Artificial intelligence–enabled assessment of right ventricular to pulmonary artery coupling in patients undergoing transcatheter tricuspid valve intervention

Abstract

Aims

Right ventricular to pulmonary artery (RV-PA) coupling has been established as a prognostic marker in patients with severe tricuspid regurgitation (TR) undergoing transcatheter tricuspid valve interventions (TTVI). RV-PA coupling assesses right ventricular systolic function related to pulmonary artery pressure levels, which are ideally measured by right heart catheterization. This study aimed to improve the RV-PA coupling concept by relating tricuspid annular plane systolic excursion (TAPSE) to mean pulmonary artery pressure (mPAP) levels. Moreover, instead of right heart catheterization, this study sought to employ an extreme gradient boosting (XGB) algorithm to predict mPAP levels based on standard echocardiographic parameters.

Methods and results

This multicentre study included 737 patients undergoing TTVI for severe TR; among them, 55 patients from one institution served for external validation. Complete echocardiography and right heart catheterization data were available from all patients. The XGB algorithm trained on 10 echocardiographic parameters could reliably predict mPAP levels as evaluated on right heart catheterization data from external validation (Pearson correlation coefficient R: 0.68; P value: 1.3 × 10⁻⁸). Moreover, predicted mPAP (mPAP_predicted) levels were superior to echocardiographic systolic pulmonary artery pressure (sPAP_{echocardiography}) levels in predicting 2-year mortality after TTVI [area under the curve (AUC): 0.607 vs. 0.520; P value: 1.9 × 10⁻⁶]. Furthermore, TAPSE/mPAP_predicted was superior to TAPSE/sPAP_{echocardiography} in predicting 2-year mortality after TTVI (AUC: 0.633 vs. 0.586; P value: 0.008). Finally, patients with preserved RV-PA coupling (defined as TAPSE/mPAP_predicted > 0.617 mm/mmHg) showed significantly higher 2-year survival rates after TTVI than patients with reduced RV-PA coupling (81.5% vs. 58.8%, P < 0.001). Moreover, independent association between TAPSE/mPAP_predicted levels and 2-year mortality after TTVI was confirmed by multivariate regression analysis (P value: 6.3 × 10⁻⁴).

Conclusion

Artificial intelligence–enabled RV-PA coupling assessment can refine risk stratification prior to TTVI without necessitating invasive right heart catheterization. A comparison with conservatively treated patients is mandatory to quantify the benefit of TTVI in accordance with RV-PA coupling.

Graphical Abstract

Artificial intelligence–enabled assessment of right ventricular to pulmonary artery coupling promises to improve prognostication in patients with severe tricuspid regurgitation undergoing transcatheter tricuspid valve intervention. Invasive assessment of right ventricular to pulmonary artery coupling is superior to non-invasive, echocardiographic assessment, because echocardiography tends to underestimate pulmonary artery pressure levels in patients with severe tricuspid regurgitation. However, invasive assessment by right heart catheterization is no feasible option in an outpatient setting, where the first line of patient care takes place. The application of artificial intelligence for mean pulmonary artery pressure level prediction based on echocardiographic parameters holds the promise to refine non-invasive right ventricular to pulmonary artery coupling assessment.

Open in new tabDownload slide

Introduction

The severity of untreated tricuspid regurgitation (TR) is independently associated with elevated mortality in a broad range of patient populations, including those presenting with heart failure with reduced ejection fraction, pulmonary hypertension, and atrial fibrillation.¹ Since isolated tricuspid valve surgery is reputed to be a high-risk surgery with a 10% in-hospital mortality rate,² transcatheter tricuspid valve interventions (TTVI) including leaflet approximation, direct annuloplasty, and heterotopic caval valve implantation have been established as a safe treatment alternative.³ Two propensity score–matched analyses demonstrate that TTVI is superior to conservative (standard medical) treatment with regard to the reduction of 1-year mortality and the rate of re-hospitalization for heart failure.^4,5 Moreover, first data from the TRILUMINATE Pivotal trial as the first randomized controlled analysis show that TTVI improves the quality of life in an otherwise highly symptomatic patient population; however, no differences in rates of all-cause death or hospitalization for heart failure were evident at 1 year after randomization.⁶

Considering that patients with severe TR present with vast heterogeneity owing to differences in comorbidities, pathophysiology, and progression of disease, it should be acknowledged that not all patients may equally benefit from TTVI. After TTVI, the right ventricle is acutely forced to eject blood into the high-pressure pulmonary circulation, as the regurgitant blood flow through the tricuspid valve to the low-pressure right atrium is abruptly reduced. Therefore, it is of paramount importance that the right ventricle has a sufficiently preserved contractile function to compensate for the sudden increase in afterload burden. Right ventricular to pulmonary artery (RV-PA) coupling has therefore emerged as a prognostically relevant marker for survival after TTVI capturing the contractility of the right ventricle related to pulmonary artery pressure levels.^7–9

RV-PA coupling was initially described as the ratio expressed as tricuspid annular plane systolic excursion (TAPSE) related to systolic pulmonary artery pressure (sPAP) levels as assessed by transthoracic echocardiography. However, echocardiography tends to underestimate sPAP levels in patients with severe TR, because a huge tricuspid valve regurgitant orifice area results in rapid pressure equalization between right ventricle and right atrium.^10,11 Current guidelines therefore recommend right heart catheterization to measure pulmonary artery pressure levels and to diagnose pulmonary hypertension.¹²

To circumvent the necessity to perform invasive right heart catheterization, this study sought to employ a machine learning algorithm to predict pulmonary artery pressure levels from echocardiographic input parameters in patients with severe TR. Using those predicted pulmonary artery pressure levels, this study further aimed at refining the RV-PA coupling concept, ultimately leading to better prognostication for survival following TTVI.

Methods

Study population

This analysis is a post hoc, multicentric examination of prospectively and systematically collected data from five high-volume centres in Germany and Switzerland. It involves the systematic assessment of data from 918 patients who underwent TTVI for severe TR between 2016 and 2022. The primary inclusion criterion was the presence of severe TR,¹³ accompanied by a high symptomatic burden despite optimal medical treatment. These patients were also deemed inoperable due to the prohibitive perioperative risks, as determined by the local heart team. The study was conducted in accordance with the Declaration of Helsinki and received approval from the local ethics committee of each participating centre. All involved patients provided written informed consent. Specifically, patients with incomplete echocardiography or right heart catheterization data were excluded from this analysis. For the training of the algorithm, we compiled patient data from four independent German institutions, namely the Heart and Diabetes Center North Rhine-Westphalia in Bad Oeynhausen, the Heart Center at the University of Cologne Hospital, the Ludwig Maximilians University Hospital of Munich, and the Heart Center at the University of Leipzig. This compilation is hereafter referred to as the derivation cohort. In addition, an external validation cohort consisting of similarly treated patients was obtained from an independent institution, the Department of Cardiology at Bern University Hospital, Switzerland. It is essential to note that at the time of echocardiography or right heart catheterization, all patients were in a fully compensated state.

Echocardiographic analysis

All echocardiographic studies were performed by experienced institutional cardiologists during clinical routine with commercially available equipment (Philips Medical and General Electric systems). Left ventricular ejection fraction (LVEF) was measured using the biplane method, which involves averaging the volumes derived from both the apical four-chamber and two-chamber views. TAPSE was assessed through an apical four-chamber view utilizing the M-mode ultrasound to measure the movement of the tricuspid annulus in the longitudinal direction. Pulmonary hypertension assessment was a routine part of the preprocedural transthoracic echocardiographic examination. The echocardiographic sPAP (sPAP_{echocardiography}) was derived by summing the RV-RA gradient (estimated from the continuous wave Doppler profile of the TR jet) with the right atrial pressure. The latter was inferred from the diameter and collapsibility of the inferior vena cava, as outlined in contemporary guidelines.^14,15 The vena contracta width of TR was determined in a right ventricle–focused apical four-chamber view, targeting the most constricted segment of the regurgitant flow proximal to or at the regurgitant orifice. Based on 2D echocardiography, tricuspid valve effective regurgitant orifice area (TV EROA) and TR volume calculations employed the flow convergence technique, also known as the proximal isovelocity surface area method.¹⁶

Invasive pulmonary hypertension assessment

Right heart catheterization represents the gold standard to assess pulmonary artery pressure levels.¹⁷ A 7-French Swan-Ganz catheter was routinely used for preprocedural right heart catheterization via femoral access. Systolic and diastolic pulmonary artery pressure (dPAP) levels were directly recorded. Mean pulmonary artery pressure (mPAP) levels were calculated as mPAP = dPAP + 1/3 × (sPAP − dPAP). Mean postcapillary wedge pressure (mPCWP) was assessed over the entire cardiac cycle. Cardiac output was determined using the indirect Fick method. Pulmonary vascular resistance (PVR) was defined as PVR = (mPAP − mPCWP)/cardiac output. Invasively measured mPAP levels are hereinafter referred to as mPAP_invasive.

Artificial intelligence–enabled mPAP prediction

For the prediction of mPAP levels using standard echocardiographic parameters, we employed an extreme gradient boosting (XGB) algorithm, which is suited for regression tasks. In the derivation cohort data, which were used for algorithm training, any missing values among the input parameters were imputed using a well-established random forest algorithm.¹⁸ However, these imputed values were subsequently not utilized further.

Our selection of input parameters was fine-tuned through recursive feature elimination, as previously detailed in a proof-of-principle study.¹¹ The echocardiographic parameters that were selected as input variables included LVEF, left ventricular end-systolic diameter, left atrial area, sPAP_{echocardiography}, basal right ventricular diameter, TAPSE, TR vena contracta width, right atrial area, and inferior vena cava diameter. Notably, we expanded our list of input parameters from 9 to 10 by incorporating the TV EROA. This addition was made to account for potential inaccuracies in the measurement of the TR vena contracta width.

Hyperparameters of the XGB algorithm, such as maximum depth and learning rate (expressed as eta), underwent optimization. We implemented a grid search approach within a five-fold cross-validation of the derivation cohort to determine the optimal settings. To evaluate model performance in this regression task, the chosen metric was the root mean square error. To mitigate overfitting risks, we incorporated a callback function designed to halt the training process if the test set loss failed to show improvement over five consecutive epochs.

For the external validation cohort data, we did not perform any data imputation. We utilized SHAP (SHapley Additive exPlanations) values, a contemporary metric rooted in cooperative game theory,¹⁹ to ascertain the contribution of each input variable to the model’s prediction.²⁰ Throughout this work, values predicted through this machine learning methodology are referred to as mPAP_predicted.

Definitions of RV-PA coupling

RV-PA coupling was evaluated by relating TAPSE to either sPAP_{echocardiography} or mPAP_predicted levels. RV-PA uncoupling was defined if TAPSE/sPAP_{echocardiography} levels ranged below 0.317 mm/mmHg.⁷ Furthermore, maximally selected log-rank statistics were employed to calculate the best cut-off value of TAPSE/mPAP_predicted levels to predict 2-year mortality after TTVI.

Procedural success definition

Procedural success was defined as a device successfully implanted and delivery system retrieved, with TR reduction by at least one grade²¹ and/or a residual TR grade ≤ II/V°⁵ as evaluated on transthoracic echocardiography before discharge (that is 2–5 days after the procedure).

Clinical endpoint definition

As a population of elderly patients was studied, postprocedural 2-year all-cause mortality was defined as a clinically meaningful primary outcome measure. Survival data were regularly obtained from the German Civil Registry or from general practitioners, hospitals, and practice cardiologists for patients from foreign countries.

Statistical analysis

Data are presented as numbers and frequencies (%) or as means ± standard deviation [and 95% confidence interval (CI) when appropriate]. χ² or Fisher’s exact test was 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 the log-rank test was applied to compare cumulative survival rates. Moreover, a Cox proportional hazards model was used to estimate hazard ratios (HR). Maximally selected log-rank statistics were employed to create a simple model for survival after TTVI in accordance with RV-PA coupling.

To identify further factors related to all-cause mortality, Cox regression analyses were performed. Variables with a P ≤ 0.05 in univariate testing were subsequently included in the multivariate logistic regression model. The assumption of proportional hazards in the multivariate Cox regression model was verified using tests based on Schoenfeld residuals.

Receiver operating characteristic (ROC) curves and the corresponding area under the curve (AUC) were calculated to assess the performance of different RV-PA coupling indices to predict 2-year mortality after TTVI.

A P ≤ 0.05 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 S1 for a complete list of employed R packages).

Moreover, the work presented in this study is in accordance with the Proposed Recommendations for Cardiovascular Imaging-Related Machine Learning Evaluation (PRIME), which were specifically developed to enhance scientific rigour in medical machine learning and to unify reporting of machine learning analyses.²² A PRIME checklist for this study can be found in Supplementary data online, Table S2. Furthermore, a preview of the R code for artificial intelligence–enabled mPAP prediction is also found in Supplementary data online, Figure S1.

Results

The study population constituted 737 patients from a multicentric registry with complete echocardiographic and haemodynamic assessment prior to TTVI for severe TR

A total of 737 patients who underwent TTVI for TR from 2016 to 2022 were included in this multicentric analysis. Since this study aimed to analyse RV-PA coupling, only patients with complete echocardiographic and haemodynamic assessment by right heart catheterization obtained prior to TTVI were included in this study (hereinafter referred to as study population). Consequently, 181 from 918 patients (19.7%) with missing measurements of mPAP_invasive, TAPSE, and sPAP_{echocardiography} levels were excluded (Figure 1A). The mean age of the study population (737 patients) was 78.4 ± 7.55 (95% CI: 77.8–78.9) years, and 44.8% of patients were male (Table 1). Patients typically presented with dyspnoea corresponding to New York Heart Association (NYHA) functional class III (76.4%) or IV (13.3%) and with a mean NT-proBNP level of 4513 (95% CI: 3950–5070) pg/mL. Transcatheter edge-to-edge repair was the predominant TTVI technique utilized, followed by annuloplasty and transcatheter valve replacement, in 594 (80.6%), 136 (18.5%), and 7 (0.9%) patients, respectively. While 254 deaths among 737 enrolled patients were recorded, survivors were traced on a median follow-up time of 1.77 years (interquartile range: 1.06–3.42 years; Figure 1B). Severe, massive, and torrential TR were diagnosed in 375 (50.9%), 238 (32.2%), and 124 (16.8%) patients (Table 2 and Figure 1C and D). A successful TR reduction by at least one grade could be achieved in 685 (92.9%) out of 737 cases (Table 3 and Figure 1C and D). Accordingly, 1- and 2-year survival rates for all patients ranged at 80.0% and 67.7%, respectively (Figure 1E).

An XGB algorithm using echocardiographic input parameters can reliably predict mPAP levels as measured by right heart catheterization, and predicted mPAP levels show a higher predictive value than sPAP levels from echocardiography regarding 2-year all-cause mortality prediction after TTVI

To train and validate an XGB algorithm to predict mPAP levels, the study population was further divided into derivation and validation cohorts (Figure 2A), presenting with largely similar clinical, echocardiographic, and haemodynamic characteristics (Tables 1 and 2). Importantly, no statistically significant differences regarding survival following TTVI were detectable between derivation and validation cohorts (Figure 2B). In total, 10 routine parameters from transthoracic echocardiography were used to train the XGB algorithm to predict mPAP levels. Echocardiographic parameters serving as input variables included LVEF, left ventricular end-systolic diameter, left atrial area, sPAP_{echocardiography}, basal right ventricular diameter, TAPSE, TV EROA, TR vena contracta width, right atrial area, and inferior vena cava diameter. Among those 10 input parameters for 737 patients, 317 (4.30%) of 7370 data points had missing values (Figure 2C), and the largest proportion of missing values was found for measurements of left atrial area (16.7% of values missing; Figure 2D). Notably, missing values were imputed only for the derivation cohort but not for the external validation cohort. The mPAP levels predicted by the XGB algorithm showed a highly significant correlation to the mPAP levels measured by right heart catheterization as confirmed both in patients from the derivation cohort (Pearson correlation coefficient R: 0.57, P < 2.2 × 10⁻¹⁶) and from the external validation cohort (Pearson correlation coefficient R: 0.68, P value: 1.3 × 10⁻⁸; Figure 3A and B). Pairwise comparisons revealed that the invasive and predicted mPAP levels were not significantly different (30.3 ± 9.1 mmHg vs. 30.0 ± 3.7 mmHg, P value: 0.978); yet, the predicted mPAP levels exhibited a narrower range compared with the invasively measured mPAP levels (Figure 3C and D). While the overall difference between the invasive and predicted mPAP levels was a mere 0.31 mmHg, larger discrepancies were observed at the extreme ends of the measurement range (see Figure 3E). The top five echocardiographic input parameters with the strongest contribution to mPAP prediction in terms of global feature importance as determined by SHAP values were as follows (in order of predictive importance): sPAP_{echocardiography}, inferior vena cava diameter, left ventricular end-systolic diameter, TAPSE, and LVEF (Figure 3F; see the figure legend for a detailed explanation on how to interpret SHAP values). The accuracy of mPAP prediction was consistent across different grades of preprocedural TR severity (see Supplementary data online, Figure S2) and did not vary in patient subsets, such as those with a history of pacemaker implantation or atrial fibrillation (see Supplementary data online, Figure S3). Furthermore, the ability to predict mPAP levels was consistent among patients treated with different types of TTVI (see Supplementary data online, Figure S4). Acknowledging that sPAP_{echocardiography} is inferior to mPAP_invasive in predicting 2-year all-cause mortality after TTVI [AUC: 0.626 (95% CI: 0.579–0.673) vs. 0.520 (95% CI: 0.471–0.570), P value: 9.6 × 10⁻⁵; Graphical Abstract], it is also important to note that ROC analysis demonstrated superiority of mPAP_predicted levels over sPAP_{echocardiography} levels regarding 2-year all-cause mortality prediction after TTVI [AUC: 0.607 (95% CI: 0.561–0.653) vs. 0.520 (95% CI: 0.471–0.570), P value: 1.9 × 10⁻⁶; Figure 3G]. Moreover, rising mPAP_predicted levels were constantly associated with an increasing HR for 2-year all-cause mortality after TTVI (Figure 3H).

Artificial intelligence–enhanced RV-PA coupling assessment reliably identifies patients with favourable survival outcome after TTVI

Notably, RV-PA coupling expressed as TAPSE/mPAP_invasive serves as a better prognostic marker for 2-year all-cause mortality after TTVI than TAPSE/sPAP_{echocardiography} [AUC: 0.657 (95% CI: 0.613–0.702) vs. 0.586 (95% CI: 0.537–0.635), P value: 6.2 × 10⁻⁴; Graphical Abstract]. However, TAPSE/mPAP_invasive assessment requires invasive measurements of pulmonary artery pressure levels by right heart catheterization. We therefore sought for an alternative and replaced mPAP_invasive with mPAP_predicted levels to assess RV-PA coupling, ultimately showing that also TAPSE/mPAP_predicted is superior to TAPSE/sPAP_{echocardiography} in predicting 2-year all-cause mortality after TTVI [AUC: 0.633 (95% CI: 0.586–0.679) vs. 0.586 (95% CI: 0.537–0.635), P value: 0.008; Graphical Abstract]. While TAPSE/mPAP_predicted also demonstrated numeric superiority over TAPSE/sPAP_{echocardiography} in predicting 2-year all-cause mortality following TTVI in a time-to-event analysis considering all 737 patients—even those who had not yet completed a 2-year follow-up [263 patients (35.7%)]—the superiority was not statistically significant [concordance index: 0.613 (95% CI: 0.574–0.649) vs. 0.568 (0.526–0.610), P value: 0.499; see Supplementary data online, Figure S5]. Applying maximally selected log-rank statistics to dichotomize the study population according to TAPSE/mPAP_predicted levels resulted in an ideal threshold of 0.617 mm/mmHg with respect to 2-year all-cause mortality. RV-PA uncoupling is hence defined by low TAPSE/mPAP_predicted levels (≤0.617 mm/mmHg), while preserved RV-PA coupling is defined by high TAPSE/mPAP_predicted levels (>0.617 mm/mmHg). In comparison to patients with preserved RV-PA coupling, patients with reduced RV-PA coupling presented with more severe signs of congestive heart failure expressed as higher levels of NT-proBNP [5420 (95% CI: 4580–6260) pg/mL vs. 3104 (95% CI: 2540–3670) pg/mL, P value: 5.0 × 10⁻⁷] and enlarged inferior vena cava diameter (26.8 ± 6.14 mm vs. 23.6 ± 6.78 mm, P value: 2.7 × 10⁻¹²; Tables 4 and 5). Notably, a higher proportion of patients with reduced RV-PA coupling were diagnosed with ventricular aetiology of TR (60.5%) compared with those with preserved RV-PA coupling (48.6%; P value: 0.002) as detailed in Table 4. The Kaplan–Meier analysis showed that patients with reduced RV-PA coupling feature a significantly lower survival after TTVI in comparison to patients with preserved RV-PA coupling [2-year survival: 58.8% (95% CI: 53.8–64.2%) vs. 81.5% (95% CI: 76.7–86.6%), HR for 2-year mortality: 2.48 (95% CI: 1.79–3.45), P < 0.001; Graphical Abstract]. To delineate the significance of RV-PA coupling in patients with pulmonary hypertension, we conducted a focused subanalysis among those with predicted mPAP levels ≥ 29.9 mmHg, a threshold previously identified for risk stratification. Of the 349 patients (47.4%) with predicted mPAP levels above this threshold, a mere 67 exhibited preserved RV-PA coupling, while the remaining 282 were characterized by reduced RV-PA coupling (see Supplementary data online, Figure S6A–C). In this selected cohort with predicted mPAP levels surpassing 29.9 mmHg, the Kaplan–Meier analysis revealed that RV-PA uncoupling is associated with an elevated HR for 2-year mortality after TTVI [1.93 (95% CI: 1.10–3.37), P value: 0.021; see Supplementary data online, Figure S6D], underscoring the pivotal role of intact right ventricular function in counteracting any increased afterload burden in the pulmonary circulation. Interestingly, procedural success rates were also significantly lower in patients with reduced RV-PA coupling (Table 4). To test the possibility that higher rates of procedural failure act as drivers for increased mortality in patients with reduced RV-PA coupling, we performed univariate and multivariate Cox regression analyses also taking into account the grade of residual TR (Table 6). The association of higher TAPSE/mPAP_predicted levels with better survival could thus be confirmed by multivariate analysis after adjusting for multiple clinical, laboratory, echocardiographic, and haemodynamic data; independent predictors for mortality were as follows: renal function, dyspnoea expressed as functional NYHA class, TAPSE/mPAP_predicted ratio (which showed the strongest association with mortality in terms of statistical significance), and postprocedural TR grade. Importantly, the validity of the proportional hazards assumption was substantiated, as testing based on Schoenfeld residuals did not reveal any statistically significant violations of this assumption for the included predictors (all P > 0.05), ensuring the robustness of the multivariate Cox regression model (see Supplementary data online, Table S3). Notably, patients exhibiting reduced RV-PA coupling demonstrated a lower LVEF compared with those with preserved RV-PA coupling (51.8 ± 12.0% vs. 57.8 ± 9.14%, P value: 6.8 × 10⁻¹¹; Table 5). Furthermore, LVEF itself was associated with 2-year all-cause mortality after TTVI, as revealed by univariate Cox regression analysis (P value: 0.004; Table 6). When stratifying patients based on left ventricular function (defining preserved and reduced function with an LVEF threshold of 50%), survival curves initially diverged but later converged (see Supplementary data online, Figure S7A). This contrasts with the distinctly separate survival curves observed for patients categorized by RV-PA coupling, presented in the Graphical Abstract. Additionally, ROC analysis affirmed the superiority of TAPSE/mPAP_predicted over LVEF in predicting 2-year mortality after TTVI (see Supplementary data online, Figure S7B).

On the discrepancy and the concomitant prognostic significance among different RV-PA coupling assessments

The possible discrepancy in RV-PA coupling assessments is illustrated using data from an exemplifying patient (Figure 4A): this patient presented with severe TR (III°/V°); echocardiography detected right ventricular dysfunction (TAPSE: 13 mm) and estimated sPAP_{echocardiography} levels at 24 mmHg. At the same time, this patient was diagnosed with pulmonary hypertension by right heart catheterization (mPAP_invasive: 26 mmHg), and mPAP_predicted levels ranged at 28.2 mmHg. Subsequently, a simple echocardiographic RV-PA coupling assessment according to TAPSE/sPAP_{echocardiography} ratio would have concluded that this patient presents with preserved RV-PA coupling, while artificial intelligence–enhanced RV-PA coupling assessment according to TAPSE/mPAP_predicted ratio would have classified this patient to suffer from reduced RV-PA coupling with poor prognosis. A discrepancy in RV-PA coupling assessment was found in 277 (37.6%) out of 737 patients (Figure 4B). Importantly, the Kaplan–Meier analysis confirmed that the prognosis was consistently poor, when artificial intelligence–enhanced RV-PA coupling assessment detected RV-PA uncoupling (Figure 4C).

Discussion

How to guide the right patient with severe TR to the right therapy?

Due to the heterogeneity as commonly encountered in patients with severe TR, it is of eminent importance to evaluate the relationship of the tricuspid valve with the right ventricle and the pulmonary vasculature. In some patients with pre-existing severely dysfunctional right ventricles, TTVI may precipitate acute right heart failure due to the sudden afterload mismatch following restoration of tricuspid valve integrity. RV-PA coupling has therefore emerged as an important determinant for survival after TTVI.⁷ The prognostic significance of RV-PA coupling expressed as TAPSE/sPAP_{echocardiography} levels has been shown in patients undergoing transcatheter intervention for severe aortic stenosis, mitral regurgitation, and TR.^23,24,7 However, echocardiography tends to underestimate sPAP levels in patients with severe TR, because a huge tricuspid valve regurgitant orifice area results in rapid pressure equalization between right ventricle and right atrium.^8,10,11 Fortmeier et al.⁹ have therefore proposed to use TAPSE/mPAP levels—yet necessitating the performance of invasive and time-consuming right heart catheterization. To enable prognostically relevant RV-PA coupling assessment in potential TTVI candidates with severe TR using echocardiography alone, this study employs an artificial intelligence–based algorithm to calculate mPAP levels using echocardiographic input parameters. Analysing data from one of the largest registries of patients undergoing TTVI to date, we show that an XGB algorithm using transthoracic echocardiography data can reliably predict mPAP levels as measured by right heart catheterization (gold standard). Moreover, those mPAP levels outperform sPAP estimations from echocardiography with regard to 2-year mortality prediction after TTVI. Finally, predicted mPAP levels could refine the pathophysiologically meaningful and prognostically relevant RV-PA coupling concept without necessitating invasive right heart catheterization. Apart from its clinical significance, this study is yet another proof that the convergence of human and machine intelligence can result in advanced knowledge from lower-level imaging modalities such as echocardiography.

How to further improve survival in patients with reduced RV-PA coupling?

Due to the lack of a randomized control group, our study is not able to draw definite conclusions about futility of TTVI in patients with reduced RV-PA coupling. Instead, we wish to raise awareness that the armamentarium to treat those patients is currently limited and therefore urgently needs to be expanded:

• Patients with reduced RV-PA coupling presented with the highest mPCWP (20.6 ± 7.32 mmHg), mPAP_invasive (32.3 ± 9.51 mmHg), and PVR (3.25 ± 2.12 WU) levels (Table 5), possibly indicating that long-standing pulmonary hypertension driven by backward transmission of elevated left-sided filling pressures had resulted in remodelling of the pulmonary vasculature (expressed as elevated PVR levels). In fact, resolving pulmonary hypertension in these patients remains a challenge, as there exists no specific pharmacotherapy to lower pulmonary artery pressure levels in patients with pulmonary hypertension due to left-sided heart disease.¹⁷ Treatment with phosphodiesterase type 5 inhibitor sildenafil in patients with persistent pulmonary hypertension after successful correction of left-sided valvular heart disease was even shown to be associated with worse clinical outcomes (death, hospital admission, worsening functional NYHA class, and global symptom burden) as compared with placebo.²⁵ Acknowledging the prognostic significance of pulmonary hypertension, it is therefore of paramount importance to develop pharmacotherapeutic options to ameliorate pulmonary artery pressure levels due to left-sided heart disease. This is particularly true as we are facing a growing category of patients with TR—that is those with persistent TR after transcatheter intervention for aortic stenosis and/or mitral regurgitation.^26–28

• Moreover, patients with reduced RV-PA coupling presented with severe right ventricular dysfunction (TAPSE: 14.4 ± 3.09 mm; Table 5). Did these patients present too late in the moment of right ventricular decompensation [also expressed by the highest NT-proBNP levels of 5420 (95% CI: 4580–6260)  pg/mL; Table 4]? Even if you would consider guideline-directed medical heart failure therapy as the treatment of choice in these patients, it remains noteworthy that left heart failure–specific therapeutics are not necessarily effective in right heart failure, as adaption of the right ventricle to pressure and volume overload differs from the left ventricle on a molecular level,²⁹ possibly reflecting a distinct embryological origin and haemodynamic physiology.³⁰

Taken together, the potential utility of our algorithm lies in its application to real-world scenarios where fine-tuning risk stratification can guide management strategies. For a patient identified by the algorithm as having reduced RV-PA coupling (despite conventional methods indicating preserved status), clinicians might opt for a more aggressive intervention, optimization in the therapy of comorbidities, and closer monitoring. The insights from the algorithm may also facilitate tailored patient counselling and shared decision-making, wherein patients are apprised of their risk status with a higher degree of precision than traditional echocardiography could provide. Finally, in the realm of research, utilizing the algorithm to identify higher-risk patients could enable more targeted enrolment in clinical trials, ensuring that interventions are evaluated in populations where the potential benefits of TTVI remain to be substantiated.

On the variability in defining RV-PA coupling thresholds

There is growing evidence that a one-size-fits-all threshold for RV-PA coupling might not be the most effective approach. For instance, Fortmeier et al.⁹ demonstrated in previous research that men and women with severe TR show distinct patterns in RV-PA coupling. This naturally brings forth the following question: what exactly delineates a ‘healthy’ RV-PA coupling index? It appears plausible that diverse thresholds might be necessary based on gender, resting vs. exercise conditions, or the presence or absence of right ventricular dysfunction. While we recognize the potential value in exploring various RV-PA coupling thresholds in future research, we also caution that an excessive number of thresholds could potentially complicate clinical interpretation.³¹

The added value of our algorithm to assess RV-PA coupling in accordance with TAPSE/mPAP_predicted levels instead of TAPSE/sPAP_{echocardiography} levels becomes evident in Figure 4C. For the large subset of patients who were discrepantly classified as having preserved RV-PA coupling via echocardiography but were identified as having reduced RV-PA coupling by our algorithm [275 out of 737 patients (37.3%), as represented by the orange curve], the survival rate of these patients paralleled that of individuals unequivocally labelled as having reduced RV-PA coupling by both methods. Even though one may argue that for the entire population of patients (including both discrepant and concordant assessments of RV-PA coupling), the superiority of our algorithm over traditional echocardiography expressed as AUC was only 0.633 vs. 0.586 for 2-year mortality prediction (P value: 0.008), the sizable fraction of patients identified differently by our model underscores its potential significance in refining clinical assessments.

On the importance to predict procedural success

We have shown by multivariate logistic regression analysis that residual TR severity acts as an important determinant of survival (Table 6). Prediction of procedural success and concomitant reduction of TR is therefore of great interest, especially as predictors for procedural success may vary depending on the applied technique. For transcatheter tricuspid valve edge-to-edge repair, which is so far the most common technique aiming at improving leaflet coaptation,³² small coaptation gap size and central/anteroseptal TR jet location independently predict procedural success.²¹ If this observation also holds true in patients treated with the Cardioband system, which mimics the surgical approach by implanting an annular reduction system and hence targets annular dilatation as the central pathology in most patients with TR,^33,34 it is still elusive. Filling this knowledge gap is pivotal in order to pave the way to tailored medicine, where the selection of TTVI technique is optimized in accordance with individual characteristics of each patient.

Limitations

As a retrospective, observational, non-randomized registry study, our analysis has inherent limitations. Among them, four major limitations merit consideration.

• Input data quality: We acknowledge that our echocardiography data were obtained during clinical routine without the oversight of a central core lab, also implying that no inter- or intra-observer variability in echocardiographic measurement was assessed. One may argue that this inherent ‘noise’ in the data, arising from day-to-day clinical variability and measurement inaccuracies, could potentially enhance the robustness of the algorithm, preparing it for a wider array of real-world applications. Future iterations of our mPAP prediction algorithm should explore whether the inclusion of more sophisticated parameters, such as RV free wall strain and 3D echocardiography to better capture the intricate geometry of the tricuspid valve, can further augment the algorithm’s predictive capabilities. While the current version of the algorithm predicts a narrower range of mPAP values compared with invasive measurements, it brings about two critical clinical implications: underestimating mPAP levels in patients with severe pulmonary hypertension may mask those at highest risk for irreversible pulmonary hypertension, and overestimating mPAP levels in those without pulmonary hypertension could inadvertently categorize them as affected. Such possible misclassifications in the extreme range of mPAP values carry significant clinical implications, and clinicians should therefore exercise caution when implementing this preliminary algorithm into patient care. The algorithm’s limited accuracy in the extreme range of mPAP values could be improved by training on even more extensive data sets. At present, these extreme values are indeed outliers, given the rarity of patients presenting with such values in our data set. This observation highlights the imperative of consistently refining and updating algorithms as we accumulate more data.

• Study endpoint: This study focused on mortality as an endpoint, taking advantage of its objective nature and the robustness of data acquisition from civil registries (ensuring that no patient was lost to follow-up in this study). However, specific causes of death, such as cardiovascular, aging related, or oncological, are not detailed. Moreover, in a multimorbid patient population, relief of symptoms could be considered an equally desirable treatment aim, and future studies should therefore evaluate symptom alleviation after TTVI in accordance with RV-PA coupling. Additionally, elucidating the postprocedural trajectory of cardiac function in line with RV-PA coupling, and correlating this with both symptom improvement and survival, presents a promising avenue for subsequent investigations.

• Clinical perspective: This study was designed to develop an easily comprehensible, pathophysiologically reasonable, and prognostically meaningful mechanistic model to anticipate which patient will experience the best survival outcome following TTVI. Importantly, our model did not intend to forecast treatment futility, thereby suggesting a complete withholding of TTVI from patients with reduced RV-PA coupling. To better understand the impact of TTVI on longevity in such patients, a randomized study juxtaposing TTVI with a conservative treatment regimen is mandatory. New insights on this topic are eagerly awaited from the CLASP II TR (NCT04097145) and the TRILUMINATE (NCT03904147) trials. Independent of the absence of a control group to measure TTVI’s definitive benefit for those with reduced RV-PA coupling in this study, our model paves the way to identify patients warranting intensified follow-up care.

• Generalizability of the algorithm: Our model was derived from a highly selected cohort of patients with severe TR who had already been identified as candidates for TTVI. Thus, the generalizability of this algorithm to broader populations requires careful consideration. Key questions for future research include:

  • Ethnic variability: Will an algorithm trained on Western patients demonstrate the same reliability when predicting mPAP levels in patients of different ethnicities, such as Japanese individuals with notably smaller heart cavities?

  • Treatment modalities: What is the prognostic value of this algorithm in patients who undergo surgical intervention for severe TR or in those who are managed conservatively without intervention? It is imperative to emphasize that our model, while promising, is not currently designed for broader clinical decision-making regarding surgical indications.

  • Valvular diseases: Is the superiority of TAPSE/mPAP_predicted over TAPSE/sPAP_{echocardiography} maintained in patients with other valvular heart diseases, for example severe aortic stenosis or mitral regurgitation?

Conclusion

Four main conclusions can be drawn from this comprehensive multicentre analysis:

• An XGB algorithm, using echocardiographic input parameters, is proficient at predicting pulmonary artery pressure levels, as validated by right heart catheterization.

• Predicted mPAP levels surpass sPAP levels from echocardiography when it comes to predicting 2-year mortality after TTVI.

• Without requiring invasive right heart catheterization, predicted mPAP levels refine the pathophysiologically meaningful and prognostically relevant RV-PA coupling concept.

• Patients with preserved RV-PA coupling, defined by TAPSE/mPAP_predicted levels exceeding 0.617 mm/mmHg, show significantly better survival rates than patients with reduced RV-PA coupling.

Our algorithm may prove particularly valuable in outpatient settings, addressing the clinical challenges posed by patients suffering from severe TR. Since traditional echocardiography would systematically underestimate pulmonary hypertension in those patients, while right heart catheterization is barely available in an outpatient setting, a reliable, non-invasive assessment of pulmonary hypertension and pertinent insights into the patient’s RV-PA coupling status are crucial. Our algorithm aspires to empower clinicians by providing answers to two critical questions before considering patients for TTVI:

• Does the patient in question suffer from pulmonary hypertension?

• Does the right ventricle demonstrate a sufficiently preserved contractile force, represented as preserved RV-PA coupling, to compensate for afterload challenges?

Taken together, artificial intelligence–enabled RV-PA coupling assessment has the potential to identify ideal TTVI candidates. Randomized controlled studies are mandatory to quantify the net benefit of TTVI vs. conservative treatment in accordance with RV-PA coupling. Moreover, devising treatment strategies aimed at ameliorating survival outcomes for critically ill patients with reduced RV-PA coupling is of paramount importance.

Supplementary data

Supplementary data are available at European Heart Journal - Cardiovascular Imaging online.

Funding

M.L. has received funding from the Technical University of Munich (clinician scientist grant) and from the Else Kröner-Fresenius Foundation (clinician scientist grant). A.H. received funding from the German Cardiac Society (DGK; Otto Hess Doctoral Scholarship).

Data availability

The data underlying this article will be shared upon reasonable request to the corresponding author. All requests for raw and analysed data and related materials, excluding programming code, will be reviewed by the Ethics Committee at Ruhr University Bochum, Germany. Any data and materials that can be shared will be released via a Material Transfer Agreement. The R code for artificial intelligence–enabled mPAP prediction is available at a public GitHub repository (https://github.com/Lachmann1990/Artificial-intelligence-enabled-mPAP-prediction).

References

Author notes