—such as the identification of inconsistent methods of measurement, overlooked abnormalities that may suggest an underlying disease, or the impact of making a change in how ECGs are processed across an institution.
Retrospective Analysis Often Reveals Errors in Initial Results
A recently published study in Cureus compared Bazett's QT interval formula (QTcB) with Fridericia's formula (QTcFri) to determine the extent of difference in those identified as "prolonged QT." Previous research indicates Bazett's formula is actually more sensitive to heart rate, and most modern healthcare data management systems automatically use Fridericia's formula in its place. After examining more than 44,000 ECGs retrospectively, researchers found that 21% of all the ECGs corrected to normal QTc when Fridericia's formula was applied.4 This equates to 57% fewer ECGs identified as having a QTc greater than 500ms, which is the threshold for identifying patients at risk for drug-induced torsade de pointes (TdP).4, 5
Obviously, a change of 57% impacting thousands of ECGs in this hospital-based study has important clinical consequences impacting patient care, such as the administration or withholding of medication. Consequently, when comparing ECGs over time, it is vital that a consistent method be used, including the correction formula for correcting QT for heart rate.
Even without accounting for initial errors in ECG interpretation, a retrospective analysis may yield previously overlooked information important for predicting risk for mortality. A study in the International Journal of Cardiology reviewed ECG data from over 342,000 primary care patients over a period of 10 years. Upon review of the ECG data, researchers found an association between abnormal T waves (those with asymmetry, flatness, or notching) and mortality risk independent of other patient-specific factors, such as heart rate, QTc, and other baseline comorbidities.6
Additional Benefits Help Workflow and Healthcare Costs
Retrospective analysis of ECGs may also be useful for determining cost-effectiveness and optimizing workflow. A 2015 study in The Journal of Pediatrics examined clinical data from over 8,600 ECGs performed as part of athletic pre-participation evaluations (PPE) to assess children and young adults' readiness to participate in sports. From 2005 to 2010, pediatricians selectively performed ECG during the PPE based on suspicious cardiac symptoms, such as dizziness or syncope. Only 0.5% of individuals receiving PPE alone were referred to cardiologists based on clinical suspicion, but 13% of individuals receiving PPE and an ECG were referred to cardiologists for further evaluation.7
Evaluation of the same study's data years later further supports the use of ECG for diagnostic purposes, as PPE alone identified cardiac disease with a sensitivity of only 44%. Within one year of the PPE, cardiac disease was diagnosed in 0.5% of people receiving the test. However, incorporating the ECG into a PPE helped identify underlying cardiac disease in 18% of participants, allowing more patients to receive therapeutic procedures.7
AI's Role in Evaluating Data
The expanding use of AI can also benefit clinicians seeking to analyze previously recorded ECGs. The machine-learning techniques involved go beyond evaluating only a small portion of the information contained in an ECG readout; rather, researchers are aiming to develop algorithmic frameworks that support large-scale ECG analysis. Ultimately, this may help physicians save time and effort when interpreting ECG findings and making treatment decisions.
Algorithms Increasingly Aid in Trend Identification and Diagnosis
Researchers created this type of algorithm and published their findings in Circulation: Cardiovascular Quality and Outcomes with the ultimate goal of developing an automated, scalable, and interpretable method of characterizing cardiac structure and diastolic function as well as detecting and tracking disease via patient-specific ECG profiles.2
In the examination of standard 12-lead ECGs collected from 2010 to 2017, the model was restricted to ECGs presenting normal sinus rhythm. Using continuous metrics, the AI approach enabled estimation of the severity of structural abnormalities, which were verified against reference electrocardiographic measurements from GE's MUSE Cardiology Information System. The algorithm also helped to classify four example diseases (pulmonary arterial hypertension, hypertrophic cardiomyopathy, cardiac amyloidosis, and mitral valve prolapse), even identifying new ECG predictors for each disease.2
Another machine learning platform, known as "rECHOmmend", was developed to analyze data from any ECG system to predict various types of structural heart disease, including clinically significant valvular disease, reduced left ventricular ejection fraction, and pathologically increased septal thickness. Using over 1 million ECG traces combined with information about each patient's age and biological sex, the platform demonstrated an overall positive predictive value of 42% for clinically meaningful structural disease with 90% sensitivity and 73% specificity. These findings, published in Circulation, could eventually be used to identify high risk patients using only ECG data.8
Stay on top of cardiology trends and best practices by browsing our Diagnostic ECG Clinical Insights Center.
Deep Neural Networks Augment Physician Expertise
Other research supports the use of machine learning techniques for interpreting ECG findings. Deep neural networks (DNNs)—which have already been used to classify images and in speech recognition tests—combine two or more artificial neural networks to predict specific outcomes more accurately.9
Research in Nature Communications examined the use of DNNs in classifying six types of abnormalities (first-degree AV block, right bundle branch block, left bundle branch block, sinus bradycardia, atrial fibrillation, and sinus tachycardia) by employing data from over 2 million standard 12-lead ECGs.10 The data was split into a training set to develop the DNN and a validation set to confirm the results.
When compared with interpretations from two fourth-year cardiology residents, two third-year emergency residents, and two fifth-year medical students, the DNN matched or outperformed all people in identifying the six cardiac abnormalities.10 Other work published in Nature Medicine supports these conclusions, showing that another DNN exceeded average cardiologist sensitivity for all rhythm classes in a sample of over 91,000 single-lead ECGs.11
However, it is important to remember that neither AI-based algorithms or DNNs will replace physician expertise and clinical judgment. The future availability of large-scale ECG tracing databases, combined with DNNs performing automatic ECG analysis, could instead help save clinicians time and prevent incorrect diagnoses.
ECG Storage in the COVID-19 Era
AI applications must have raw data that can be plugged into algorithms; simply storing ECG images is not sufficent for accurate interpretation of the data. Management of ECG data is especially important as healthcare providers grapple with the COVID-19 pandemic, because the health consequences of the infection are not yet fully understood. Current research, such as that in the Archives of Academic Emergency Medicine, show various abnormal findings common among patients with COVID-19, including sinus tachycardia, abnormal T waves, ST-segment depression, prolonged QT interval, bifascicular block, and left anterior hemiblock.12 These findings were observed both at the time of admission and during the hospital stay, leading to better prognostic predictions for cardiologists.
Clinicians should continue to analyze ECGs using best practices and consider the potential of comprehensive data management systems and machine learning techniques to supplement current understanding and knowledge. Insights gained from ECG analysis may prove indispensable in future diagnostic and prognostic cases.
- Zhu H, Cheng C, Yin H, et al. Automatic multilabel electrocardiogram diagnosis of heart rhythm or conduction abnormalities with deep learning: a cohort study. The Lancet Digital Health. 2020;2(7):e348-e357. https://www.thelancet.com/journals/landig/article/PIIS2589-7500(20)30107-2/fulltext
- Tison G, Zhang J, Delling F, et al. Automated and interpretable patient ECG profiles for disease detection, tracking, and discovery. Circulation: Cardiovascular Quality and Outcomes. 2019;12(9):e005289. https://doi.org/10.1161/CIRCOUTCOMES.118.005289
- University of Pittsburgh School of Health and Rehabilitation Sciences. What is health care data management? University of Pittsburgh. https://online.shrs.pitt.edu/blog/what-is-health-care-data-management/. Accessed October 7, 2022.
- Rosenblum A, Dremonas A, Stockholm S, et al. A retrospective analysis of hospital electrocardiogram auto-populated QT interval calculation. Cureus. 2020;12(7):e9317. https://doi.org/10.7759/cureus.9317.
- Drew B, Ackerman M, Funk M, et al. Prevention of torsade de pointes in hospital settings. Circulation. 2010;121(8):1047-1060. https://doi.org/10.1161/CIRCULATIONAHA.109.192704
- Isaksen JL, Ghouse J, Graff C, et al. Electrocardiographic T-wave morphology and risk of mortality. International Journal of Cardiology. 2021;328:199-205. https://www.internationaljournalofcardiology.com/article/S0167-5273(20)34264-9/fulltext
- Burns K, Encinosa W, Pearson G, et al. Electrocardiogram in preparticipation athletic evaluations among insured youths. The Journal of Pediatrics. 2015;167(4):804-809. https://doi.org/10.1016/j.jpeds.2015.06.011
- Ulloa-Cerna AE, Jing L, Pfeifer JM, et al. RECHOmmend: an ecg-based machine learning approach for identifying patients at increased risk of undiagnosed structural heart disease detectable by echocardiography. Circulation. 2022;146(1):36-47. https://www.ahajournals.org/doi/10.1161/CIRCULATIONAHA.121.057869
- Shamshirband S, Fathi M, Dehzangi A, et al. A review on deep learning approaches in healthcare systems: Taxonomies, challenges, and open issues. Journal of Biomedical Informatics. 2021;113:103627. https://doi.org/10.1016/j.jbi.2020.103627.
- Ribeiro A, Ribeiro M, Piaxão G, et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nature Communications. 2020;11:1760. https://doi.org/10.1038/s41467-020-15432-4.
- Hannun A, Rajpurkar P, Haghpanahi M, et al. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature Medicine. 2019;25(1):65-69. https://doi.org/10.1038%2Fs41591-018-0268-3.
- Aghajani M, Toloui A, Aghamohammadi M, et al. Electrocardiographic findings and in-hospital mortality of COVID-19 patients; a retrospective cohort study. Archives of Academic Emergency Medicine. 2021;9(1);e45. https://doi.org/10.22037/aaem.v9i1.1250.
Sarah Handzel, BSN, RN, has been writing professionally since 2016 after spending over nine years in clinical practice in various specialties.
The opinions, beliefs and viewpoints expressed in this article are solely those of the author and do not necessarily reflect the opinions, beliefs and viewpoints of GE Healthcare. The author is a paid consultant for GE Healthcare and was compensated for creation of this article.