The problem with your mitral valve is that it looks, on paper, a lot like somebody else's mitral valve. Both of you have severe regurgitation. Both of you have echocardiograms filled with numbers that cross the same clinical thresholds. Both of you sit in the same risk category according to guidelines published by the American College of Cardiology and the European Society of Cardiology. And yet one of you will thrive after surgery, while the other might have been better off waiting. The guidelines cannot tell you which is which. That is not a minor oversight. It is, increasingly, the central unsolved problem in the management of primary mitral regurgitation.
Mitral regurgitation is the most common valvular heart abnormality in the world, affecting more than two per cent of the global population. Degenerative mitral valve disease alone accounts for an estimated 24 million people worldwide, according to a 2021 review in Nature Reviews Cardiology. A 2024 analysis using Global Burden of Disease data, published in the Journal of the American Heart Association, reported an estimated 13.3 million cases of non-rheumatic valvular disease globally in 2021, with the absolute burden continuing to rise as populations age. The condition occurs when the mitral valve, that crucial flap of tissue separating the left atrium from the left ventricle, fails to close properly, allowing blood to leak backwards with every heartbeat. In primary mitral regurgitation, the valve itself is the culprit, typically due to myxomatous degeneration or prolapse. Left untreated, severe cases can lead to heart failure, atrial fibrillation, and death.
The standard clinical approach relies on a set of echocardiographic measurements and symptomatic triggers. Operate when the left ventricular ejection fraction drops below 60 per cent. Operate when the left ventricular end-systolic dimension exceeds 40 millimetres. Operate when symptoms appear. These thresholds have guided cardiac surgeons and cardiologists for decades, and they are not wrong, exactly. They are simply insufficient. Roughly 20 per cent of patients who undergo mitral valve surgery with a pre-operative ejection fraction above 60 per cent still develop post-operative left ventricular dysfunction. The numbers that were supposed to guarantee a good outcome did not deliver.
Now a series of studies is suggesting that artificial intelligence, applied to the same echocardiographic data that clinicians already collect, can identify hidden patient subpopulations whose surgical trajectories diverge in ways that traditional risk stratification completely misses. The implications are striking, not because AI is replacing the cardiologist, but because it is revealing that the disease we call primary mitral regurgitation is actually several diseases masquerading as one.
When Twenty-Four Numbers Tell a Story That Guidelines Cannot
The study that brought this idea into sharp focus was published in JACC: Cardiovascular Imaging in 2023 by Julien Bernard, Naveena Yanamala, and colleagues. The work was supported by the National Science Foundation, the National Institute of General Medical Sciences at the National Institutes of Health, and the Canadian Institutes of Health Research. Their approach was deceptively simple in concept, though computationally sophisticated in execution. They took 24 standard echocardiographic parameters from 400 patients with primary mitral regurgitation across two cohorts, one from France with 243 patients followed for a median of 3.2 years, and one from Canada with 157 patients followed for a median of 6.8 years. These were not exotic measurements requiring specialised equipment. They were the routine numbers that any competent echocardiography laboratory produces during a standard examination: chamber volumes, valve gradients, Doppler velocities, wall thicknesses, strain measurements.
The team then applied unsupervised machine learning, specifically hierarchical clustering, to let the data organise itself into groups without any preconceived notions about what those groups should look like. No clinician told the algorithm what “severe” means. No guideline threshold was imposed. The algorithm simply looked at all 24 measurements simultaneously, something no human clinician can do with equal rigour, and sorted patients into phenogroups based on the mathematical relationships between those parameters.
What emerged were two distinct phenogroups: a high-severity cluster and a low-severity cluster. The French development cohort split into 117 high-severity and 126 low-severity patients; the Canadian validation cohort divided into 87 and 70, respectively. So far, that might sound unremarkable. But the remarkable part came when the researchers examined what happened to patients in each group who did and did not undergo mitral valve surgery.
In the high-severity phenogroup, surgical patients had significantly improved event-free survival compared to non-surgical patients, in both the French cohort (P = 0.047) and the Canadian validation cohort (P = 0.020). Surgery clearly helped these patients. The model suggested that assignment to the high-severity phenogroup predicted a reduction in risk of all-cause mortality following mitral valve surgery.
In the low-severity phenogroup, however, there was no statistically significant difference between surgical and non-surgical patients in either cohort (P = 0.70 and P = 0.50, respectively). The algorithm had identified a population in whom surgery conferred no measurable survival benefit, a finding invisible to conventional risk stratification.
The critical insight is this: many of the patients in the low-severity phenogroup would have been classified as having severe or moderate-to-severe mitral regurgitation by traditional guideline criteria. They crossed the same thresholds. They appeared, by every conventional measure, to need surgery. But the machine learning model, by integrating all 24 parameters simultaneously rather than applying sequential threshold cutoffs, recognised a pattern that human interpretation had missed. These patients shared a combination of chamber dimensions, flow characteristics, and myocardial properties that, taken together, indicated a fundamentally different disease trajectory from their guideline-matched counterparts in the high-severity group.
The Black Box Opens Up
One of the persistent criticisms of machine learning in medicine is the black box problem. If an algorithm says a patient belongs to one group rather than another, but nobody can explain why, clinicians have every right to be sceptical. The stakes in cardiac surgery are too high for blind trust in opaque computational processes. Bernard and colleagues anticipated this concern by employing SHapley Additive exPlanations, or SHAP, an explainable AI technique rooted in cooperative game theory that quantifies the contribution of each individual feature to a given prediction.
SHAP values, derived from a framework originally developed to fairly distribute payouts in cooperative games, assign each input variable a numerical importance score for each individual prediction. This means a clinician can interrogate not just which variables matter in general, but which variables mattered for this specific patient. The technique has become one of the most widely adopted explainability methods in clinical AI, precisely because it bridges the gap between computational complexity and human interpretability.
The SHAP analysis revealed that left ventricular end-diastolic volume, the Doppler E/e-prime ratio (a marker of diastolic filling pressure), mitral regurgitation regurgitant volume, and interventricular septal thickness were the most important parameters driving the phenogroup classification. These are not obscure research variables. They are measurements that echocardiographers record routinely. The difference is that the algorithm weighted and combined them in ways that current guidelines do not. Where guidelines apply binary thresholds to individual parameters, the machine learning model captured continuous, nonlinear interactions between all 24 variables simultaneously, recognising patterns of co-occurrence that sequential threshold-checking inherently misses.
This matters enormously for clinical adoption. A cardiologist looking at a SHAP plot can see exactly which measurements pushed a particular patient into the high-severity or low-severity phenogroup, and by how much. The model is not asking clinicians to trust a mysterious oracle. It is showing them, in quantitative terms, what the data actually say when all the measurements are considered together rather than in isolation. A 2024 review of explainable AI evaluation approaches in cardiology, published in BMC Medical Informatics, noted that SHAP remains the most frequently applied interpretability technique in cardiovascular AI research, though it cautioned that only a minority of studies have involved cardiologists in evaluating the clinical relevance of the explanations produced.
The Bernard study also demonstrated incremental prognostic value over conventional approaches. When the researchers looked specifically at patients who were classified as having severe or moderate-to-severe mitral regurgitation by traditional methods, the phenogrouping approach improved the categorical net reclassification index significantly (P = 0.002). In practical terms, this means the algorithm correctly reclassified patients whose outcomes had been mischaracterised by guideline-based stratification.
Why Ejection Fraction Lies
To understand why machine learning phenogrouping works where guidelines falter, it helps to understand exactly why left ventricular ejection fraction, the cornerstone of current surgical decision-making, is such a flawed metric in the context of mitral regurgitation.
Ejection fraction measures the percentage of blood ejected from the left ventricle with each heartbeat. In a healthy heart, that number typically sits above 55 per cent. Both the ACC/AHA and the ESC/EACTS guidelines recommend surgery when it drops below 60 per cent, reasoning that declining pump function signals irreversible myocardial damage. The problem is that ejection fraction, in a leaking valve, is a fundamentally misleading measurement.
In mitral regurgitation, the left ventricle ejects blood through two outlets simultaneously: forward into the aorta, where it belongs, and backwards through the leaking valve into the left atrium, where it does not. The ejection fraction calculation captures both directions of flow without distinguishing between them. A patient can have a seemingly normal or even elevated ejection fraction while their actual forward output, the blood reaching the rest of their body, is dangerously low. The metric is flattering a failing heart. As a 2024 review in Clinical Cardiology by Neveu and colleagues observed, clinical guidelines remain anchored to ejection fraction despite its well-recognised limitations, including its lack of a consistent pathophysiological basis and its dependency on haemodynamic loading conditions.
This is not a new observation. Research published in the Journal of the American Heart Association by Gaasch and Meyer has shown that forward left ventricular ejection fraction, a calculation that accounts only for antegrade flow, is superior to total ejection fraction in predicting outcomes in primary mitral regurgitation. Patients with a forward ejection fraction below 50 per cent face significantly higher risk of adverse events. A pre-operative forward ejection fraction below 40 per cent was associated with increased risk of post-surgical left ventricular systolic dysfunction. Some researchers have argued that the occurrence of post-operative left ventricular dysfunction was 9 per cent when ejection fraction was 64 per cent or above and left ventricular end-systolic dimension was below 37 millimetres, but jumped to 33 per cent when ejection fraction fell below 64 per cent and end-systolic dimension exceeded 37 millimetres. These data suggest that the current guideline threshold of 60 per cent may itself be set too low.
Similarly, a 2024 study in Frontiers in Cardiovascular Medicine demonstrated that among asymptomatic patients with primary mitral regurgitation and preserved ejection fraction above 60 per cent, machine learning models using ejection fraction, mid-left ventricular circumferential strain rate, left ventricular end-systolic dimension, and left ventricular sphericity predicted that 30 per cent of those patients would develop ejection fraction below 50 per cent after surgery. Nearly a third. These patients looked fine by guideline criteria. They were not fine. The subclinical dysfunction was there, but the conventional measurements were not sensitive enough to detect it.
The machine learning phenogrouping approach sidesteps this problem not by replacing ejection fraction with a better single metric, but by refusing to rely on any single metric at all. By integrating dozens of parameters simultaneously, it captures the complex, nonlinear interactions between chamber volumes, filling pressures, valve haemodynamics, and myocardial function that no individual measurement can represent.
Five Faces of a Single Disease
The Bernard study identified two phenogroups, but the broader body of evidence suggests that primary mitral regurgitation fractures into even more distinct subpopulations when examined through a machine learning lens.
A 2023 study published in Heart by Sungho Kwak and colleagues at three tertiary hospitals in South Korea used latent class analysis on 2,321 patients with severe primary mitral regurgitation who underwent valve surgery, with a separate validation cohort of 692 patients. The analysis incorporated 15 variables spanning demographics, laboratory values, surgical factors, and echocardiographic measurements. Five distinct phenogroups emerged, each with dramatically different long-term outcomes over a median follow-up of 6.0 years, during which 149 patients (9.1 per cent) in the derivation cohort died.
Group 1 consisted of younger patients with the fewest comorbidities, and their five-year survival after surgery was 98.5 per cent. Group 2 comprised predominantly men with left ventricular enlargement, surviving at 96.0 per cent. Group 3, mostly women with rheumatic mitral regurgitation, had a five-year survival of 91.7 per cent. Group 4 were low-risk older patients at 95.6 per cent. And Group 5, high-risk older patients, survived at just 83.4 per cent (P less than 0.001 across all groups). In univariable Cox analysis, age, female sex, atrial fibrillation, left ventricular end-systolic dimensions and volumes, ejection fraction, left atrial dimension, and tricuspid regurgitation peak velocity were all significant predictors of mortality following surgery.
The phenogroups performed comparably to the Mitral Regurgitation International Database score, a validated risk prediction tool, achieving a three-year concordance index of 0.763 versus 0.750 (P = 0.602). But crucially, the phenogroups identified these risk strata through an entirely data-driven process, without relying on the predetermined assumptions baked into existing scoring systems. Patients in Group 3, for example, comprised a subpopulation whose specific risk profile, predominantly female with rheumatic aetiology, might be inadequately weighted by conventional tools designed primarily around degenerative valve disease in Western populations. The findings were reproduced in the validation cohort, lending credibility to the phenogroup structure.
Fibrosis, Strain, and the Hidden Damage
Perhaps the most clinically provocative work in this space comes from Olivier Huttin and colleagues, whose 2023 study in JACC: Cardiovascular Imaging used machine learning phenogrouping in 429 patients with mitral valve prolapse (mean age 54 plus or minus 15 years) to identify profiles associated with myocardial fibrosis and cardiovascular events. Mitral regurgitation was severe in 195 patients, or 45 per cent of the cohort.
Their unsupervised clustering analysis identified four distinct groups. Cluster 1 showed minimal cardiac remodelling with mainly mild regurgitation. Cluster 2 was a transitional group with moderate regurgitation and left atrial enlargement. Clusters 3 and 4 both featured significant left ventricular and left atrial remodelling with severe regurgitation, but they diverged in a critical way: Cluster 4 showed a drop in left ventricular systolic strain, a marker of impaired myocardial contractility, while Cluster 3 did not.
When the researchers correlated these clusters with cardiac magnetic resonance imaging data, Clusters 3 and 4 showed significantly more myocardial fibrosis than Clusters 1 and 2 (P less than 0.0001). Patients in these higher-risk clusters also experienced higher rates of cardiovascular events. The fibrosis finding is particularly important because myocardial fibrosis is largely irreversible. A patient who has already developed significant fibrosis may still benefit from valve repair, but the window for achieving optimal outcomes is narrower. Traditional echocardiographic parameters cannot reliably detect fibrosis, yet the machine learning phenogroups, derived entirely from echocardiographic data, identified patients whose myocardial tissue was silently scarring.
This is where the mechanism underlying improved event-free survival becomes clearer. The phenogrouping algorithm does not merely predict who is at higher risk. It identifies a specific physiological pattern, combining valve haemodynamics, chamber geometry, and myocardial mechanics, that corresponds to a particular stage and trajectory of disease. Patients in the high-severity phenogroup of the Bernard study, or in the fibrosis-associated clusters of the Huttin study, are experiencing a specific constellation of adaptations that makes surgical correction both timely and effective. The surgery works because the underlying tissue has not yet passed the point of irreversible damage, or because the haemodynamic burden is severe enough that its relief produces measurable benefit. In the low-severity phenogroup, by contrast, the disease trajectory is more benign, the tissue damage less advanced, and the natural history more favourable even without intervention.
Huttin's team then translated their complex clustering results into a strikingly simple clinical algorithm based on just three variables: severity of mitral regurgitation, indexed left atrial volume, and left ventricular systolic contractility assessed by strain. This decision-tree classification, validated in an external replication cohort, predicted cardiovascular events better than conventional regression models. An accompanying editorial by Nozomi Kagiyama in the same issue of JACC: Cardiovascular Imaging praised the translation from complex machine learning to a three-variable bedside tool, demonstrating that AI-derived insights do not need to remain trapped inside computational infrastructure. They can reshape clinical practice directly.
Beyond the Valve Itself
The scope of machine learning phenotyping extends beyond primary mitral regurgitation into the broader landscape of mitral valve disease. Teresa Trenkwalder and Mark Lachmann, working from the German Heart Center Munich and the German Centre for Cardiovascular Research (DZHK), published a study in European Heart Journal: Cardiovascular Imaging in 2023 that applied unsupervised agglomerative clustering to 609 patients undergoing transcatheter edge-to-edge repair for mitral regurgitation, with external validation in 817 patients from two additional institutions.
Their analysis, based on eight echocardiographic variables, identified four clusters characterised not by the valve lesion itself, but by the pattern of extra-mitral cardiac damage. Cluster 1 showed isolated mitral valve disease with preserved left ventricular function (ejection fraction 56.5 plus or minus 7.79 per cent) and the best five-year survival at 60.9 per cent. Cluster 2 presented with preserved ventricular function (ejection fraction 55.7 plus or minus 7.82 per cent) but the largest regurgitant orifice area (0.623 plus or minus 0.360 square centimetres) and the highest systolic pulmonary artery pressures (68.4 plus or minus 16.2 millimetres of mercury), surviving at 43.7 per cent. Cluster 3 featured impaired ventricular function (ejection fraction 31.0 plus or minus 10.4 per cent) and enlarged end-systolic dimensions, with five-year survival of 38.3 per cent. Cluster 4, characterised by biatrial dilatation (left atrial volume 312 plus or minus 113 millilitres), had the worst prognosis at 23.8 per cent despite only slightly reduced ventricular function (ejection fraction 51.5 plus or minus 11.0 per cent).
The transcatheter repair significantly reduced pulmonary artery pressure and improved survival in Cluster 1 but did not improve outcomes in Cluster 4, where significant diastolic dysfunction rendered the intervention insufficient. This finding mirrors the surgical pattern observed by Bernard: certain patient subpopulations derive clear benefit from intervention, while others do not, and the distinction is invisible to conventional classification.
What makes the Trenkwalder study particularly illuminating is its demonstration that cardiac damage in mitral regurgitation does not follow a neat, sequential progression. A clinician might assume that patients move from mild disease to moderate remodelling to severe dysfunction in an orderly fashion, but the clustering analysis showed that biatrial dilatation (Cluster 4) could occur even with relatively preserved ventricular function, and that pulmonary hypertension (Cluster 2) could develop independently of ventricular impairment. The machine learning model captured a multidimensional reality that linear clinical reasoning tends to oversimplify.
A Proof of Concept With Precedent
The phenogrouping approach in mitral regurgitation does not exist in a vacuum. It draws on a methodology that has already demonstrated its value in other areas of cardiovascular medicine. A landmark 2019 study by Maja Cikes and colleagues, published in the European Journal of Heart Failure, used unsupervised machine learning to phenogroup 1,106 heart failure patients from the MADIT-CRT trial and identify those most likely to respond to cardiac resynchronisation therapy. Their analysis identified four phenogroups with significantly different baseline characteristics, biomarker values, and treatment responses. Two phenogroups were associated with substantially better treatment effects from CRT (hazard ratios of 0.35 and 0.36, P = 0.0005 and P = 0.001, respectively), while the others showed no significant benefit.
The CRT study established that unsupervised clustering of clinical and imaging data could meaningfully stratify patients for therapeutic response in ways that conventional selection criteria could not. The mitral regurgitation studies extend this principle to surgical and transcatheter valve interventions, applying the same logic: not all patients who meet guideline criteria for treatment will benefit equally, and the differences between responders and non-responders are encoded in patterns that only multidimensional analysis can detect.
Automating the Front Door
While phenogrouping addresses the question of who benefits from intervention, a parallel stream of AI research is tackling an equally important upstream problem: making the initial echocardiographic assessment more accurate, more reproducible, and vastly more scalable.
In 2024, Amey Vrudhula, David Ouyang, and colleagues at Cedars-Sinai Medical Centre published a study in Circulation describing EchoNet-MR, a fully automated, open-source deep learning pipeline for detecting clinically significant mitral regurgitation from transthoracic echocardiograms. The system was trained on 58,614 studies comprising 2,587,538 individual videos and required no manual input, processing raw echocardiographic studies from start to finish. Internally, it achieved an area under the curve of 0.916 for detecting moderate or greater regurgitation and 0.934 for severe regurgitation. When tested externally at Stanford Healthcare on 915 studies comprising 46,890 videos, performance actually improved, with an area under the curve of 0.951 for moderate or greater regurgitation and 0.969 for severe regurgitation.
Building on this, Anita Sadeghpour and colleagues published a study in JACC: Cardiovascular Imaging in January 2025 describing an automated machine learning workflow for grading mitral regurgitation severity using 16 American Society of Echocardiography-recommended parameters. The preferred model used nine parameters, was feasible in 99.3 per cent of cases, completed analysis in approximately 80 seconds per case, and achieved accuracy of 0.97 for distinguishing significant from non-significant regurgitation, with sensitivity of 0.96 and specificity of 0.98. Patients graded as having severe regurgitation by the model had significantly higher one-year mortality (adjusted hazard ratio 5.20, 95 per cent confidence interval 1.24 to 21.9, P = 0.025 compared with mild).
The convergence of these two streams, automated detection and severity grading on one hand, and phenogrouping for surgical decision support on the other, points towards a future in which the entire pathway from echocardiographic acquisition to treatment recommendation could be substantially augmented by artificial intelligence. Bo Xu and Alejandro Sanchez-Nadales, writing in a 2025 editorial in JACC: Cardiovascular Imaging, described this as a “paradigm shift” in how cardiac imaging is performed, interpreted, and applied in patient care.
The Reclassification Problem
The most disquieting implication of these studies is not that AI can predict outcomes. It is that AI reveals how many patients were being misclassified all along.
The Bernard study's categorical net reclassification improvement of P = 0.002 in conventionally severe or moderate-to-severe patients means that a meaningful number of patients were being placed in the wrong prognostic category by existing methods. Some patients classified as needing urgent surgery may not have derived benefit from it. Others who appeared to be safely managed with watchful waiting may have been silently accumulating the kind of cardiac damage, ventricular remodelling, atrial dilatation, myocardial fibrosis, that narrows the window for successful intervention.
The Huttin study compounds this concern. Patients in Cluster 4, those with severe regurgitation and impaired systolic strain, had significantly more myocardial fibrosis. Yet by conventional echocardiographic criteria, many of these patients might not have appeared markedly different from Cluster 3, which shared similar regurgitation severity and remodelling but without the strain impairment. The fibrosis, undetectable by standard measurements alone, was the distinguishing feature, and it was the machine learning algorithm that flagged it.
This reclassification problem is not academic. Mitral valve surgery, whether repair or replacement, carries real operative risk. The debate between early surgery and watchful waiting in asymptomatic patients is one of the most contentious in cardiology precisely because the consequences of getting it wrong run in both directions. Operate too early, and you expose a patient to surgical risk for a condition that might have been safely monitored. Operate too late, and irreversible myocardial damage may have already occurred, diminishing the benefit of intervention. Guidelines from the ACC/AHA and the ESC/EACTS do not fully agree on the thresholds for surgery in asymptomatic patients, a disagreement that itself reflects the inadequacy of current risk stratification. The American guidelines consider mitral valve repair reasonable when the likelihood of a successful repair exceeds 95 per cent with expected mortality below one per cent, whereas the European guidelines consider watchful waiting a safe strategy except in the presence of atrial fibrillation or pulmonary hypertension exceeding 50 millimetres of mercury.
Machine learning phenogrouping offers a potential resolution by replacing the binary question of whether regurgitation meets a severity threshold with the more nuanced question of which pattern of cardiac adaptation a particular patient exhibits. It reframes surgical candidacy from a one-dimensional threshold problem into a multidimensional pattern recognition exercise, one in which the data-driven phenotype carries prognostic information that the individual measurements, taken in isolation, do not.
What Stands Between Here and the Clinic
For all its promise, AI-driven phenogrouping in mitral regurgitation remains in its early stages. The Bernard study, while validated in two independent cohorts, involved a total of only 400 patients. The Kwak study was larger at 2,321 patients, but it was retrospective and limited to three South Korean hospitals. The Huttin study comprised 429 patients. None of these represents the kind of large-scale, prospective, multi-ethnic randomised trial that would be needed to change clinical guidelines.
A 2025 scoping review published in npj Cardiovascular Health examined the landscape of unsupervised machine learning applied to valvular heart disease and concluded that while these approaches consistently provided more detailed insights than traditional guideline-based severity classifications, significant barriers remain. Feature selection varies widely between studies. Validation cohorts are often small and geographically limited. The relationship between computationally derived phenogroups and the biological mechanisms underlying disease progression requires further elucidation.
There is also the question of integration. Even if a phenogrouping model proves robust in large trials, it needs to be embedded in clinical workflow software, connected to echocardiography machines, and made accessible to the cardiologists and cardiac surgeons who make treatment decisions. The automated severity grading systems being developed by groups such as Us2.ai and the Cedars-Sinai team suggest that the infrastructure for AI-augmented echocardiography is taking shape, but the pathway from research algorithm to routine clinical deployment is neither short nor straightforward.
Perhaps the most profound barrier is cultural. Cardiology, like all of medicine, operates on a framework of evidence-based guidelines developed through decades of clinical trials, consensus conferences, and expert deliberation. Machine learning phenogrouping does not simply add a new variable to existing risk scores. It fundamentally challenges the paradigm of threshold-based decision-making that underpins current practice. Asking a clinician to trust a clustering algorithm over guidelines they have trained with for their entire career is asking them to accept a different epistemology of disease, one in which diagnosis is not a categorical label but a position within a multidimensional space of continuous variables.
Yet the data are increasingly hard to ignore. When a machine learning model identifies a patient subpopulation in whom surgery confers no survival benefit, and that finding replicates in an independent cohort on a different continent, and the explanatory AI reveals exactly which echocardiographic parameters drove the classification, the question ceases to be whether this technology has value. The question becomes how quickly clinical practice can adapt to the reality that mitral regurgitation is not one disease, and never was.
References and Sources
Bernard, J., Yanamala, N., Shah, R., et al. “Integrating Echocardiography Parameters With Explainable Artificial Intelligence for Data-Driven Clustering of Primary Mitral Regurgitation Phenotypes.” JACC: Cardiovascular Imaging, 16(10), 1253-1267 (2023). DOI: 10.1016/j.jcmg.2023.02.016.
Kwak, S., Lee, S.A., Lim, J., et al. “Long-term outcomes in distinct phenogroups of patients with primary mitral regurgitation undergoing valve surgery.” Heart, 109(4), 305-313 (2023). DOI: 10.1136/heartjnl-2022-321305.
Huttin, O., Girerd, N., Jobbe-Duval, A., et al. “Machine Learning-Based Phenogrouping in Mitral Valve Prolapse Identifies Profiles Associated With Myocardial Fibrosis and Cardiovascular Events.” JACC: Cardiovascular Imaging, 16(10), 1271-1284 (2023). DOI: 10.1016/j.jcmg.2023.03.009.
Trenkwalder, T., Lachmann, M., et al. “Machine learning identifies pathophysiologically and prognostically informative phenotypes among patients with mitral regurgitation undergoing transcatheter edge-to-edge repair.” European Heart Journal: Cardiovascular Imaging, 2023. DOI: 10.1093/ehjci/jead013.
Vrudhula, A., Duffy, G., Vukadinovic, M., et al. “High-Throughput Deep Learning Detection of Mitral Regurgitation.” Circulation, 150(12), 923-933 (2024). DOI: 10.1161/CIRCULATIONAHA.124.069047.
Sadeghpour, A., Jiang, Z., Hummel, Y.M., et al. “An Automated Machine Learning-Based Quantitative Multiparametric Approach for Mitral Regurgitation Severity Grading.” JACC: Cardiovascular Imaging, 18(1), 1-12 (2025). DOI: 10.1016/j.jcmg.2024.06.011.
Xu, B., Sanchez-Nadales, A. “Artificial Intelligence in Echocardiographic Evaluation of Mitral Regurgitation: Envisioning the Future.” JACC: Cardiovascular Imaging, 18(1), 13-15 (2025). DOI: 10.1016/j.jcmg.2024.05.026.
Coffey, S., Cairns, B.J., Iung, B. “The global epidemiology of valvular heart disease.” Nature Reviews Cardiology, 18, 853-864 (2021). DOI: 10.1038/s41569-021-00570-z.
ACC/AHA. “2020 Guideline for the Management of Patients With Valvular Heart Disease.” Circulation, 143(5), e72-e227 (2021). DOI: 10.1161/CIR.0000000000000923.
Baumgartner, H., et al. “2021 ESC/EACTS Guidelines for the management of valvular heart disease.” European Heart Journal, 43(7), 561-632 (2022). DOI: 10.1093/eurheartj/ehab395.
Neveu, D., et al. “Primary mitral regurgitation: Toward a better quantification on left ventricular consequences.” Clinical Cardiology, 2024. DOI: 10.1002/clc.24190.
Gaasch, W.H., Meyer, T.E. “Forward Left Ventricular Ejection Fraction: A Simple Risk Marker in Patients With Primary Mitral Regurgitation.” Journal of the American Heart Association, 6(11), e006309 (2017). DOI: 10.1161/JAHA.117.006309.
“Phenotyping valvular heart diseases using the lens of unsupervised machine learning: a scoping review.” npj Cardiovascular Health, 2025. DOI: 10.1038/s44325-025-00077-3.
Li, Z., et al. “Global, Regional, and National Burden of Valvular Heart Disease, 1990 to 2021.” Journal of the American Heart Association, 2024. DOI: 10.1161/JAHA.124.037991.
Cikes, M., et al. “Machine learning-based phenogrouping in heart failure to identify responders to cardiac resynchronization therapy.” European Journal of Heart Failure, 21, 74-85 (2019). DOI: 10.1002/ejhf.1333.
Kagiyama, N. “Translating Complex Machine-Learning Phenogrouping Into Simple Algorithm: Atrium, Ventricle, and Fibrosis in Mitral Valve Prolapse.” JACC: Cardiovascular Imaging, 16(10), 1285-1287 (2023). DOI: 10.1016/j.jcmg.2023.07.010.
Tim Green
UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
ORCID: 0009-0002-0156-9795
Email: tim@smarterarticles.co.uk
