http://orcid.org/0000-0003-2618-4285
?Mathematical formulae have been encoded as MathML and are displayed in this HTML version using MathJax in order to improve their display. Uncheck the box to turn MathJax off. This feature requires Javascript. Click on a formula to zoom.Abstract
Objectives. In heart failure, invasive angiography is often employed to differentiate ischaemic from non-ischaemic cardiomyopathy. We aim to examine the predictive value of echocardiographic strain features alone and in combination with other features to differentiate ischaemic from non-ischaemic cardiomyopathy, using artificial neural network (ANN) and logistic regression modelling. Design. We retrospectively identified 204 consecutive patients with an ejection fraction <50% and a diagnostic angiogram. Patients were categorized as either ischaemic (n = 146) or non-ischaemic cardiomyopathy (n = 58). For each patient, left ventricular strain parameters were obtained. Additionally, regional wall motion abnormality, 13 electrocardiographic (ECG) features and six demographic features were retrieved for analysis. The entire cohort was randomly divided into a derivation and a validation cohort. Using the parameters retrieved, logistic regression and ANN models were developed in the derivation cohort to differentiate ischaemic from non-ischaemic cardiomyopathy, the models were then tested in the validation cohort. Results. A final strain-based ANN model, full feature ANN model and full feature logistic regression model were developed and validated, F1 scores were 0.82, 0.79 and 0.63, respectively. Conclusions. Both ANN models were more accurate at predicting cardiomyopathy type than the logistic regression model. The strain-based ANN model should be validated in other cohorts. This model or similar models could be used to aid the diagnosis of underlying heart failure aetiology in the form of the online calculator (https://cimti.usj.edu.lb/strain/index.html) or built into echocardiogram software.
Introduction
Heart failure has been estimated to affect 1–2% of the adult population in developed countries, increasing to greater than 10% in people aged 70 and over [Citation1]. There are numerous underlying aetiologies of heart failure including hypertension, viral infections, chemotherapy and inherited cardiomyopathies. However, ischaemic coronary artery disease (CAD) is the most common underlying cause [Citation2]. Invasive angiography is often required to differentiate underlying myocardial disease due to CAD from other causes [Citation3]. However, non-invasive imaging is being increasingly utilized to make this distinction [Citation4].
Clinical features obtained from history and examination, echocardiography and electrocardiography are key for the elucidation of the diagnosis underlying heart failure [Citation5]. A number of specific electrocardiographic features, criteria and scores have been identified that can help characterise certain types of myocardial disease [Citation6–10]. For example, the Selvester QRS score takes into account various Q wave parameters and is a validated tool for both the identification [Citation11,Citation12] and quantification of ischaemic myocardial damage [Citation7]. Further, a number of characteristic electrocardiogram (ECG) features associated with the most common cardiomyopathy phenotypes (hypertrophic and dilated) have been identified and conglomerated into predictive criteria and scores [Citation8–10,Citation13]. Additionally, echocardiographic parameters can be used to examine patterns of cardiac dysfunction and refine the underlying cause of heart failure [Citation3]. Measuring strain parameters via speckle tracking echocardiography is not widely used as a routine part of clinical heart failure investigation, despite strain parameters emerging as a sensitive method for detecting subtle cardiac dysfunction in several clinical contexts [Citation14–16]. To date, little has been done to investigate the effectiveness of strain patterns in differentiating types of myocardial disease underlying heart failure, despite the improved sensitivity and reproducibility of these measures. However, strain parameters have been demonstrated to independently predict prognosis in ischaemic cardiomyopathy, using logistic regression modelling [Citation17].
The use of logistic regression models to predict complex biological relationships may be limited in some contexts, partly because many biological relationships are non-linear [Citation18,Citation19]. Artificial Neural Networks (ANN) allow for modelling of both linear and non-linear relationships and have been shown to be a more accurate method than regression modelling for predicting risk/outcome in several clinical settings [Citation20–22].
Here, in a cohort of patients with an ejection fraction <50% we developed and validated both a logistic regression model and ANN models for differentiating individuals with ischaemic cardiomyopathy (CM) from those with non-ischaemic CM. These models were constructed using: (a) speckle tracking strain patterns and (b) speckle tracking strain patterns in combination with demographic and ECG features. In summary, we aim to develop, validate and compare ANN and logistic regression models for differentiating ischaemic from non-ischaemic CM.
Materials and methods
Patient selection
From the American University of Beirut Medical Center medical records, we retrospectively identified 231 consecutive patients with systolic dysfunction (defined as an ejection fraction (EF) <50% on echocardiography) who received a subsequent coronary angiogram and ECG within 90 days and did not have atrial fibrillation or a pacemaker, between 1 July 2013 and 1 December 2015. Of these, 204 had echocardiogram images of suitable quality to allow post hoc left ventricular strain measurements to be made, and were subsequently included in the analysis. Demographic variables and co-morbidities were retrieved from the medical records.
Electrocardiogram
All ECG were reviewed blindly by a board certified cardiologist. Thirteen ECG variables previously linked to specific types of myocardial disease were obtained and recorded including: left bundle branch block, left anterior fascicular block, right bundle branch block, right bundle branch block with left anterior fascicular block, Cornell criteria, Sokolow-lyon criteria, right atrial overload, Selvester score, hypertrophic cardiomyopathy score and prolonged p wave duration [Citation6–8,Citation11–13].
Strain and other echocardiographic parameters
For each patient and for each of the 18 left ventricular segments, the following parameters were derived: (1) longitudinal peak systolic strain (peak strain), presence of early systolic lengthening (ESL; continued lengthening of a segment following the initiation of systole); (2) presence of post systolic shortening (PSS; as defined by the PSS peak being >90 ms after aortic valve closure) and (3) duration of PSS (PSSdtime; as defined by milliseconds between aortic valve closure and peak systolic shortening). Further, the presence or absence of a regional wall motion abnormalities and left ventricular internal dimension at end diastole (mm) were recorded. These parameters were derived using DICOM images and analysed using EchoInsight® software. This is a non-vendor specific software and therefore since all images are interrogated with one software this reduces error from using different vendor software. Images were obtained using Phillips and GE cardiovascular ultrasound machines.
Defining ischaemic and non-ischaemic cardiomyopathy
Diagnostic angiograms in addition to the presence or absence of previous myocardial infarction or revascularisation were used to categorise patients as having either ischaemic CM (n = 146) or non-ischaemic CM (n = 58), as per Felker et al.’s definition [Citation23]. Patients with significant coronary artery disease (>50% stenosis) in two or more epicardial coronary arteries, or significant single-vessel disease in the left main or proximal left anterior descending artery, or significant single vessel disease with a history of revascularization or myocardial infarction were categorised as having ischaemic CM [Citation23].
Statistical methods
Three statistical learning models were developed and validated to predict the known outcomes of ischaemic or non-ischaemic CM (detailed in the data analysis section below). The three models developed were: (1) a logistic regression model utilizing all echocardiographic, electrocardiographic and demographic variables, (2) an ANN model utilizing all echocardiographic, electrocardiographic and demographic variables (full feature models) and (3) an ANN model utilizing peak strain measurements and presence of regional wall motion abnormality only (strain-based model).
In the data description, the average and the standard deviations of each group of patients (ischaemic versus non-ischaemic) were computed. In comparing the obtained models, the positive and negative predictive values were used as well as the positive and negative likelihood ratios. Finally note that all accuracies, sensitivities and specificities were reported with their respective 95% confidence intervals.
Data analysis
Artificial neural network models
A detailed introduction to ANN has been described previously [Citation19]. The architecture design for the ANN employed a systematic method where the number of neurons was changed incrementally, and bootstrapping was performed to evaluate the accuracy of the models. We derived one model to predict if a patient has ischaemic CM present by peak strain and wall motion abnormality data only (ANN strain-based model) and another model by using all strain data (including PSS, PSSdtime and ESL), regional wall motion abnormality data, ECG data and demographics (ANN full feature model). The covariates used in the ANN full feature model are present in and the covariates used in the ANN strain model are highlighted in bold in . A multilayer network was used for the predicted outcomes consisting of one input layer one hidden layer and one output layer. See Supplement 1 for an explanation of how the artificial neural network models were developed.
Table 1. Demographic and electrocardiographic variables used in predictive models.
Table 2. Echocardiographic variables used in predictive models.
Logistic regression
This method is the starting point for devising any best-fitting model, and is essential because it gives an indication about the relevance of each of the predictors used for risk prediction. See Supplement 1 for an explanation of how the logistic regression model was developed. The MATLAB implementation for logistic regression outputs the p value for every predictor coefficient. Predictors that had a p value <.05 were selected for the logistic regression model.
Prediction model for ischaemic cardiomyopathy present
The derivation cohort was randomly chosen as follows: 29 out of the 58 patients who tested negative (non-ischaemic CM) were added randomly to the derivation cohort and 73 out of the remaining 146 patients who tested positive were also added randomly to the derivation cohort. The remaining 102 patients were all added to the validation cohort and were used only to test the performance of the model derived using the derivation cohort.
ANN bootstrapping
ANN is a special type of non-linear regression that presents multiple local minima; hence, every time we run the training algorithm it will converge in general to a different model. In order to choose the best model, the training process is repeated 200 times. In every iteration the algorithm converges to a point that give the highest accuracy on the validation set. Then, every model of the 200 generated models is tested on the testing cohort and the sensitivity and specificity of the model are recorded. Some models presented good specificity but poor sensitivity, some models the opposite, and some presented good specificity and sensitivity. The results of the 200 models are presented as sensitivity versus 1-specificity of each model. However, if two models had the same specificity but different sensitivities, then only the model with higher sensitivity was included.
Statistical analysis
The sensitivity and specificity was used to assess the strength of each model. The F1-score defined as the geometric mean of sensitivity and specificity is computed by the following formula:
The F1 score will have a highest value of 1 when both sensitivity and specificity are 1, which means a perfect model, and a value of 0 if sensitivity or specificity is 0.
Finally, the percentage of avoided tests corresponding to each model to predict ischaemic CM was then plotted against the sensitivity of the prediction model. The number of tests avoided was calculated assuming that the results of the corresponding predictive model were respected. The %avoided was then calculated as
Percentage ordered is 100 − %avoided.
Ethical approval was obtained from the institutional review board at the American University of Beirut Medical Center. The study abided by the Declaration of Helsinki.
Results
The cohort included 204 patients, 78% male, mean age 69 years. Patients in the ischaemic CM groups (n = 153) were older, had more comorbidities and were more likely to be male (). Eight of the 13 ECG parameters examined showed significant differences between groups (). Global longitudinal strain (GLS) and EF did not differ between groups ().
Examining individual left ventricular segments:
Peak strain did not differ significantly between groups, however there was a trend towards reduced peak strain in the apical segments in the ischaemic CM group.
Presence of ESL often differed significantly between the groups (p value ≤.0001), however whether it was more prevalent in the ischaemic or non-ischaemic groups was heterogeneous between segments ().
PSS tended to be more prevalent and longer in duration in the ischaemic CM group.
The variables found to be independent predictors of ischaemic CM for the cohort by logistic regression were older age (p value = .039), having hypertension (p value = .05), a smoking history (p value = .004) and the presence of a regional wall motion abnormality (p value = .001).
Predictive models
For the development of the predictive models, the derivation cohort consisted of 73 patients with confirmed ischaemic CM and 29 patients with confirmed non-ischaemic CM, selected randomly. The validation cohort comprised the remaining 102 patients.
shows sensitivity, specificity, negative and positive predictive values, likelihood ratios and F1 scores for the full feature ANN and logistic regression models, and strain-based ANN model. Both ANN models were more accurate than the full feature logistic regression model. There was little difference in the predictive accuracy between the full feature and strain-based models. The stain based ANN model demonstrated good sensitivity, specificity, positive likelihood ratio and positive predictive value for identifying patients with ischaemic CM, however negative predictive value was weaker (). shows the plot of the sensitivity of the strain-based ANN model versus 1-specificity to predict ischaemic CM. The cut-off threshold was determined as the one that maximizes the F1 score as it is the measure of reference in our research. shows the ROC curve for the ANN strain-based model. The probabilistic output of the ANN has been thresholded at different levels to yield different sensitivities and specificities. For each sensitivity–specificity pair, the F1-score was calculated as defined by its equation. The operation cutoff was chosen as the one that maximizes the F1-score as shown in .
Figure 1. Sensitivity versus 1-specificity for strain-based artificial neural network (ANN) model to predict ischaemic cardiomyopathy in systolic heart failure.
Figure 2. ROC curve for the strain-based ANN model.
Table 3. Accuracy of predictive models.
Diagnostic angiograms avoided
shows the sensitivity of a given model to predict ischaemic CM increases, diagnostic angiograms avoided decreases exponentially and diagnostic angiograms ordered increases. This is based on the assumption that the results of our predictive models are accepted, with the outcome of ischaemic CM resulting in an invasive angiogram being ordered and the outcome of non-ischemic CM resulting in an invasive angiogram being avoided. Applied to our strain-based model this would mean ∼40% of diagnostic angiograms would be avoided.
Figure 3. Diagnostic angiograms avoided in relationship with model sensitivity. The number of diagnostic angiograms avoided decreased exponentially as the sensitivity of the test increased.
Discussion
In this study, we developed a logistic regression model and ANN models to predict ischaemic CM in patients with systolic dysfunction, as demonstrated by coronary angiography. We constructed and validated three predictive models: (1) a logistic regression “full feature model” incorporating all echocardiographic, electrocardiographic and demographic features gathered; (2) a full feature ANN model and (3) a strain-based model incorporating only segmental peak strain and the presence of a wall motion abnormality.
Both ANN models were significantly more accurate than the logistic regression model. Logistic regression has been the traditional method for predicting outcome in medicine, for example the use of variables from the Framingham population to predict cardiovascular risk [Citation24]. Increasingly, artificial intelligence or machine learning methods such as support vector machines and artificial neural networks are being used for outcome prediction in medicine [Citation25–27]. Contexts in which artificial neural networks have demonstrated enhanced predictive accuracy over logistic regression models include: prediction of acute coronary syndromes using clinical features [Citation20] and significant coronary artery disease in patients with normal ECGs and biomarkers using clinical features [Citation27]. Here, we have demonstrated that the differentiation of underlying ischaemic and non-ischaemic CM in heart failure is another clinical context in which artificial intelligence offers enhanced predictive value over traditional predictive methods. Non-linear modelling such as ANN is likely to become more utilized in medicine. Similar to other statistical methods, tools for this analysis are available on several commercial statistical programs, making the design of ANN analysis widely accessible.
Interestingly, the strain-based ANN model was as good as the full feature ANN model for predicting ischaemic CM. Therefore, ECG features, demographics, PSS and ESL did not add incremental validity.
ECG parameters often show poor predictive accuracy when used to predict cardiovascular outcome [Citation6,Citation28,Citation29]. Our study demonstrates another context in which ECG features provide poor incremental validity in predicting outcome.
PSS has been shown to be present in acute ischaemia, however it tends not to persist for a prolonged period following recovery [Citation30,Citation31]. In our cohort, many of the patients may not have had their echocardiogram during or soon after an episode of acute ischaemia. Further, PSS may not be specific to the acute ischaemic myocardium, as indicated by our data () and other work demonstrating PSS in non-ischaemic cardiomyopathies [Citation32,Citation33]. These factors may explain a lack of incremental validity offered by PSS in our ANN model. ESL has been linked to coronary artery disease [Citation34], however in our cohort, although generally more prevalent in the ischaemic group it was also prevalent in the non-ischaemic group (), which may explain why ESL did not provide added predictive value in our model.
Our strain-based ANN model likely picks up on subtle patterns in peak strain, such as the pattern of relatively reduced apical strain in comparison to basal strain in the ischaemic CM group. The outcome is also likely reinforced or weakened with the presence or absence of a regional wall motion abnormality, a feature that alone has been shown to be a poorly sensitive and specific predictor of ischaemic CM [Citation3]. This strain-based ANN model showed good sensitivity (89%), specificity (76%), positive predictive value (96%), but a weaker negative predictive value (66%), meaning a positive result of ischaemic CM makes a true positive highly likely. However, the high positive predictive value and low negative predictive value are likely influenced by the relatively high prevalence of ischaemic CM in the study population.
Several possibilities could be explored to enhance the predictive accuracy of the strain-based ANN model. Firstly, additional features could be explored to add incremental validity to the model. As ECG and demographic features showed poor incremental validity in this study, perhaps the best candidates for increasing predictive accuracy would be other echocardiographic features or features identified in other imaging modalities associated with specific types of CM [Citation3]. Secondly, the non-ischaemic outcome represents a heterogeneous group. Perhaps generating a model that can identify more specific types of non-ischaemic CM would enhance accuracy in addition to more precisely guiding subsequent investigation and management. However, this would require a larger sample size than used in the current study due to the large number of diagnoses encapsulated within non-ischaemic CM.
Alternatively, perhaps there would be value in building predictive models that prioritise either specificity or sensitivity. In this study, models were designed to maximise overall predictive accuracy (F1 score), but perhaps in this context a useful model might be one prioritising a high sensitivity for identifying ischaemic cardiomyopathy or a high specificity for identifying non-ischaemic cardiomyopathy, so one can identify patients that do not need an angiogram to a high degree of confidence, accepting as excess of “unnecessary” angiograms.
Some key limitations of this study should be noted. Patients were categorised into either the ischaemic or non-ischaemic CM groups based on burden of coronary artery disease as assessed by invasive coronary angiography and history of myocardial infarction or revascularisation. However, a recent cardiac MRI study found that 11.8% of patients traditionally categorised as ischaemic CM (by angiography) and 1.5% of patients categorised as non-ischaemic CM, demonstrated a pattern of mixed CM on MRI [Citation4]. It is likely, therefore, that a proportion of patients in our cohort have mixed ischaemic and non-ischaemic CM, which would not fit with the binary outcomes of our ANN model. Patterns of mixed CM detected by stain could result in a false positive or false negative in our ANN model, reducing its predictive accuracy.
A further limitation is that this was a single centre study, with a preponderance of subjects from an Eastern Mediterranean ethnic background. Therefore, the strain-based ANN model would require validation in other populations before being utilized. An additional limitation is the potential introduction of attrition bias, as 27 of the 231 eligible subjects could not have speckle-tracking strain analysis due to suboptimal images.
Finally, patients included in the study had an EF <50%, a subsequent coronary angiogram and an ECG, therefore, selection bias may exist. For example, patients not receiving an angiogram due to low risk of coronary artery disease or deemed not fit for invasive coronary angiography would not have been included in the study population. However, a mitigating factor for this is that during the study period at this institution most patients presenting with new onset systolic heart failure (EF <50%) received an invasive coronary angiogram in accordance with appropriateness criteria [Citation35]. Nonetheless, we acknowledge that the high prevalence of ischaemic CM in our cohort does affect the NPV and PPV generated and therefore limits the generalizability of our results. However, the impact of prevalence of ischaemic CM has a minor effect – indeed may not affect – the negative and positive likelihood ratios.
In conclusion, we have developed and validated a strain-based ANN model with good predictive accuracy to differentiate ischaemic and non-ischaemic CM in systolic heart failure. This model has been incorporated into an application: https://cimti.usj.edu.lb/strain/index.html, which will be utilized to validate this ANN model in larger cohorts before it can be used to guide the diagnosis of underlying heart failure and potentially reduce the burden of invasive coronary angiograms.
Supplemental Material
Download MS Word (14.9 KB)Disclosure statement
The authors declare no conflict of interest.
References
- Mosterd A, Hoes AW. Clinical epidemiology of heart failure. Heart. 2007;93(9):1137–1146.
- Ziaeian B, Fonarow GC. Epidemiology and aetiology of heart failure. Nat Rev Cardiol. 2016;13(6):368–378.
- Chrysohoou C, Greenberg M, Stefanidis C. Non-invasive methods in differentiating ischaemic from non-ischaemic cardiomyopathy. A review paper. Acta Cardiol. 2006;61(4):454–462.
- Kim EK, Chang S-A, Choi J-O, et al. Concordant and discordant cardiac magnetic resonance imaging delayed hyperenhancement patterns in patients with ischemic and non-ischemic cardiomyopathy. Korean Circ J. 2016;46(1):41–47.
- McMurray JJV, Adamopoulos S, Anker SD, et al. ESC guidelines for the diagnosis and treatment of acute and chronic heart failure 2012: the task force for the diagnosis and treatment of acute and chronic heart failure 2012 of the European Society of Cardiology. Eur Heart J. 2012;33(14):1787–1847.
- Carey MG, Al-Zaiti SS, Canty JM. Jr. High-risk electrocardiographic parameters are ubiquitous in patients with ischemic cardiomyopathy. Ann Noninvasive Electrocardiol. 2012;17(3):241–251.
- Carey MG, Luisi AJ, Baldwa S, et al. The Selvester QRS Score is more accurate than Q waves and fragmented QRS complexes using the Mason-Likar configuration in estimating infarct volume in patients with ischemic cardiomyopathy. J Electrocardiol. 2010;43(4):318–325.
- Chatard JC, Mujika I, Goiriena JJ, et al. Screening young athletes for prevention of sudden cardiac death: practical recommendations for sports physicians. Scand J Med Sci Sports. 2016;26(4):362–374.
- Lopez C, Ilie CC, Glancy DL, et al. Goldberger’s electrocardiographic triad in patients with echocardiographic severe left ventricular dysfunction. Am J Cardiol. 2012;109(6):914–918.
- Pelto H, Owens D, Drezner J. Electrocardiographic findings suggestive of cardiomyopathy: what to look for and what to do next. Curr Sports Med Rep. 2013;12(2):77–85.
- Jaarsma C, Bekkers SC, Haidari Z, et al. Comparison of different electrocardiographic scoring systems for detection of any previous myocardial infarction as assessed with cardiovascular magnetic resonance imaging. Am J Cardiol. 2013;112(8):1069–1074.
- Wiiala J, Hedstrom E, Kraen M, et al. Diagnostic performance of the Selvester QRS scoring system in relation to clinical ECG assessment of patients with lateral myocardial infarction using cardiac magnetic resonance as reference standard. J Electrocardiol. 2015;48(5):750–757.
- Delcre SD, Di Donna P, Leuzzi S, et al. Relationship of ECG findings to phenotypic expression in patients with hypertrophic cardiomyopathy: a cardiac magnetic resonance study. Int J Cardiol. 2013;167(3):1038–1045.
- Ishii K, Suyama T, Imai M, et al. Abnormal regional left ventricular systolic and diastolic function in patients with coronary artery disease undergoing percutaneous coronary intervention: clinical significance of post-ischemic diastolic stunning. J Am Coll Cardiol. 2009;54(17):1589–1597.
- Adda J, Mielot C, Giorgi R, et al. Low-flow, low-gradient severe aortic stenosis despite normal ejection fraction is associated with severe left ventricular dysfunction as assessed by speckle-tracking echocardiography a multicenter study. Circ Cardiovasc Imaging. 2012;5(1):27–35.
- Tan YT, Wenzelburger F, Lee E, et al. The pathophysiology of heart failure with normal ejection fraction: exercise echocardiography reveals complex abnormalities of both systolic and diastolic ventricular function involving torsion, untwist, and longitudinal motion. J Am Coll Cardiol. 2009;54(1):36–46.
- Bertini M, Ng AC, Antoni ML, et al. Global longitudinal strain predicts long-term survival in patients with chronic ischemic cardiomyopathy. Circ Cardiovasc Imaging. 2012;5(3):383–391.
- Zhang Z. A gentle introduction to artificial neural networks. Ann Trans Med. 2016;4(19):370.
- Hagan MT, Demuth HB, Beale MH. Neural network design. 1st ed. Boston (MA): PWS Publishing.; 1996.
- Harrison RF, Kennedy RL. Artificial neural network models for prediction of acute coronary syndromes using clinical data from the time of presentation. Ann Emerg Med. 2005;46(5):431–439.
- Isma'eel HA, Sakr GE, Habib RH, et al. Improved accuracy of anticoagulant dose prediction using a pharmacogenetic and artificial neural network-based method. Eur J Clin Pharmacol. 2014;70(3):265–273.
- Purwanto , Eswaran C, Logeswaran R, et al. Prediction models for early risk detection of cardiovascular event. J Med Syst. 2012;36(2):521–531.
- Felker GM, Shaw LK, O’Connor CM. A standardized definition of ischemic cardiomyopathy for use in clinical research. J Am Coll Cardiol. 2002;39(2):210–218.
- Truett J, Cornfield J, Kannel W. A multivariate analysis of the risk of coronary heart disease in Framingham. J Chronic Dis. 1967;20(7):511–524.
- Yu KH, Zhang C, Berry GJ. Predicting non-small cell lung cancer prognosis by fully automated microscopic pathology image features. Nat Comm. 2016;7:12474.
- Dilsizian SE, Siegel EL. Artificial intelligence in medicine and cardiac imaging: harnessing big data and advanced computing to provide personalized medical diagnosis and treatment. Curr Cardiol Rep. 2014;16(1):441.
- Isma'eel HA, Cremer PC, Khalaf S, et al. Artificial neural network modeling enhances risk stratification and can reduce downstream testing for patients with suspected acute coronary syndromes, negative cardiac biomarkers, and normal ECGs. Int J Cardiovasc Imaging. 2016;32(4):687–696.
- Laszlo R, Kunz K, Dallmeier D, et al. Accuracy of ECG indices for diagnosis of left ventricular hypertrophy in people >65 years: results from the ActiFE study. Aging Clin Exp Res. 2017;29:875–884.
- Rodrigues JCL, McIntyre B, Dastidar AG, et al. Electrocardiographic detection of hypertensive left atrial enlargement in the presence of obesity: re-calibration against cardiac magnetic resonance. J Hum Hypertens. 2016;30(3):197.
- Asanuma T, Uranishi A, Masuda K, et al. Assessment of myocardial ischemic memory using persistence of post-systolic thickening after recovery from ischemia. JACC Cardiovasc Imaging. 2009;2(11):1253–1261.
- Asanuma T, Nakatani S. Myocardial ischaemia and post-systolic shortening. Heart. 2015;101(7):509–516.
- Sato H, Yoshitomi H, Watanabe N, et al. Visually confirmed post-systolic shortening during the recovery period in four cases of Takotsubo cardiomyopathy. J Echocardiogr. 2014;12(4):159–161.
- Mast TP, Teske AJ, Walmsley J, et al. Right ventricular imaging and computer simulation for electromechanical substrate characterization in arrhythmogenic right ventricular cardiomyopathy. J Am Coll Cardiol. 2016;68(20):2185–2197.
- Smedsrud MK, Sarvari S, Haugaa KH, et al. Duration of myocardial early systolic lengthening predicts the presence of significant coronary artery disease. J Am Coll Cardiol. 2012;60(12):1086–1093.
- Patel MR, Bailey SR, Bonow RO, et al. Appropriate use criteria for diagnostic catheterization: a report of the American College of Cardiology Foundation Appropriate Use Criteria Task Force, Society for Cardiovascular Angiography and Interventions, American Association for Thoracic Surgery, American Heart Association, American Society of Echocardiography, American Society of Nuclear Cardiology, Heart Failure Society of America, Heart Rhythm Society, Society of Critical Care Medicine, Society of Cardiovascular Computed Tomography, Society for Cardiovascular Magnetic Resonance, and Society of Thoracic Surgeons. J Am Coll Cardiol. 2012;59(22):1995–2027.