Identification of risk phenogroups among patients with moderate-to-severe tricuspid regurgitation by unsupervised cluster analysis

Abstract Background Assessing the individual risk of patients with secondary tricuspid regurgitation (STR) is challenging as it requires integrating disease severity and mechanism with different clinical and imaging characteristics. Several scores and cluster analysis have recently been proposed, almost all of them using visual estimation or conventional echocardiography for the evaluation of right ventricle (RV) size and function. However, advanced echocardiography has been scarcely used so far in this context. Purpose We sought to identify the different phenogroups of STR using unsupervised cluster analysis and assess their association with clinical outcomes. Methods We included 558 consecutive patients (mean age 74±14 years, 54.7% women) with moderate and severe STR who underwent comprehensive two-, three-dimensional, and speckle-tracking echocardiography. The primary endpoint was a composite of heart failure hospitalization and all-cause mortality. Four unsupervised algorithms were used to cluster STR patients based on 16 variables recruited from demographics or echocardiographic domains, routinely obtained for the assessment of STR or having prognostic value. Results Over a median follow-up of 16.5 months, 215 patients reached the composite endpoint. A cluster analysis divided the patients into three phenogroups with distinct characteristics: Phenogroup 1 "lowest-risk STR" (rate of event-free survival at two years: 80±3%) with moderate STR, preserved RV size and function, and moderately dilated but normofunctional RA; Phenogroup 2 "intermediate-risk STR (hazard ratio [HR] 2.37, 95% confidence interval [CI] 1.14-4.92, p=0.02) characterized by older age, severe STR, mildly dilated but uncoupled RV; and Phenogroup 3 "highest-risk STR" (HR 3.83, 95% CI 1.94-7.54, p<0.001), characterized by younger age, massive-torrential TR, severely dilated and overtly dysfunctional RV, and dysfunctional RA. After adjusting for multiple variables, the clustering analysis remained independently associated with the composite endpoint (HR 1.40, 95% CI 1.13–1.70, p=0.002, for each increased-risk phenogroup). Additionally, a supervised machine learning model was developed to assist clinicians to assign patients to one of the three phenogroups. The model successfully matched each patient with an accuracy of 0.91, precision 0.91, recall 0.91, and F1 score of 0.91. Conclusions The unsupervised cluster analysis has identified three risk phenogroups with distinct characteristics and different risk of experience death or hospitalization for HF which could potentially assist clinicians in developing more personalized treatment and follow-up strategies for STR patients. Figure 1 Figure 2