Artificial intelligence-derived risk score for mortality in secondary mitral regurgitation treated by transcatheter edge-to-edge repair: the EuroSMR risk score

Abstract

Background and Aims

Risk stratification for mitral valve transcatheter edge-to-edge repair (M-TEER) is paramount in the decision-making process to appropriately select patients with severe secondary mitral regurgitation (SMR). This study sought to develop and validate an artificial intelligence-derived risk score (EuroSMR score) to predict 1-year outcomes (survival or survival + clinical improvement) in patients with SMR undergoing M-TEER.

Methods

An artificial intelligence-derived risk score was developed from the EuroSMR cohort (4172 and 428 patients treated with M-TEER in the derivation and validation cohorts, respectively). The EuroSMR score was validated and compared with established risk models.

Results

The EuroSMR risk score, which is based on 18 clinical, echocardiographic, laboratory, and medication parameters, allowed for an improved discrimination of surviving and non-surviving patients (hazard ratio 4.3, 95% confidence interval 3.7–5.0; P < .001), and outperformed established risk scores in the validation cohort. Prediction for 1-year mortality (area under the curve: 0.789, 95% confidence interval 0.737–0.842) ranged from <5% to >70%, including the identification of an extreme-risk population (2.6% of the entire cohort), which had a very high probability for not surviving beyond 1 year (hazard ratio 6.5, 95% confidence interval 3.0–14; P < .001). The top 5% of patients with the highest EuroSMR risk scores showed event rates of 72.7% for mortality and 83.2% for mortality or lack of clinical improvement at 1-year follow-up.

Conclusions

The EuroSMR risk score may allow for improved prognostication in heart failure patients with severe SMR, who are considered for a M-TEER procedure. The score is expected to facilitate the shared decision-making process with heart team members and patients.

Structured Graphical Abstract

The EuroSMR risk score. The EuroSMR risk score was developed from the extended EuroSMR registry, including a total of 4600 patients in the derivation (4172 patients) and validation (428 patients) cohorts. The artificial intelligence-based algorithm identified 18 risk parameters associated with 1-year survival after mitral valve transcatheter edge-to-edge repair. The application of the EuroSMR risk score performed similarly with respect to survival prediction in the derivation and validation cohorts. With an increasing EuroSMR risk score, the predicted probability for mortality with or without New York Heart Association class improvement is increasing. BMI, body mass index; COPD, chronic obstructive pulmonary disease; eGFR, estimated glomerular filtration rate; EROA, effective regurgitant orifice area; GDMT, guideline-directed medical therapy; LA, left atrial; LVEDV, left ventricular end-diastolic volume; LVEF, left ventricular ejection fraction; M-TEER, mitral valve transcatheter edge-to-edge repair; NT, proBNP, N-terminal pro-brain natriuretic peptide; NYHA, New York Heart Association; RA, right atrial; RV, right ventricular; SHAP, SHapley Additive exPlanations; sPAP, systolic pulmonary artery pressure; TAPSE, tricuspid annular plane systolic excursion; TR, tricuspid regurgitation.

Open in new tabDownload slide

See the editorial comment for this article ‘Scoring systems developed by machine learning: intelligent but simple to use?’, by A. Banerjee and P. Leeson, https://doi.org/10.1093/eurheartj/ehae053.

Introduction

Secondary mitral regurgitation (SMR) in patients with heart failure is a multifaceted valve disorder with a strong impact on life expectancy and quality of life.^1,2 While mitral valve transcatheter edge-to-edge repair (M-TEER) has become the first-line interventional treatment option for SMR in selected high-risk patients, the two most recent randomized trials in this setting indicated a high mid-term mortality rate, with one study failing to demonstrate a survival benefit following M-TEER compared with optimal guideline-directed medical therapy (GDMT).^3,4 Thus, better characterization and stratification are needed to identify SMR patients who will benefit clinically and in terms of survival from a M-TEER procedure.

Current risk scores for SMR patients evaluated for M-TEER are still suboptimal.⁵ Scores predicting perioperative mortality for surgical patients, including the EuroSCORE II and Society of Thoracic Surgeons score, showed weak stratification capabilities in M-TEER populations. As for scores specifically designed for M-TEER patients, the MitraScore⁶ for instance does not differentiate between patients with primary and secondary mitral regurgitation (MR), two different pathophysiologic entities with different survival prognosis.⁷ The recently developed Cardiovascular Outcomes Assessment of the MitraClip Percutaneous Therapy for Heart Failure Patients with Functional Mitral Regurgitation (COAPT) score has not been validated in a separate and independent patient cohort with SMR.⁵ Furthermore, all dedicated M-TEER scores were derived from relatively small patient cohorts, which may also limit the clinical applicability to a broader patient population.

The purpose of the current study was to develop and validate a comprehensive risk score for the prediction of mortality and clinical outcomes from the large extended EuroSMR patient population. The herein proposed EuroSMR risk score employs machine learning, an approach that differs from conventional statistical methodologies. Machine learning enables the integration of multifactorial relationships amongst numerous predictors, thereby enhancing a model’s ability to fit complex data patterns associated with increased mortality following M-TEER. However, it should be noted that while machine learning can incorporate these relationships, it does not explicitly reveal the specific underlying predictor combinations. To bridge this gap, we have also applied techniques to better understand the model’s predictions, aligning with the imperative of providing explainable patient information. Furthermore, this EuroSMR risk score, which will allow for a detailed patient information and shared decision-making within the interdisciplinary heart team regarding planned M-TEER procedure, was also developed to identify an SMR patient subgroup at extreme risk for mortality, for which the procedure may be considered futile.

Methods

Study cohort and procedural technique

The study population comprised SMR patients undergoing a M-TEER procedure and were included in several patient cohorts, which formed the extended EuroSMR (European Registry of Transcatheter Repair for Secondary Mitral Regurgitation) registry. Patients from the initial EuroSMR registry were included at 12 cardiac centres across Europe from 2008 to 2019.⁸ Additionally, SMR patients from the GIOTTO [Italian Society of Interventional Cardiology (GIse) registry Of Transcatheter treatment of mitral valve regurgitaTiOn] registry,⁹ SMR patients from the Spanish MitraScore registry,⁶ and SMR patient from the German MitraPro registry were included in the extended EuroSMR cohort. The data from the above four registries defined the derivation cohort in this analysis. Data from study sites, who participated in more than one of the above registries, were only considered once. Patients from two additional study sites, which recently joined the initial EuroSMR registry, were exclusively evaluated as a validation cohort and not included in the process of risk score development. Of note, the EuroSMR registry retrospectively collected M-TEER patients who were treated at the respective study sites according to internal standards and international guidelines without a dedicated prospective study protocol. The EuroSMR registry was registered at ‘Deutsches Register Klinischer Studien’ (DRKS00017428).

Patients, who remained symptomatic despite GDMT and, if applicable, cardiac resynchronization therapy, were discussed in an interdisciplinary heart team at the respective study centre. Treatment indications for interventional treatment were established by heart team consensus considering individual history, comorbidities, and age. The M-TEER treatment was performed using the MitraClip device (Abbott Structural Heart, Santa Clara, CA, USA) as previously described.^10–12

Study variables

All clinical, laboratory, and echocardiographic parameters as well as medication data were collected at the index hospitalization or within 4 weeks prior to the M-TEER procedure. Clinical patient characteristics included demographic data (age, sex, and body mass index), comorbidities (diabetes mellitus, arterial hypertension, coronary artery disease, atrial fibrillation, and chronic obstructive pulmonary disease), renal function (estimated glomerular filtration rate), N-terminal pro-brain natriuretic peptide (NT-proBNP) and haemoglobin at baseline, information on GDMT, and New York Heart Association (NYHA) functional class.

Triple GDMT was defined as concurrent prescription of three heart failure medication classes, i.e. renin–angiotensin system inhibitors (RASIs), mineralocorticoid receptor antagonists, and beta-blockers at the time of M-TEER after heart team evaluation. The prescription of an angiotensin receptor–neprilysin inhibitor (ARNI) was added to guideline-recommended medication in the later inclusion period of this registry. Consequently, ARNI was considered in the category of RASI for this analysis. The use of sodium–glucose co-transporter 2 inhibitors was not analysed in this registry due to late adoption in recent HF guidelines.

Transthoracic echocardiographic evaluation was performed by experienced operators at each study site in line with respective guidelines.¹³ SMR severity was assessed applying an integrative four-grade approach of quantitative, semi-quantitative, and qualitative parameters according to the European recommendations for the assessment of native valvular regurgitation.¹³ Baseline evaluation of SMR by transthoracic echocardiography included quantitative [effective regurgitant orifice area (EROA) and regurgitant volume by proximal isovelocity surface area method] and semi-quantitative parameters (vena contracta width and 3D vena contracta area from multi-planar reconstruction by transoesophageal echocardiography, if applicable). Left ventricular (LV) function and size were described by LV ejection fraction (LVEF) and LV end-diastolic volume (LVEDV). A LVEF of ≥50% was used for defining a SMR patient subgroup with an atrial SMR phenotype, while a LVEF of <50% defined a ventricular SMR phenotype. Left atrial volume was assessed using the Simpson’s biplane method. Evaluation of the right heart included right ventricular midventricular diameter, right atrial area, tricuspid annular plane systolic excursion (TAPSE), and tricuspid regurgitation (TR) grade. Systolic pulmonary artery pressure (sPAP) was estimated by the addition of peak systolic TR pressure gradient to estimated right atrial pressure derived from the inferior vena cava width and respiratory variability. Right ventricular to pulmonary artery coupling was calculated as the ratio of TAPSE to sPAP, as described previously.^8,14

Endpoints, follow-up, and definitions

One-year all-cause mortality was the main endpoint for the development of the EuroSMR score in the derivation cohort. This endpoint was also used for validation of the EuroSMR score, as well as for comparison with other risk scores in the validation cohort. In an additional secondary analysis, we assessed the combined endpoint of mortality and need for heart failure hospitalization up to 2 years following the procedure in the validation cohort for comparability with the COAPT risk score. Furthermore, we assessed the composite endpoints of 1-year mortality and NYHA class IV or mortality and NYHA class ≥III at 1-year follow-up in the extended EuroSMR cohort for the assessment of M-TEER utility and futility. Depending on the protocol of the individual study centres, follow-up was conducted with patients or their next of kin, either during regular patient visits or by telephone interview. Follow-up studies included at least NYHA class and survival status.

The COAPT risk score was assessed as described by Shah et al.⁵ for patients without missing data. The following parameters were used to identify patients as ‘COAPT-like’¹⁵: SMR severity of ≥3+, NYHA functional class ≥II, LVEF of ≥20% and ≤50%, LV end-systolic diameter of ≤70 mm, TR of ≤2+, sPAP of ≤70 mmHg (echocardiographic assessment), and preserved right ventricular function as assessed by TAPSE of ≥15 mm.¹⁶ The evaluation of COAPT likeliness was only undertaken for patients with comprehensive data for all these parameters. Patients who adhered to all the specified criteria were termed ‘COAPT-like’. In contrast, ‘non-COAPT-like’ patients diverged from at least one of the listed criteria. For those patients lacking any data, the COAPT risk score could not be computed, and those patients were excluded from further sub-analyses.

Statistical analysis

Categorical variables are presented as numbers and frequencies (%), whilst continuous variables are given as means ± standard deviation. Chi-square and Fisher’s exact tests were used to evaluate the association between categorical variables and independent samples. Wilcoxon test was used for comparison of continuous variables. Survival was illustrated using the Kaplan–Meier method, and a Cox proportional hazard model was used to estimate hazard ratios (HRs) including 95% confidence intervals (CIs).

An extreme gradient boosting (XGB) algorithm (R package ‘xgboost’) was selected as the machine learning technique of choice for mortality prediction. For our modelling approach, we treated the prediction task as a binary classification problem (distinguishing between patients who were alive and those who were deceased at the 1-year follow-up after M-TEER). While our data set contains follow-up time information, these data were not directly incorporated into our prediction model (meaning that our model does not directly address right censoring or incorporate varying lengths of follow-up in its current form). Instead of random data set splitting, the algorithm was developed on data from the derivation cohort and validated on a never-before-seen data set from two independent institutions (validation cohort; giving an estimate of the algorithm’s performance in future patients). Moreover, patients in the derivation cohort with unknown 1-year survival status after M-TEER were not included in the development of the model. Initially, 28 variables covering clinical, echocardiographic, laboratory, and medication data were considered as potential input parameters (see Supplementary data online, Table S1). Individually missing values for the input parameters, used to train the model on the derivation cohort, were imputed using an established random forest algorithm.¹⁷ This approach enabled to retain as many patient records as possible while ensuring that the data set for training of the model was complete. However, in subsequent analyses such as the comparison of baseline characteristics, we strictly used available, non-imputed data to ensure the integrity and accuracy of these comparisons. Input parameter selection was refined by recursive feature elimination, meaning that the XGB algorithm was initially trained on all 28 variables as input parameters (see Supplementary data online, Figure S1). Each parameter was then ranked based on its global feature importance in predicting the target variable, i.e. 1-year all-cause mortality. Contrary to a pre-specified number of parameters, our approach involved iteratively discarding the least important feature, defined as having a global feature importance of zero. This process of iterative model re-fitting continued until only parameters with a SHapley Additive exPlanations (SHAP) value of at least 0.001 remained. Hyperparameters of the XGB algorithm were optimized by testing various settings in a five-fold cross-validation within the derivation cohort using a grid search approach. The metric of choice to evaluate the model performance was the area under the curve (AUC) from receiver operating characteristic (ROC) analysis. The SHAP values were calculated to quantify the contribution of each input variable to the model’s prediction (R package ‘SHAPforxgboost’). This approach provides a clear and interpretable understanding of how each parameter influences the model’s outcome.^16,18 Finally, to facilitate the clinical implementation of the XGB algorithm, which was named ‘EuroSMR risk score’, a free online calculator was created based on Shiny R (www.eurosmr.com). ROC analyses were conducted to compare the predictive performance of the EuroSMR risk score against previously established risk markers and dedicated scores for 1-year mortality after M-TEER. To statistically compare the AUC of our EuroSMR risk score against other scores and parameters, we applied the roc.test function, which compares two ROC curves based on the method proposed by DeLong et al.¹⁹

A P-value of ≤.05 was considered to indicate statistical significance. All statistical analyses were performed using R statistical software (R version 3.6.3; R Foundation for Statistical Computing, Vienna, Austria; see Supplementary data online, Table S2 for a complete list of employed R packages).

Results

Overall study cohort

A total of 4600 patients who underwent M-TEER for SMR were included in this analysis. The derivation cohort included 4172 patients, while the validation cohort comprised 428 patients (Figure 1; Structured Graphical Abstract). The patients’ demographic, clinical, laboratory, and echocardiographic characteristics are summarized in Tables 1 and 2. The median follow-up was 1.7 years (interquartile range: 1.0–2.7 years). The estimated 1- and 2-year survival rates for all patients were 82.1% (95% CI 80.9%–83.2%) and 70.1% (95% CI 68.6%–71.6%), respectively, without statistically relevant differences amongst derivation and validation cohorts (see Supplementary data online, Figure S2).

The 1-year survival status was known in 3449 (82.7%) patients from the derivation cohort. From the initial 28 candidate input parameters (see Supplementary data online, Table S1 and Figure S1), recursive feature elimination reduced the amount to 18 final input parameters, representing a comprehensive set of clinical, echocardiographic, laboratory, and medication data. The 1-year mortality prediction of the XGB algorithm is hereinafter referred to as EuroSMR risk score, and it ranged between 12.7 and 85.5 points in the derivation cohort (theoretical range 0–100 points). As shown by ROC analysis, the algorithm was trained until reaching an AUC of 0.768 (95% CI 0.749–0.786) in the derivation cohort (see Supplementary data online, Figure S3). The top five parameters with the strongest contribution to 1-year mortality prediction were (in order of predictive importance): NT-proBNP, NYHA class, haemoglobin, TAPSE, and age (Figure 2; Structured Graphical Abstract). Choosing a balance between sensitivity and specificity to predict 1-year mortality, a Youden index-based cut-off value for the EuroSMR risk score was identified at 52.5 points, yielding the best discrimination of surviving and non-surviving patients (see Supplementary data online, Figure S3). Non-survivors were detected with a sensitivity of 63.1% (95% CI 59.5%–66.6%) and a specificity of 75.6% (95% CI 73.9%–77.2%) (see Supplementary data online, Table S3). The respective Kaplan–Meier analysis revealed that patients who were classified as likely to not survive the first year after M-TEER had an increased HR for 1-year mortality of 4.3 (95% CI: 3.7–5.0, P < .001; Figure 3).

Validation of the risk score

In the validation cohort which comprised 428 patients, the calculated EuroSMR risk scores ranged from 11.0 to 84.7 points and the score performance yielded an AUC of 0.789 (95% CI 0.737–0.842) for 1-year mortality prediction. When the Youden index-based cut-off of 52.5 score points was applied, the risk score identified 322 (77.6%) and 96 (22.4%) patients with a lower and higher risk for 1-year mortality, respectively (see Supplementary data online, Figure S4). In the Kaplan–Meier curves, the 1-year survival rates were 89.1% (95% CI 85.8%–92.6%) and 60.3% (95% CI 51.2%–70.9%) in the low- and high-risk groups, respectively, with a HR of 4.5 (95% CI: 2.8–7.1; P < .001; Figure 4). The score achieved a diagnostic accuracy in 1-year mortality prediction of 77.8% (95% CI 73.5%–81.7%) with sensitivity, specificity, and positive and negative predictive values of 52.1% (95% CI 40.6%–63.5%), 83.4% (95% CI 79.4%–87.4%), 40.4% (95% CI 30.5%–50.3%), and 88.9% (95% CI 85.5%–92.4%), respectively (see Supplementary data online, Table S4). As shown by calibration plots, the performance of the EuroSMR risk score to predict 1-year mortality was particularly good for patients with low- to intermediate-risk profiles; however, our algorithm tended to underestimate mortality in those with high-risk profiles (see Supplementary data online, Figure S5).

The EuroSMR risk score calculation is shown for two representative patient examples in Supplementary data online, Figure S6; Patient #1 with a EuroSMR risk score of 47.2 points was still alive at 35 months after M-TEER. Patient #2 with a EuroSMR risk score of 83.7 points had died within the first month after M-TEER. The prediction of mortality for Patient #2 was predominantly influenced by a high level of NT-proBNP (35 000 pg/mL), severe dyspnoea as indicated by NYHA class IV, and an advanced age of 80.7 years.

Performance of the EuroSMR risk score in subgroups

The performance of the EuroSMR risk score was evaluated in several patient subgroups, including LVEF ≥ 50% vs. LVEF < 50%, ischaemic vs. non-ischaemic MR, NYHA class ≤III vs. IV, TR grade ≤2+ vs. ≥3+, and sPAP ≤ 35 mmHg vs. >35 mmHg. The results are summarized in Supplementary data online, Figure S7.

Comparison of the EuroSMR risk score with other scores

The 1-year mortality prediction from the EuroSMR risk score was compared with EuroSCORE II, the MitraScore, and the COAPT risk score in the validation cohort. The EuroSMR risk score performed significantly better than all three risk scores as displayed in the respective ROC curves in Figure 5 as well as in in time-dependent ROC analyses (see Supplementary data online, Figure S8). The EuroSMR risk score also significantly outperformed the COAPT score when 2-year mortality or the combined endpoint of 2-year mortality and heart failure hospitalization were selected as an endpoint (see Supplementary data online, Figure S9A). The performance of the COAPT risk score in the extended EuroSMR patient cohort is visualized in Supplementary data online, Figure S9B.

Assessment of mitral valve transcatheter edge-to-edge repair utility and futility

A total of 3209 patients from the entire extended EuroSMR cohort (derivation and validation cohort) with known 1-year follow-up status were ranked according to their EuroSMR risk score. Afterwards, 20 groups of approximately 160 patients were formed and the event rates of (i) mortality, (ii) mortality or NYHA class IV at 1-year follow-up, and (iii) mortality or NYHA class ≥III at 1-year follow-up were calculated and displayed for each group in Figure 6 (and Structured Graphical Abstract). The frequency distribution of EuroSMR risk scores is also provided. The median EuroSMR risk score of the extended cohort was 45.0 points, which translated into event rates of approximately 21.9%, 24.4%, and 49.5% for the endpoints of mortality, mortality or NYHA class IV, and mortality or NYHA class ≥III at 1-year follow-up, respectively. Above a EuroSMR risk score of 60 points, the likelihood for the occurrence of each endpoint increased excessively. In the 5% of patients with the highest EuroSMR risk score (mean score 74.4 ± 3.7 points), the respective event rates for mortality, mortality or NYHA class IV, and mortality or NYHA class ≥III at 1-year follow-up were 72.7%, 73.5%, and 83.2%, respectively.

Extreme-risk prediction

In addition, we derived a ‘EuroSMR extreme-risk cut-off’ for mortality prediction from the derivation cohort, which is associated with extremely high 1-year mortality, i.e. with a very low survival likelihood despite M-TEER. When the EuroSMR extreme-risk cut-off of 70.9 points (instead of the Youden index-based cut-off) was applied to the validation cohort, the EuroSMR risk score identified 11 (2.6%) patients with an extremely high risk for 1-year mortality (Figure 7). At 1 year, the predicted survival rate was only 36.4% (95% CI 16.6%–79.5%) and all patients had died at 1.5 years after the procedure. In this extreme-risk patient group, the EuroSMR risk score achieved an accuracy of 82.9% (95% CI 78.9%–86.4%) for 1-year mortality prediction (see Supplementary data online, Table S5). Importantly, the specificity reached 98.8% (95% CI 97.7%–100%), indicating a high reliability to identify patients who will survive beyond 1 year.

Impact of COAPT likeliness in 1-year mortality prediction

Based on seven COAPT eligibility criteria, the extended EuroSMR registry was divided into ‘COAPT-like’ and ‘non-COAPT-like’ patient groups. Those seven COAPT eligibility criteria were fully available for evaluation in 1822 patients. All seven COAPT eligibility criteria were met by 37.5% of COAPT-like patients. Notably, non-COAPT-like patients showed higher EuroSMR risk scores than COAPT-like patients (47.0 ± 14.4 vs. 43.2 ± 13.7; P < .001; Figure 8). Furthermore, non-COAPT-like patients displayed a significantly higher risk for 1-year mortality, when compared with COAPT-like patients (HR for 1-year mortality: 1.4, 95% CI 1.1–1.7; P = .005; Figure 8).

In an additional secondary analysis, we compared the predictive value of the SMR proportionality and the EuroSMR risk score for 1-year mortality prediction. The ratio of EROA_mitral to LVEDV, known as the SMR proportionality quotient, was significantly outperformed by the EuroSMR risk score in the validation cohort [AUC 0.536 (95% CI 0.440–0.633) vs. 0.789 (95% CI 0.737–0.842); P < .001; Supplementary data online, Figure S10].

Discussion

Based on the extended EuroSMR cohort, which is to our knowledge the largest international retrospective registry on outcomes of M-TEER in patients with SMR, an artificial intelligence (AI)-derived, comprehensive risk score for mortality and futility prediction was developed and validated for SMR patients considered for M-TEER. The main findings from this study can be summarized as follows: firstly, the EuroSMR risk score, which was validated in a separate patient cohort, allows for a significantly improved 1-year mortality prediction using clinical, echocardiographic, laboratory, and medication data over established risk scores. Secondly, the score identifies SMR patients at extremely high risk for mortality and allows for estimating the likelihood of M-TEER futility, which might help for identifying patients who benefit most from M-TEER. Thirdly, the EuroSMR score enables for an improved characterization of SMR patient groups undergoing M-TEER, as evidenced by the differences between COAPT-like vs. non-COAPT-like patient groups.

Our ability to predict which SMR patients will derive a meaningful benefit from a M-TEER procedure is often limited by the patients’ age, frailty, and large burden of comorbidities. The management of SMR is challenging and should be discussed in an interdisciplinary team on an individual basis.²⁰ However, careful personalized risk stratification and appropriate patient selection for M-TEER are warranted due to the relatively high procedural costs as well as the considerable residual morbidity and mortality. The currently available scores for risk prediction in M-TEER present with a number of limitations: the often applied EuroSCORE II has been developed for in-hospital mortality prediction of cardiac surgical procedures but has not been designed to predict 1-year survival for catheter-based mitral interventions; the MitraScore predicts the 1-year mortality rates after M-TEER but does not differentiate between patients with primary and secondary MR, representing two different disease entities with different outcomes. The recently developed COAPT score estimates the 2-year risk for mortality and the need of heart failure hospitalizations. Yet, the derivation cohort included 614 highly selected SMR patients, amongst which only 302 patients were treated by M-TEER, which may limit the diagnostic accuracy of the COAPT score in daily practice. Furthermore, a validation of the COAPT score is lacking thus far.

With the availability of new AI applications,²¹ we sought to develop a novel AI-derived risk score for 1-year mortality prediction from a large cohort of SMR patients undergoing M-TEER. The mortality prediction from our EuroSMR risk score, which is based on 18 clinical, echocardiographic, laboratory variables, and medication data, ranged from 2.3% for patients in the lowest 5th percentile to 74.3% for patients in the highest 5th percentile of the extended EuroSMR cohort. The score was validated in a separate validation cohort and significantly outperforms the above-mentioned other risk scores. Of the 18 variables, most readily available in patients evaluated for M-TEER, the first eight variables determine >80% of the EuroSMR score value, while the next 10 variables refine the individual risk. Interestingly, three laboratory markers were amongst the top input parameters with the highest global feature importance: NT-proBNP, haemoglobin, and estimated glomerular filtration rate (Structured Graphical Abstract). While anaemia might reflect iron deficiency due to low iron intake, mal-absorption, or gastrointestinal blood loss, the reduced estimated glomerular filtration rate might indicate the cardio-renal syndrome with forward failure and renal mal-perfusion. The prognostic significance of those laboratory markers has also been confirmed by a machine learning-based survival tree analysis in conservatively treated patients²² further strengthening the reliability of our model.

Importantly, the EuroSMR risk calculator (www.eurosmr.com) also has the ability to handle missing variables; however, the accuracy and reliability of the risk assessment are naturally enhanced with more complete information. The availability of an online EuroSMR risk calculator will facilitate the widespread use in daily clinical practice for an individual risk assessment. The EuroSMR risk score will support heart team members involved in shared decision-making regarding M-TEER, because the score enables the identification of patients, who have a low or high risk for 1-year mortality.

The EuroSMR risk score outperformed other scores in predicting 1-year mortality after M-TEER. Furthermore, ‘high-risk patients’ identified by the EuroSMR risk score exhibited a significantly higher mortality rate compared with those classified as ‘high risk’ by the COAPT score or MitraScore criteria (see Supplementary data online, Figure S11). In this regard, it is important to emphasize that the EuroSMR risk score tends to slightly underestimate the mortality risk amongst ‘high-risk patients’ (see calibration plots in Supplementary data online, Figure S5). This slight underestimation appears to be acceptable, considering that an overestimation may inadvertently lead to M-TEER being withheld from patients perceived as being at too high risk.

Besides mortality risk, assessment of risk/benefit ratios and estimation of procedural futility in the individual patient, which is usually defined as a low probability for a meaningful survival or symptomatic improvement, have been challenging due to the lack of appropriate risk prediction models. In this regard, the easy-to-use EuroSMR risk score may help clinicians and heart team members with the ability to perform an adequate risk stratification before the M-TEER procedure in two important aspects: firstly, the EuroSMR risk score allows for the identification of patients at an extremely high risk for 1-year mortality, a threshold often recommended by current guidelines for the definition of procedural futility, e.g. in the presence of malignancies. A heart team discussion on deferring a M-TEER procedure in such extreme-risk patients appears to be indicated, because a modification of the individual risk parameters in SMR patients who are considered to be already on maximized GDMT, is unlikely to result in lower risk estimates. Secondly, the EuroSMR risk score provides heart team members with a probability for predicting patients’ survival with a clinical improvement, represented by NYHA class I or II at follow-up. This probability drops from 75% to <20% over the wide range of EuroSMR risk scores in the studied SMR population. By providing important information beyond mortality prediction, the EuroSMR risk score will aid in discussing the patient’s expectations from a M-TEER procedure, ultimately leading to a more informed and personalized approach to care. Finally, the EuroSMR score allows characterization of patient cohorts with SMR. The overall risk for mortality was significantly lower in COAPT-like patients, when compared with non-COAPT-like patients, which might indicate that differences in the risk profiles of the enrolled patient cohorts might have contributed to the divergent results of the above-mentioned two randomized controlled M-TEER trials: COAPT and MITRA-FR. Similarly, the EuroSMR risk score might be used to characterize the risk profile of patient cohorts in future clinical trials addressing SMR in heart failure, including the ongoing RESHAPE-HF2 trial (A RandomizEd Study of tHe MitrACliP DEvice in Heart Failure Patients With Clinically Significant Functional Mitral Regurgitation; NCT02444338), to describe the studied cohort and estimate its risk.

Study limitations

The EuroSMR score was developed from a large patient cohort, for which the 1-year survival information was available in 3449 out of 4172 patients (82.7%). The restriction to patients from the derivation cohort with known 1-year survival status was necessary as the algorithm was trained to solve a binary classification task (i.e. prediction of survival or death at 1 year after M-TEER). The partially missing survival information in the derivation cohort might have, therefore, impacted the development of the EuroSMR risk score by introducing some unintentional selection bias; however, as shown by Kaplan–Meier survival analyses, the comparable survival rates for patients with predicted 1-year mortality in both the derivation and validation cohorts support the validity and generalizability of the risk model within the extended EuroSMR registry covering a broad, real-world spectrum of European patients with SMR (Figures 3 and 4). While the EuroSMR score has been validated using a ‘never-before-seen’ data set from two independent European institutions, we must acknowledge the possibility that this validation cohort may contain the same selection biases as the patient cohort used for model development. Both the derivation and validation cohorts consisted of patients predominantly from the same ethnicity and possibly subjected to similar treatment standards. Therefore, while our validation represents a notable step beyond internal validation, further prospective validation in diverse patient populations, particularly from different ethnic backgrounds and healthcare systems like those in North America or Japan, is crucial. Such broader validation would not only confirm the clinical utility and robustness of the EuroSMR score but also mitigate the risk of systemic biases. It would thereby enhance the generalizability of our findings across various global healthcare settings and diverse patient demographics.

The EuroSMR registry is a large, real-word M-TEER registry, which is lacking a control group of patients. Thus, a post hoc analysis of patient groups from the randomized clinical trials, COAPT, MITRA-FR, and RESHAPE-HF2, would provide important information on the value of M-TEER in patients with different EuroSMR risk profiles, which would support the clinical importance of a pre-procedural risk stratification by the EuroSMR risk score. The M-TEER in EuroSMR was performed with predominantly first- and second-generation MitraClip systems, which were also used in COAPT and MITRA-FR. It remains unclear, if iterations and modifications of the EuroSMR score might be needed to reflect future advances in the field of M-TEER.

Conclusions

In conclusion, the AI-derived EuroSMR score is a novel comprehensive risk calculator for the prediction of mortality and clinical improvement in heart failure patients with SMR considered for M-TEER. The score also facilitates the identification of patients at extreme risk for mortality and/or lack of symptomatic improvement. Accordingly, this algorithm will support clinicians and patients in daily practice in the shared decision-making process on the utility or futility of a planned M-TEER procedure.

Supplementary data

Supplementary data are available at European Heart Journal online.

Declarations

Disclosure of Interest

J.H. reports research grant support and speaker honoraria from Edwards Lifesciences. M.L. received funding from the Technical University of Munich (Clinician Scientist Grant) and from the Else Kröner-Fresenius Foundation (Clinician Scientist Grant). L.S. received speaker honoraria from Edwards Lifesciences. R.E.-L. serves as a consultant for Abbott Vascular and Edwards Lifesciences and has received speaker fees from Boston Scientific and Venus Medtech. S.R.-R. received speaker fees from AstraZeneca, Bayer, Boehringer, Daiichi, Pfizer, Amgen, and Abbott. N.K. received consulting fees from Medtronic, Edwards Lifesciences, and Abbott Vascular. V.R. serves on the advisory board of Edwards Lifesciences and Abbott Vascular, and he received research grants from Edwards Lifesciences and Abbott Vascular. The remaining authors have nothing to disclose.

Data Availability

In general, we highly appreciate further collaborations to extend the results of this study and optimize the treatment of SMR. Due to multiple registries and centres being included in this study, each site will need to agree with potential data sharing for further projects. Therefore, we request contacting the author with appropriate requests.

Funding

All authors declare no funding for this contribution.

Ethical Approval

The study received proper ethical oversight.

Pre-registered Clinical Trial Number

DRKS00017428.

Appendix

List of contributing physicians (for PubMed listing)

GERMANY:

Medizinische Klinik und Poliklinik I, Klinikum der Universität München, Munich, Germany:

• Thomas Stocker, MD; [email protected]

• Mathias Orban, MD; [email protected]

• Daniel Braun, MD; [email protected]

• Michael Näbauer, MD; [email protected]

• Steffen Massberg, MD; [email protected]

Zentrum für Kardiologie, Johannes-Gutenberg-Universität, Mainz, Germany:

• Aniela Popescu, MD; [email protected]

• Tobias Ruf, MD; [email protected]

• Ralph Stephan von Bardeleben, MD; [email protected]

Department III of Internal Medicine, Heart Center, University of Cologne, Cologne, Germany:

• Christos Iliadis, MD; [email protected]

• Roman Pfister, MD; [email protected]

• Stephan Baldus, MD; [email protected]

Department of Cardiology, Heart Center Leipzig, University of Leipzig, Leipzig, Germany:

• Christian Besler, MD; [email protected]

• Tobias Kister, MD; [email protected]

• Karl Kresoja, MD; [email protected]

• Philipp Lurz, MD; [email protected]

• Holger Thiele, MD; [email protected]

Department of Cardiology, University Heart and Vascular Center Hamburg, University Medical Center Hamburg-Eppendorf, Hamburg, Germany:

• Benedikt Koell, MD; [email protected]

• Niklas Schofer, MD; [email protected]

• Daniel Kalbacher, MD; [email protected]

Herzzentrum Brandenburg, Medizinische Hochschule Brandenburg Theodor Fontane, Bernau, Germany:

• Michael Neuss, MD; [email protected]

• Christian Butter, MD; [email protected]

Medizinische Klinik, Klinikum rechts der Isar, Munich, Germany:

• Karl-Ludwig Laugwitz, MD; [email protected]

Department of Cardiology, Deutsches Herzzentrum München, Munich, Germany:

• Teresa Trenkwalder, MD; [email protected]

• Eroion Xhepa, MD; [email protected]

• Michael Joner, MD; [email protected]

Department of Cardiology, Herz- und Diabeteszentrum, Bad Oeynhausen, Germany:

• Hazem Omran, MD; [email protected]

• Vera Fortmeier, MD; [email protected]

• Muhammed Gerçek, MD; [email protected]

Department of Cardiology, Helios Klinikum Siegburg, Siegburg, and Department of Cardiology, Faculty of Health, School of Medicine, Witten, Witten/Herdecke University, Germany:

• Harald Beucher MD; [email protected]

Department of Cardiology, Elisabeth Krankenhaus, Essen, Germany

• Thomas Schmitz, MD; [email protected]

Department of Cardiology, Heart Centre Krefeld, University Witten/Herdecke, Witten, Germany:

• Alexander Bufe, MD; [email protected]

Department of Cardiology and Angiology II, University Heart Center Freiburg, Bad Krozingen, Germany:

• Jürgen Rothe, MD; [email protected]

Department of Cardiology, University Hospital Helios Wuppertal, Wuppertal, Germany:

• Melchior Seyfarth, MD; [email protected]

Department of Cardiology, University Heart Center Lübeck, Lübeck, Germany:

• Tobias Schmidt, MD; [email protected]

• Christian Frerker, MD; [email protected]

Department of Cardiology, Krankenhaus Porz am Rhein, Cologne, Germany and Department of Cardiology, Faculty of Health, School of Medicine, Witten/Herdecke University, Witten, Germany:

• Dennis Rottländer, MD; [email protected]

Department of Cardiology, University Hospital Düsseldorf, Düsseldorf, Germany

• PD Dr. med. Patrick Horn; [email protected],

• Dr. med. Maximilian Spieker; [email protected]

• Dr. med. Elric Zweck; [email protected]

SWITZERLAND:

Universitätsklinik für Kardiologie, Inselspital Bern, Bern, Switzerland:

• Mohammad Kassar, MD; [email protected]

• Fabien Praz, MD; [email protected]

• Stephan Windecker, MD; [email protected]

FRANCE:

Department of Cardiology, European Hospital Georges Pompidou and Paris Cardiovascular Research Center (INSERM U970), Paris, France:

• Tania Puscas, MD: [email protected]

ITALY:

Cardiac Catheterization Laboratory and Cardiology, ASST Spedali Civili and University of Brescia, Brescia, Italy:

• Marianna Adamo, MD; [email protected]

• Laura Lupi, MD; [email protected]

• Marco Metra, MD; [email protected]

Valve Center and Cardiac Surgery Unit, Poliambulanza Foundation Hospital, Brescia, Italy:

• Emmanuel Villa, MD; [email protected]

Invasive Cardiology Unit, Pineta Grande Hospital, Castelvolturno, Italy:

• Giuseppe Biondi Zoccai, MD, PhD; [email protected]

Division of Cardiology, Centro Alte Specialità e Trapianti (CAST), Azienda Ospedaliero-Universitaria Policlinico-Vittorio Emanuele, University of Catania, Catania, Italy:

• Corrado Tamburino, MD, PhD; [email protected]

• Carmelo Grasso, MD; [email protected]

Interventional Cardio-Angiology Unit, Villa Maria Cecilia Hospital, Cotignola, Italy:

• Fausto Catriota, MD; [email protected]

Department of Cardiology, IRCCS Policlinico San Donato, San Donato Milanese, Milan, Italy:

• Luca Testa, MD, PhD; [email protected]

• Maurizio Tusa, MD; [email protected]

Clinical Cardiology Unit, Faculty of Medicine, IRCCS San Raffaele Scientific Institute, Milan, Italy:

• Cosmo Godino, MD, PhD; [email protected]

• Michele Galasso, MD, PhD; [email protected]

• Matteo Montorfano; MD; [email protected]

• Eustachio Agricola, MD, PhD; [email protected]

Cardiac Surgery Department, San Raffaele University Hospital, Milan, Italy:

• Paolo Denti, MD, [email protected]

Centro Cardiologico Monzino, IRCCS, Milan, Italy:

• Federico De Marco, MD; [email protected]

Department of Cardiac, Thoracic and Vascular Science, Interventional Cardiology Unit, University of Padua, Padua, Italy:

• Giuseppe Tarantini, MD, PhD; [email protected]

• Giulia Masiero, MD; [email protected]

Fondazione IRCCS Policlinico San Matteo, Pavia, Italy:

• Gabriele Crimi, MD; [email protected]

• Andrea Raffaele Munafò, MD; [email protected]

Cardiac Catheterization Laboratory, Cardiothoracic and Vascular Department, University of Pisa, Pisa, Italy:

• Christina Giannini, MD; [email protected]

• Anna Petronio, MD; [email protected]

Division of Cardiology, Department of Medical Science, University of Turin, Città della Salute e Della Scienza, Turin, Italy:

• Stefano Pidello, MD; [email protected]

• Paolo Boretto, MD; [email protected]

• Antonio Montefusco, MD, PhD; [email protected]

• Simone Frea, MD, PhD; [email protected]

• Filippo Angelini, MD, PhD; [email protected]

• Pier Paolo Bocchino, MD, PhD; [email protected]

Division of Interventional Cardiology, Azienda Ospedaliera S. Camillo Forlanini, Rome, Italy:

• Francesco De Felice, MD; [email protected]

University Hospital San Giovanni di Dio e Ruggi d'Aragona, Salerno, Italy:

• Rodolfo Citro, MD; [email protected]

SPAIN

Hospital Álvaro Cunqueiro, Vigo, Spain:

• Berenice Caneiro-Queija, MD; [email protected]

Hospital Clinic, Barcelona, Spain:

• Xavier Freixa, MD, PhD; [email protected]

• Ander Regueiro, MD, PhD; [email protected]

• Laura Sanchís, MD, PhD; [email protected]

• Manel Sabaté, MD, PhD; [email protected]

Interventional Cardiology Unit, Hospital Sant Pau i Santa Creu, Barcelona, Spain:

• Dabit Arzamendi, MD, PhD; [email protected]

• Lluís Asmarats, MD, PhD; [email protected]

• Estefanía Fernández Peregrina, MD, PhD; [email protected]

Complejo Asistencial Universitario de León, León, Spain:

• Tomas Benito-González, MD; [email protected]

• Felipe Fernández-Vázquez, MD, PhD; [email protected]

Interventional Cardiology Unit, Hospital Universitario Central de Asturias, Oviedo, Spain:

• Isaac Pascual, MD, PhD; [email protected]

• Pablo Avanzas, MD, PhD; [email protected]

Cardiovascular Institute, Hospital Clinico San Carlos, IdISSC, Madrid, Spain:

• Luis Nombela-Franco, MD, PhD; [email protected]

• Gabriela Tirado-Conte, MD; [email protected]

• Eduardo Pozo, MD, PhD; [email protected]

Cardiology Department, Hospital Universitario Puerta de Hierro, Majadahonda, Madrid:

• Antonio Portolés-Hernández, MD, PhD; [email protected]

• Vanessa Moñivas Palomero, MD, PhD; [email protected]

PORTUGAL:

Centro Hospitalar Vila Nova de Gaia, Espinho, Portugal:

• Francisco Sampaio, MD; [email protected]

• Bruno Melica, MD; [email protected]

CANADA:

Cardiology Department, Quebec Heart and Lung Institute, Laval University, Quebec City, Quebec, Canada:

• Josep Rodes-Cabau, MD, PhD; [email protected]

• Jean-Michel Paradis, MD, PhD; [email protected]

• Alberto Alperi, MD, PhD; [email protected]

ISRAEL:

Heart Institute, Hadassah-Hebrew University Medical Center, Jerusalem, Israel:

• Mony Shuvy, MD, PhD; [email protected]

• Dan Haberman MD PhD; [email protected]

TWITTER:

@maorban, @DanielKalbacher, @chriliadis, @KP_Kresoja, @TobiasKister, @PhilippLurz, @Hazem_Omran_MD, @benediktkoell, @T_Trenkwalder, @MGercek40, @DRottlander, @mkassar90, @FabienPraz, @MariannaAdamo1, @DrLucaTesta, @GodinoCosmo, @MMontorfanoOSR, @G_Tarantini01, @giuliamasiero3, @andrea_muna, @Fil_Angelini, @BocchinoPier, @CitroRodolfo, @b_caneiro, @anderregueiro, @StructuralBCN, @EstefaniaFdezP3, @AsmaratsL, @TdBenito, @Isaacpascual79, @pabloavanzas, @ConteTirado, @epozoosinalde, @VMonivas, @a_portoles, @Falmeidasampaio, @lsanchisruiz, @brmelica, @mshuvy, @haberdan, @EusAgricolaMD, @gbiondizoccai.

DISCLOSURES:

M. Näbauer received lecture and proktoring fees from Edwards Lifesciences. C. Iliadis has received travel support by Abbott and consultant honoraria by Abbott and Edwards Lifesciences; R. Pfister reports consultancy and speaker fee by Edwards Lifesciences and speaker fees by Abbot Vascular; S. Baldus received a research grant from Abbott Vascular; N. Schofer reports proctor fees, speaking honoraria and travel support from Edwards Lifesciences, speaking honoraria from HighLife SAS, travel support from Abbott Vascular, and speaking honoraria and travel support from Boston Scientific; D. Kalbacher has received personal fees from Edwards Lifesciences, Abbott Medical and Pi-Cardia Ltd.; F. Praz received travel expenses from Edwards Lifesciences, Abbott Vascular, Medira, Polares Medical, and Siemens Healthineers; Stephan Windecker reports research, travel or educational grants to the institution from Abbott, Abiomed, Amgen, Astra Zeneca, Bayer, Biotronik, Boehringer Ingelheim, Boston Scientific, Bristol Myers Squibb, Cardinal Health, CardioValve, Corflow Therapeutics, CSL Behring, Daiichi Sankyo, Edwards Lifesciences, Guerbet, InfraRedx, Janssen-Cilag, Johnson & Johnson, Medicure, Medtronic, Merck Sharp & Dohm, Miracor Medical, Novartis, Novo Nordisk, Organon, OrPha Suisse, Pfizer, Polares, Regeneron, Sanofi-Aventis, Servier, Sinomed, Terumo, Vifor, V-Wave; Stephan Windecker serves as advisory board member and/or member of the steering/executive group of trials funded by Abbott, Abiomed, Amgen, Astra Zeneca, Bayer, Boston Scientific, Biotronik, Bristol Myers Squibb, Edwards Lifesciences, Janssen, MedAlliance, Medtronic, Novartis, Polares, Recardio, Sinomed, Terumo, V-Wave and Xeltis with payments to the institution but no personal payments. He is also member of the steering/executive committee group of several investigator-initiated trials that receive funding by industry without impact on his personal remuneration. M. Adamo received speaker fees from Abbott Vascular and Medtronic; G. Biondi-Zoccai has consulted for Amarin, Balmed, Cardionovum, Crannmedical, Endocore Lab, Eukon, Innovheart, Guidotti, Meditrial; M. Montorfanohas the following financial interests/arrangements: consultant fees from Abbott, Boston Scientific, Edwards, Kardia and Medtronic; P. Denti reports receiving speaker honoraria from Abbott Vascular and Edwards Lifesciences; G. Tarantini received speaker fees for Abbott Vascular and Edwards Lifesciences; C. Grasso, MD serves as proctor for Abbott Vascular and Boston Scientific; L. Sanchis serves as proctor for and received speaker honararia from Abbott Vascular; G. Tirado-Conte holds a research-training contract ‘Rio Hortega’ (CM21/00091) from the Spanish Ministry of Science and Innovation (Instituto de Salud Carlos III); V. Moñivas Palomero serves as clinical proctor for Abbott Vascular; M. Shuvy serves as clinical proctor for Abbott Vascular. The remaining contributing colleagues have nothing to disclose.

References

Author notes