The derivation cohort was obtained from a retrospective analysis of echocardiographic examinations prospectively collected between October 2020 and March 2022 (FUTURE 3D. ClinicalTrials.gov Identifier: NCT05747404). We included consecutive patients over 18 years of age with moderate-to-severe STR, good acoustic windows, complete echocardiographic studies, and a minimum follow-up period of 6 months. Exclusion criteria included primary tricuspid regurgitation (TR), cardiac implantable electronic device-related TR, previous surgical or transcatheter tricuspid valve (TV) interventions, pregnancy, inadequate echocardiographic images, and lack of follow-up data. For external validation, we included a cohort of patients prospectively enrolled at Yonsei University College of Medicine, Seoul, South Korea, between January 2022 and February 2024, adhering to the same inclusion and exclusion criteria. Patients with STR were classified into atrial or ventricular phenotypes according to recent recommendations.6,8

This retrospective analysis was approved by the Ethics Committee of our Institute (record #2021_05_18_13, approved on May 18, 2021). Patients’ informed consent was obtained and archived for the publication of their cases. The investigation conforms to the principles outlined in the Declaration of Helsinki.

A set of 16 variables, derived from demographic or echocardiographic domains routinely obtained for the assessment of STR severity or with proven prognostic value, were considered as candidates for analysis (appendix A of the supplementary data). To determine the optimal number of covariates and eliminate redundant information among the variables, those with poor representation in the first 4 dimensions of principal component analysis—defined by squared cosine values and a correlation coefficient >0.8—were filtered out.

To identify the optimal number of phenogroups within the study cohort, the elbow method was used, with the total within the sum of squares serving as the metric. Unsupervised clustering was performed using 3 methods: a partitioning method (K-means), a hierarchical method (agglomerative; balanced iterative reducing and clustering hierarchies [BIRCH]), and a model-based clustering method (Gaussian mixture model).

For internal validation, commonly used cluster validity analysis indexes—including Silhouette, Davies-Bouldin, Dunn, and S-Dbw, which combine measures of cluster compactness and separation—were evaluated to determine the optimal clustering technique and the number of clusters in the final solution. The optimal unsupervised clustering technique was identified as the one with the highest Silhouette and Dunn indexes and the lowest Davies-Bouldin and S-Dbw indexes. The cluster labels were then used as input for supervised machine learning.

We developed a machine learning model using XGBoost (a gradient boosting tree-based algorithm, appendices B and C of the supplementary data) to predict the differential phenogroups of individual participants. The goal was to assist clinicians in identifying the risk profile and prognosis of STR in each patient. Twelve routinely available variables were selected as inputs for the machine learning algorithm (appendix D of the supplementary data). We evaluated the model's performance using balanced accuracy, macro precision, macro recall, and macro F1-score (appendix E of the supplementary data). We used the SHAP (SHapley Additive exPlanations) interpretability method to determine the importance of ranking features from the model's predictive output and to provide an explanation for individual phenogroup predictions (appendix F of the supplementary data). As part of the internal validation, which involves running the model in a randomly-selected portion of study cohort, a Monte Carlo cross-validation with 100 iterations was carried out.

One-way ANOVA, chi-squared, or Fisher exact tests were used, as appropriate, to compare the distribution of continuous and categorical variables between phenogroups. Continuous variables were summarized as means±standard deviation or medians [interquartile range (IQR)], while categorical variables were summarized as frequencies and percentages. The data were standardized to remove differences in variable scales, although the original forms were retained for interpretation. Pearson correlation coefficients were calculated to analyze collinearity.

To visualize differences in parameters between the phenogroups, we generated a heatmap and used a radar plot to illustrate echocardiographic characteristics across phenogroups. The risk of the composite study endpoint among phenogroups was assessed using time-to-event analysis (Kaplan-Meier method) with the log-rank test. Cox proportional hazards models were used to estimate hazard ratios (HR) between phenogroups, with results presented as HR and 95% confidence intervals (95%CI).

The multivariable model was adjusted for NYHA class, presence of chronic kidney dysfunction (CKD), effective regurgitant orifice area, and RA and RV volume and function. Variables were selected for multivariable analysis based on their clinical significance (ie, known risk factors) and their significant association with outcomes in univariable analysis. To address the competing risk of all-cause death, the Fine-Gray competing risk regression model was applied, and subdistribution HRs (sHR) with 95%CIs were provided.

We used the likelihood ratio test and changes in the C-index, calculated with the DeLong test, the integrated discrimination improvement, and continuous net reclassification index to determine the improved prognostic performance of the model with the inclusion of the phenogroup variable. Additionally, we assessed the incremental prognostic value of the identified phenogroups over known clinical and echocardiographic parameters using the chi-squared Omnibus test of model coefficients. The statistical analysis was performed using Python version 3.10.12 and R version 4.3.2 (details of the packages used are provided in appendix G of the supplementary data).

A total of 558 and 200 patients were included in our derivation and validation cohorts, respectively (figure 1 of the supplementary data, table 1 of the supplementary data and table 2 of the supplementary data). No patient underwent surgical/transcatheter TV interventions during the study period.

Eleven out of the 16 discriminatory variables were used in the unsupervised cluster analysis (figure 2 of the supplementary data).

The first 4 dimensions of a principal component analysis accounted for 72.7% of the total variance in the dataset. According to the results from the elbow method, the optimal number of clusters was 3. K-means clustering was selected as the best algorithm for clustering based on validity indices.

Table 1 and figure 1 show the clinical and echocardiographic differences among the patients in the derivation cohort included in the 3 phenogroups. There was a large overlap of TR severity among the patients included in the 3 phenogroups (figure 2)

Figure 1.

Differences in clinical and echocardiographic variables among the lowest-, intermediate, and highest-risk phenogroups. A, heatmap of phenotypic characteristics across the selected clusters. Data are given as color-coded standard deviations from the mean after centering and scaling. B, the radar chart depicting the echocardiographic features and age, highlights the differences among the clusters. The lines represent the cluster average values across the standardized features. EROA, effective regurgitant orifice area; ESV, end-systolic volume; fSV, forward stroke volume; FWS, free-wall longitudinal strain; MR, mitral regurgitation; NYHA, New York Heart Association; PASP, pulmonary artery systolic pressure; RA, right atrial; RV, right ventricular; TAPSE, tricuspid annular plane systolic excursion; TR, tricuspid regurgitation.

(0.37MB).

Figure 2.

Sankey plot showing the distribution of tricuspid severity grades among the 3 phenogroups. EROAc, corrected effective regurgitant orifice area.

(0.25MB).

The phenogroup at the lowest risk (with an incidence of 21.3% for the composite endpoint and 12.8% for both all-cause death and HHF at 2 years) included mildly symptomatic patients with predominantly moderate STR, a higher prevalence of atrial STR, a low prevalence of CKD, normal RV size and function, and mildly dilated RA with preserved function. Patients included in the intermediate-risk phenogroup had a higher incidence of the composite endpoint (32.9%), death (18.2%), and HHF (17.6%) than the previous phenogroup. They were the oldest and mildly symptomatic, with severe STR, a higher prevalence of atrial STR and atrial fibrillation (AF), and moderately dilated RV. Although conventional RV function indexes were normal, the RV-pulmonary artery coupling and the effective RV ejection fraction were borderline low. The RA was severely dilated and dysfunctional. Finally, patients in the highest-risk phenogroup (with an incidence of the composite endpoint, death, and HHF of 64.4%, 31.1%, and 42.9%, respectively) were the youngest and exhibited massive or torrential STR, along with the highest prevalence of ventricular STR. They also had a significantly higher prevalence of CKD, high pulmonary artery systolic pressure values, and severely dilated and dysfunctional RV and RA. In addition, high-risk STR patients had a significantly higher prevalence of left ventricular dysfunction and moderate or severe mitral regurgitation (figure 3).

Figure 3.

Central illustration. Characterization of 3 phenogroups of patients with secondary tricuspid regurgitation using unsupervised clustering analysis. RA, right atrium; RV, right ventricle; STR, secondary tricuspid regurgitation.

(0.41MB).

After a median follow-up of 17 months [IQR 9-23], 215 patients in the derivation cohort reached the combined endpoint, which included 133 HHF and 113 deaths. The event-free survival was 76±2% after the first year and 58±3% after the second year of follow-up (figure 3 of the supplementary data, left panel). In the derivation and validation cohorts, 50% of events occurred within 502 and 432 days after the baseline echocardiogram, respectively (figure 3 of the supplementary data). The median time to survival was 20 months in the derivation cohort and 16 months in the validation cohort, respectively (figure 4 of the supplementary data).

The risk of adverse clinical events during follow-up significantly increased from the lowest- to the highest-risk phenogroup (P<.001 for all) (figure 4).

Figure 4.

Kaplan-Meier plots for the analysis of cumulative event-free survival of patients from the derivation (left panel) and validation (right panel) cohorts included in the lowest-, intermediate-, and highest-risk clusters. 95%CI, 95% confidence interval; HR, hazard ratio.

(0.36MB).

Compared with the lowest-risk phenogroup, patients in the intermediate-risk phenogroup doubled the risk of the composite endpoint (HR, 2.20; 95%CI, 1.44-3.37) and death (HR, 2.01; 95%CI, 1.15-3.53; P=.015). Similarly, patients in the highest-risk phenogroup had almost a 5-fold increased risk of experiencing the composite endpoint (HR, 4.67; 95%CI, 3.20-6.82), 4 times the risk of HHF (HR, 4.21; 95%CI, 2.68-6.62, and a 3-fold increased risk of death (HR, 2.98; 95%CI, 1.78-4.99) compared with those in the lowest-risk phenogroups (P<.001 for all) (figure 4, figure 5 of the supplementary data). On multivariable analysis (adjusted for NYHA class >2, CKD, effective regurgitant orifice area, RV ejection fraction, RV end-diastolic volume index [RVEDVi], right atrial longitudinal strain [RALS], and right atrial maximal volume index), the phenogroup variable was independently associated with the composite endpoint (HR, 1.57; 95%CI, 1.23-2.0; P<.001, for each risk increase) (table 2, model 1). However, in the competing risk analysis, the cluster variable showed an sHR of 1.32; 95%CI, 0.97-1.78; P=.075).

The additional prognostic value of the clustering analysis in patients with STR was further evaluated along with clinical parameters (NYHA class, CKD), STR severity, right atrial volume, RALS, RVEDVi, and RV ejection fraction, using hierarchical model chi-squared analysis (figure 5).

Figure 5.

Hierarchical Cox regression analysis showing the incremental clinical value of clusters. Model 0 was a multivariate basal model including clinical parameters and tricuspid regurgitation severity. New York Heart Association (NYHA) class III-IV, chronic kidney disease (CKD), and tricuspid regurgitation effective regurgitant orifice area. In model 1, we added right atrial volume (RAVI) and longitudinal strain (RALS) to model 0. In model 2, we added right ventricular volume (RVEDVI) and ejection fraction (RVEF) to model 1. Finally, in model 3 we added the 3 clusters to model 2.

(0.17MB).

Finally, the C-index for phenogroups for predicting the 2-year composite endpoint was 0.72 (95%CI, 0.68-0.76 at 2 years). When added to the multivariable model, as described above, there was a significant improvement in the model's goodness of fit (likelihood ratio χ2=15.9, P<.001) and discrimination statistic (change in C-index from 0.75 to 0.76, P by DeLong test=.048; integrated discrimination improvement=0.03, 95%CI, 0.007-0.059; P<.001; and reclassification net reclassification index=0.176, 95%CI, 0.044-0.29; P<.001).

The severity and etiology of TR were found to be independently predictive of the composite endpoint through Cox regression analysis. When adjusting for TR severity or atrial STR, the phenogroup remained a significant independent risk factor for the endpoint (table 2, models 2,3).

The composite endpoint occurred in 90 (42%) men and 125 (58%) women (table 1 of the supplementary data). After adjusting for sex, the phenogroups were associated with a higher risk of the outcome (table 2, model 4). Furthermore, after adjusting for TR severity, the phenogroups remained independent predictors of the endpoint, in both men (HR, 1.95; 95%CI, 1.46-2.61; P<.001) and women (HR, 1.83; 95%CI, 1.45-2.31; P<.001).

During a median follow-up of 15 months [IQR 7-17], patients in the external validation cohort experienced 27 deaths and 27 HHF. Kaplan-Meier analysis at 18 months showed that cluster 3 in the validation cohort had significantly lower event-free survival rates compared with clusters 1 and 2 (figure 4, figure 6 of the supplementary data).

The patients were split randomly into training (70%) and test (30%) sets, ensuring an equal distribution of events in both sets. The machine learning model (figure 6) demonstrated excellent clinical performance both in the derivation test set (accuracy=0.91, precision=0.91, recall=0.91, and F1 score=0.91) and in the external validation cohort (accuracy=0.80, precision=0.78, recall=0.78, F1 score=0.77) (figure 7 of the supplementary data).

Figure 6.

This analysis uses data obtained from 2 patients to illustrate the predictions made by the machine learning model via a force plot. Features contributing positively to predicting the phenogroup are highlighted in red, while those indicating that a patient will not belong to that phenogroup are shown in blue. The size of the arrows represents the relative impact of each feature on the model's predictions. EROA, effective regurgitant orifice area; ESV, end-systolic volume; fSV, forward stroke volume; PASP, pulmonary artery systolic pressure; RALS, right atrial longitudinal strain; RAVI, right atrial maximal volume index;RegFr, regurgitant fraction; RV, right ventricular; RVEDVi, right ventricular end-diastolic volume index; RVEF, right ventricular ejection fraction; TAPSE, tricuspid annular plane systolic excursion; TR, tricuspid regurgitation.

(0.22MB).

Similar results were observed from the stability analysis, with the model achieving a mean accuracy, precision, recall, and F1 score of 87%, with a 95%CI, of 86-87% across Monte Carlo cross-validation (figure 8 of the supplementary data).

The impact of the included parameters on the model's ability to predict inclusion in a specific phenogroup is shown in figure 9 of the supplementary data.

In addition to the severity of STR, it is important to consider RV and RA remodeling, patient age, and kidney function when stratifying the risk and managing patients with moderate-to-severe STR. Our results highlight the clinical heterogeneity in cardiovascular risk among these patients which is primarily associated with different extents of RA and/or RV remodeling, going beyond the current atrial and ventricular STR classification. Unlike previous studies,23 we were able to translate the results of the clustering analysis into a clinically applicable model that accurately predicts the STR patient phenogroup and assists clinicians in identifying patients in the intermediate-risk phenogroup. Although these patients do not have an indication for valve repair according to current guidelines, they exhibited a 59% incidence of either death or HHF at the 2-year follow-up.

This study is an observational and retrospective analysis that utilizes prospectively collected data, which inherently comes with limitations due to its design. Additionally, 11% of the echocardiographic studies were of poor quality, and 8% of the patients were lost to follow-up. These factors may have introduced selection bias and affected the generalizability of our findings. Following the T-VARC recommendations, we used all-cause mortality as the endpoint.35 Issues related to missing data prevented us from including some important biomarkers such as N-terminal pro-B-type natriuretic peptide and troponin.