Artificial Intelligence for Left Ventricular Diastolic Function Assessment: A New Paradigm on the Horizon

Left ventricular (LV) diastolic function assessment is an important part of the echocardiographic evaluation of patients, especially in those with undifferentiated dyspnea in whom heart failure with preserved ejection fraction (HFpEF) could be the cause. The most recent guidelines aimed to simplify the assessment of diastolic function, but this assessment remains complex and depends on the integration of multiple clinical and echocardiographic variables.1Nagueh S.F. Smiseth O.A. Appleton C.P. et al.Recommendations for the evaluation of left ventricular diastolic function by echocardiography: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging.J Am Soc Echocardiogr. 2016; 29: 277-314Abstract Full Text Full Text PDF PubMed Google Scholar The reliance on numerous parameters, many of which are often either unavailable or inaccurately measured, has made the current algorithm prone to misclassification and far too often rendered “indeterminate,” which provides little added value to the clinician. Given the complexity involved in the assessment of diastolic function, there has been a growing interest in the use of artificial intelligence (AI) to automate the process. In this issue of JASE, Chen et al.2Chen X. Yang F. Zhang P. et al.Artificial intelligence assisted left ventricular diastolic function assessment and Grading：Multi-view versus single-view.J Am Soc Echocardiogr. 2023; 36: 1064-1078Abstract Full Text Full Text PDF Google Scholar present 3 different AI-assisted approaches of assessing diastolic function, all within a single study. The objectives of this editorial are to (1) provide a summary and critical appraisal of this study, (2) position it within the broader context of comparable studies using AI methods for diastolic function evaluation, and (3) highlight some of the remaining unresolved challenges in diastolic function assessment that could potentially be overcome by AI in future studies. In this study, the authors used a retrospective data set of 3,485 studies compiled from 3 hospitals to develop 3 distinct AI models to assess diastolic function: (1) a multiview approach based on guideline algorithms, (2) a single-view approach based on left atrial (LA) and LV metrics, (3) and a single-view approach using a convolutional neural network model of two-dimensional (2D) videos. Investigators then used a prospective data set of 527 studies from 1 of the 3 hospitals to validate the models. The multiview approach consisted of 3 AI-assisted steps: (1) view classification, (2) view segmentation, and (3) input of the machine-derived measurements from the view classification and segmentation models into a rule-based decision tree algorithm based on the recent guidelines. From the retrospective data set of 3,485 studies, 1,304 and 2,238 studies were randomly selected to develop the view classification and view segmentation models, respectively, which were further randomly divided 4:1 for training and validation. The view classification model included an automated image quality control feature that was trained to identify studies that contained 6 essential views with image quality sufficient to allow subsequent analysis. The view segmentation model was then developed to make independent measurements of the highest-yield parameters for diastolic function assessment based on the guidelines: LV ejection fraction (LVEF), LA volume, mitral inflow E and A velocities, maximum tricuspid regurgitation velocity, and mitral annulus septal and lateral e’ velocities. Output variables from the view classification and segmentation models were then used as input variables for the rule-based decision tree algorithm according to the guidelines. Two single-view approaches based on the apical 4-chamber view alone were then developed to overcome the need for multiple views, which can either be unavailable or introduce several sources of potential measurement error. Of the 3,485 studies in the retrospective cohort, 2,150 (62%) were used to train the single-view approaches after excluding cases with conditions that interfere with diastolic function assessment including those with incomplete data or insufficient image quality. The models were trained by a supervised learning method that used the diastolic function grade determined by 2 experts based on the most recent guidelines as the ground truth. The first single-view approach involved two AI-assisted steps: (1) segmentation of LA and LV volumes and measurement of LVEF and longitudinal strain and (2) training the model (using a support vector machine algorithm) to identify which combination of LA and LV metrics resulted in the most accurate identification and grading of diastolic dysfunction. The second single-view approach used a deep learning three-dimensional convolutional neural network to train the model to identify and grade diastolic dysfunction directly from the 2D apical 4-chamber videos without expert segmentation. In this black box approach, the model presumably identifies elements within the 2D videos that lead to the accurate identification and grading of diastolic dysfunction. Insights into the key elements used by the model were then provided by Grad-CAM heat maps that highlight the regions of interest. To better appreciate the contribution of the current study to the field, it is important to place it in the context of other similar studies. Shown in Table 1 is a selection of prior studies that have demonstrated a novel application of AI to guide the assessment of LV diastolic function. The multiview approach in the current study combines elements from 2 prior studies: Tromp et al.3Tromp J. Seekings P.J. Hung C.L. et al.Automated interpretation of systolic and diastolic function on the echocardiogram: a multicohort study.Lancet Digit Health. 2022; 4: e46-e54Abstract Full Text Full Text PDF PubMed Scopus (36) Google Scholar developed an automated workflow that classified and segmented 2D videos and Doppler modalities to produce automated measurements of LVEF, medial and lateral e’ velocities, and average E/e’ ratio using a training data set of 1,145 echocardiograms; and Yeung et al.4Yeung D.F. Jiang R. Behnami D. et al.Impact of the updated diastolic function guidelines in the real world.Int J Cardiol. 2021; 326: 124-130Abstract Full Text Full Text PDF PubMed Scopus (9) Google Scholar used a rules-based computer algorithm to grade diastolic function in >70,000 echo studies based on measurements performed by experts. The first single-view approach presented in the current study that used automatically derived LA and LV strain measurements from the apical 4-chamber view alone to grade diastolic function represents an extension of the findings from 3 prior studies: Tabassian et al.5Tabassian M. Sunderji I. Erdei T. et al.Diagnosis of heart failure with preserved ejection fraction: machine learning of spatiotemporal variations in left ventricular deformation.J Am Soc Echocardiogr. 2018; 31: 1272-1284.e9Abstract Full Text Full Text PDF PubMed Scopus (77) Google Scholar developed a machine-learning algorithm that used spatiotemporal patterns in LV deformation based on manual tracings of the left ventricle (LV) in the apical 4-chamber, 2-chamber, and long-axis views to identify patients with HFpEF; Chiou et al.6Chiou Y.A. Hung C.L. Lin S.F. AI-assisted echocardiographic prescreening of heart failure with preserved ejection fraction on the basis of intrabeat dynamics.JACC Cardiovasc Imaging. 2021; 14: 2091-2104Crossref PubMed Scopus (16) Google Scholar developed a model that automatically segmented LA and LV length, width, area, and volume throughout the cardiac cycle from the apical 4-chamber view and used the dynamic changes in the LA and LV to identify HFpEF; and Carluccio et al.7Carluccio E. Cameli M. Rossi A. et al.Left atrial strain in the assessment of diastolic function in heart failure: a machine learning approach.Circ Cardiovasc Imaging. 2023; 16e014605Crossref PubMed Scopus (4) Google Scholar used an unsupervised learning approach that identified peak atrial longitudinal strain as having incremental prognostic value in diastolic function assessment. The second single-view approach is perhaps the most novel aspect of the current study, as it employs deep learning convolutional neural network methods to train the model to grade diastolic function from a single 2D video without segmentation. Asch et al.8Asch F.M. Poilvert N. Abraham T. et al.Automated echocardiographic quantification of left ventricular ejection fraction without volume measurements using a machine learning algorithm mimicking a human expert.Circ Cardiovasc Imaging. 2019; 12e009303Crossref Scopus (87) Google Scholar applied a similar technique to train a machine-learning algorithm to bypass the border detection and segmentation step and automatically estimate LVEF from apical 4-chamber and 2-chamber views similar to a human expert who provides a visual estimate. However, this black box technique has yet to be applied to the assessment of diastolic function until now.Table 1Notable studies with novel applications of AI to guide LV diastolic function assessmentStudyMachine-learning techniquesInput variablesGround truthFurther validationModel prediction/novel AI applicationTabassian et al. (2018)5Tabassian M. Sunderji I. Erdei T. et al.Diagnosis of heart failure with preserved ejection fraction: machine learning of spatiotemporal variations in left ventricular deformation.J Am Soc Echocardiogr. 2018; 31: 1272-1284.e9Abstract Full Text Full Text PDF PubMed Scopus (77) Google ScholarUnsupervised statistical method; supervised classifierLV velocity, strain, strain rate from the A4C, A2C, A3CPatients with HFpEF; breathless subjects with no cardiac disease; hypertensive patients with no symptoms; healthy subjectsN/AModel uses LV mechanics from apical views to identify patients with a clinical diagnosis of HFpEF.Pandey et al. (2021)9Pandey A. Kagiyama N. Yanamala N. et al.Deep-learning models for the echocardiographic assessment of diastolic dysfunction.JACC Cardiovasc Imaging. 2021; 14: 1887-1900Crossref PubMed Scopus (40) Google ScholarUnsupervised topological data analysis; supervised classifierLVEF, LVMI, E, A, E/A, e', E/e', LAVI, TRVHigh- and low-risk phenogroups labeled using unsupervised similarity clusteringLV filling pressure; HF hospitalization, all-cause death; cTnI, NT-proBNP; peak VO2, MLHFQModel identifies high- and low-risk phenogroups based on 9 diastolic function parameters, validated by: invasive measurements; clinical outcomes; cardiac biomarkers; and exercise/QoL measures.Chiou et al. (2021)6Chiou Y.A. Hung C.L. Lin S.F. AI-assisted echocardiographic prescreening of heart failure with preserved ejection fraction on the basis of intrabeat dynamics.JACC Cardiovasc Imaging. 2021; 14: 2091-2104Crossref PubMed Scopus (16) Google ScholarSupervised learningLA and LV length, width, area, and volume changes from A4CPatients with HFpEF; asymptomatic subjects without HFpEF and normal LVEFN/AModel distinguishes patients with HFpEF from those with COPD based on abnormal patterns in LA and LV intrabeat dynamics.Tromp et al. (2022)3Tromp J. Seekings P.J. Hung C.L. et al.Automated interpretation of systolic and diastolic function on the echocardiogram: a multicohort study.Lancet Digit Health. 2022; 4: e46-e54Abstract Full Text Full Text PDF PubMed Scopus (36) Google ScholarSupervised and unsupervised CNNs for view classification; supervised learning for segmentationEcho videos from A2C, A4C, PLAX; echo images from Doppler modalities including PWTDI for e', PW for E, CW for TRVExpert view classification and segmentation to measure LVESV, LVEDV, LVEF, LAESV, E, e', E/e'N/AAutomated workflow classifies and segments most of the traditional diastolic function parameters from echo videos and images.Jiang et al. (2022)11Jiang R. Yeung D.F. Behnami D. et al.A novel continuous left ventricular diastolic function score using machine learning.J Am Soc Echocardiogr. 2022; 35: 1247-1255Abstract Full Text Full Text PDF PubMed Scopus (6) Google ScholarSupervised learning including transfer learningMyocardial disease, LVEF, LAVI, E, A, E/A, e', E/e’, TRVDiastolic function grade determined by a software algorithm programmed to follow the current guidelinesN/AModel produces a continuous diastolic function score based on current diastolic function guidelines.Chao et al. (2022)10Chao C.J. Kato N. Scott C.G. et al.Unsupervised machine learning for assessment of left ventricular diastolic function and risk stratification.J Am Soc Echocardiogr. 2022; 35: 1214-1225.e8Abstract Full Text Full Text PDF PubMed Scopus (6) Google ScholarUnsupervised learningE, A, E/A, DT, e', E/e', TRVN/AHF hospitalization and all-cause mortalityModel identifies 3 phenotype clusters based on 9 diastolic function parameters corresponding to normal diastolic function, impaired relaxation, and increased filling pressure, validated by clinical outcomes.Carluccio et al. (2023)7Carluccio E. Cameli M. Rossi A. et al.Left atrial strain in the assessment of diastolic function in heart failure: a machine learning approach.Circ Cardiovasc Imaging. 2023; 16e014605Crossref PubMed Scopus (4) Google ScholarUnsupervised cluster analysisLVEF, LAESV, E, A, E/A, DT, e', E/e', TRV, PALSN/AHeart failure hospitalization and all-cause deathModel identifies 3 phenotype clusters based on traditional echo diastolic function parameters, as well as PALS, validated by clinical outcomes.Chen et al. (2023)2Chen X. Yang F. Zhang P. et al.Artificial intelligence assisted left ventricular diastolic function assessment and Grading：Multi-view versus single-view.J Am Soc Echocardiogr. 2023; 36: 1064-1078Abstract Full Text Full Text PDF Google ScholarSupervised three-dimensional CNNEcho video of A4CDiastolic function grade by 2 experts based on current guidelinesN/AModel predicts diastolic function grade based on a single echo video (A4C alone).A, peak late mitral inflow velocity; A2C, apical 2-chamber view; A3C, apical 3-chamber view; A4C, apical 4-chamber view; CNN, convolutional neural network; COPD, chronic obstructive pulmonary disease; cTnI, cardiac troponin I; CW, continuous-wave Doppler; DT, deceleration time; E, peak early mitral inflow velocity; e’, mitral annulus velocity; HF, heart failure; LAESV, LA end-systolic volume; LAVI, LA volume index; LS, longitudinal strain; LVEDV, LV end-diastolic volume; LVESV, LV end-systolic volume; LVMI, LV mass index; MLHFQ, Minnesota living with heart failure questionnaire; MV, mitral valve; N/A, not applicable; NT-proBNP, N-terminal prohormone of brain natriuretic peptide; PALS, peak atrial longitudinal strain; PLAX, parasternal long-axis view; PW, pulsed-wave Doppler; PWTDI, pulsed-wave tissue Doppler imaging; TRV, peak tricuspid regurgitation velocity; TV, tricuspid valve. Open table in a new tab A, peak late mitral inflow velocity; A2C, apical 2-chamber view; A3C, apical 3-chamber view; A4C, apical 4-chamber view; CNN, convolutional neural network; COPD, chronic obstructive pulmonary disease; cTnI, cardiac troponin I; CW, continuous-wave Doppler; DT, deceleration time; E, peak early mitral inflow velocity; e’, mitral annulus velocity; HF, heart failure; LAESV, LA end-systolic volume; LAVI, LA volume index; LS, longitudinal strain; LVEDV, LV end-diastolic volume; LVESV, LV end-systolic volume; LVMI, LV mass index; MLHFQ, Minnesota living with heart failure questionnaire; MV, mitral valve; N/A, not applicable; NT-proBNP, N-terminal prohormone of brain natriuretic peptide; PALS, peak atrial longitudinal strain; PLAX, parasternal long-axis view; PW, pulsed-wave Doppler; PWTDI, pulsed-wave tissue Doppler imaging; TRV, peak tricuspid regurgitation velocity; TV, tricuspid valve. Chen et al. should be commended for this ambitious study, which has many strengths. First, their study addresses a clinically important need, to explore ways in which AI can streamline the complex assessment of diastolic function. In addition, the concepts underlying their AI-assisted approaches are intuitive and build on the knowledge gained from prior work, combining select AI elements from previous studies to produce additional insights. Furthermore, the development of 3 distinct AI-assisted approaches to grade diastolic function within the same study and using a prospective data set for validation allowed for direct comparison of the relative merits of each method. In their current form, the multiview method still outperformed the single-view methods with respect to identifying and grading diastolic dysfunction, with accuracies of 90% and 92% for the multiview method, 83% and 75% for the strain-based single-view method, and 85% and 80% for the video-based single-view method, respectively. This is not surprising since the multiview method was modeled to perform a rule-based decision tree algorithm similar to that used by human experts, which was the chosen ground truth in this study. On the other hand, the single-view methods offer the convenience of requiring only 1 echo video, significantly reducing the average processing time (18.5 seconds for the multiview method, 8.8 seconds for the strain-based single-view method, and 4.3 seconds for the video-based single-view method). Despite the strengths of the study, it is important to recognize its limitations. The most obvious and important limitation that the authors themselves acknowledge is the lack of a more robust gold standard on which to either train or validate the models. The gold standard used in the study involved the expert interpretation of diastolic function based on the most recent guidelines. In contrast, 2 prior studies used more clinically meaningful methods for validation: Pandey et al.9Pandey A. Kagiyama N. Yanamala N. et al.Deep-learning models for the echocardiographic assessment of diastolic dysfunction.JACC Cardiovasc Imaging. 2021; 14: 1887-1900Crossref PubMed Scopus (40) Google Scholar used supervised and unsupervised machine-learning techniques to identify high- and low-risk phenotype subgroups of patients with HFpEF based on invasive hemodynamic measurements of LV filling pressure on cardiac catheterization, clinical outcomes of heart failure hospitalization and all-cause mortality, cardiac biomarkers, and exercise performance; and Chao et al.10Chao C.J. Kato N. Scott C.G. et al.Unsupervised machine learning for assessment of left ventricular diastolic function and risk stratification.J Am Soc Echocardiogr. 2022; 35: 1214-1225.e8Abstract Full Text Full Text PDF PubMed Scopus (6) Google Scholar used an unsupervised learning approach that identified 3 distinct phenotype clusters of conventional echocardiographic diastolic parameters based on clinical outcomes of heart failure hospitalization and all-cause mortality. Without a more robust clinical measure for training or validation, it is challenging to adequately evaluate the legitimacy of models presented in this study, particularly the black box video-based single-view method. While the Grad-CAM heat map feature provides some degree of reassurance that the model is focusing on the most relevant cardiac structures for diastolic function assessment, it also highlighted unexpected components such as the LV apex. It is unclear whether this is a truly novel region of interest that contributes to LV diastolic function or whether it simply represents a distracting element identified by an imperfect model. In addition, by virtue of the models being trained based on the current guidelines, they are also prone to the same tendency to grade a proportion of cases as indeterminate, which is a major limitation of the current guideline algorithm. All 3 AI-assisted models in this study still demonstrated the potential to grade diastolic function as indeterminate. Unsupervised learning models based on clinical outcomes, as demonstrated by the works of Pandey et al.9Pandey A. Kagiyama N. Yanamala N. et al.Deep-learning models for the echocardiographic assessment of diastolic dysfunction.JACC Cardiovasc Imaging. 2021; 14: 1887-1900Crossref PubMed Scopus (40) Google Scholar and Chao et al.,10Chao C.J. Kato N. Scott C.G. et al.Unsupervised machine learning for assessment of left ventricular diastolic function and risk stratification.J Am Soc Echocardiogr. 2022; 35: 1214-1225.e8Abstract Full Text Full Text PDF PubMed Scopus (6) Google Scholar offer a viable approach to optimize the classification of diastolic function with reduction of cases being labeled as “indeterminate.” Alternatively, approaches that rely on a semisupervised learning strategy may be used to generate weak labels for indeterminate cases. In addition, Jiang et al.11Jiang R. Yeung D.F. Behnami D. et al.A novel continuous left ventricular diastolic function score using machine learning.J Am Soc Echocardiogr. 2022; 35: 1247-1255Abstract Full Text Full Text PDF PubMed Scopus (6) Google Scholar developed a model that can produce a continuous diastolic function score based on inputs of traditional diastolic parameters on echocardiography, even for cases that would have been labeled indeterminate according to the current guidelines. Further validation of this model using invasive hemodynamics and clinical outcomes will establish its precision, reproducibility, and robustness. With the above considerations in mind, what are the key messages that we can take away from the current study and what should we expect from future studies? Chen et al. provided us with at least 3 important insights: (1) among patients without conditions that interfere with diastolic function assessment, interpretation of diastolic function based on the current guideline algorithm can be nearly fully automated using the multiview approach with ≥90% accuracy; (2) while some prior studies have suggested LA strain as a potential single measure surrogate to assess diastolic function, the optimal method probably involves a combination of LA and LV mechanics in addition to conventional diastolic parameters; (3) finally, and perhaps most importantly, yet another single-view method relying on an entirely different AI technique—this time using a deep learning convolutional neural network—has shown potential to effectively assess diastolic function. Establishing a reliable single-view method to assess diastolic function could have a profound clinical impact, not only by reducing the number of indeterminate cases due to insufficient views but also by facilitating point-of-care diastolic function assessment to guide clinical decision-making. Future studies should work toward refining these models. One way to improve the accuracy of the models is to minimize the risk for misclassification by designing AI models to be robust to out-of-distribution data (i.e., those echocardiography data that cannot be strongly linked to the data used for training those models). Strategies that can automatically detect such out-of-distribution data, for example, based on Bayesian deep networks, should be incorporated in the routine clinical usage of such AI models.12Kazemi Esfeh M.M. Luong C. Behnami D. et al.A deep bayesian video analysis framework: towards a more robust estimation of ejection fraction.in: Medical Image Computing and Computer Assisted Intervention – MICCAI 2020. MICCAI 2020. Lecture Notes in Computer Science. 12262. Springer, Cham2020Crossref Scopus (5) Google Scholar Another way to improve the models is to continue systematically validating them against relevant clinical outcomes such as heart failure hospitalization and mortality. Finally, future iterations would ideally incorporate the incrementally predictive features of each of the previously developed models. On the horizon, we envision a model capable of analyzing LA and LV mechanics in addition to conventional diastolic parameters to automatically generate a prognostically meaningful diastolic function score that could be used even in point-of-care devices to guide management. Until then, this study has taken yet another step toward that goal, providing further insights into how AI can facilitate diastolic function assessment. Artificial Intelligence–Assisted Left Ventricular Diastolic Function Assessment and Grading: Multiview Versus Single ViewJournal of the American Society of EchocardiographyVol. 36Issue 10PreviewClinical assessment and grading of left ventricular diastolic function (LVDF) requires quantification of multiple echocardiographic parameters interpreted according to established guidelines, which depends on experienced clinicians and is time consuming. The aim of this study was to develop an artificial intelligence (AI)–assisted system to facilitate the clinical assessment of LVDF. Full-Text PDF Open Access