Determinants of Outcomes in Patients with Severe Tricuspid Regurgitation Treated Conservatively, Surgically or Percutaneous

Background: Severe tricuspid regurgitation (TR) has seen an increase in both interventional as well as surgical treatment. Yet, prognostic factors according to treatment modality have not been investigated in those patients. Leveraging the potential of machine learning decision tree methods, we aimed to identify the most important outcome features in accordance with treatment modality in patients with severe TR. 

Methods: Patients with severe isolated functional TR which were enrolled in the TRIGISTRY were included in this study. For prediction of the primary endpoint of all-cause mortality at 2-years, three independent eXtreme gradient boosting (XGBoost) models were constructed for patients treated conservatively, surgically or interventionally and model as well as relative feature importance was compared. Patients were randomly split in a derivation (80%) and internal validation (20%) set. Cross-fold validation of the derivation cohort was used for establishing XGBoost model and to avoid overfitting. 

Results: Of 2,344 available patients 1,217 were treated conservatively (median age 74 years, IQR 65 to 81, 44% ♀, 2-year mortality 27%), 518 surgically (median age 70 years, IQR 63 to 76, 63% ♀, 2-year mortality 21%) and 609 interventionally (median age 79 years, IQR 75 to 82, 63% ♀, 2-year mortality 19%). 

The three independent XGBoost models showed good and comparable discriminatory performance for the prediction of all-cause mortality in conservatively (Concordance-index 0.83; 95% confidence interval 0.81 to 0.87), surgically (C-index 0.93; 95%CI 0.87 to 0.98) and interventionally (C-index 0.93; 95%CI 0.87 to 0.98) treated patients (Fig. 1). Among factors relevant for the prediction of all-cause mortality diuretic dose and renal function as assessed by estimated glomerular filtration rate were by far the most important factor for conservatively (relative influence [RI] 24.2% and 15.8%, respectively) managed patients. For surgically managed patients’ renal function and diuretic dose showed equal importance (RI 13.8% and 13.4%, respectively). In interventionally treated patients’ renal function (RI 12.9%), and only to a lesser extent diuretic dose (RI 5.3%) played an important role. In interventionally treated patients right ventricular function, assessed by tricuspid annulus plane systolic excursion, was the most influential factor (RI 13.0%), which was in contrast to surgical and conservative patients (RI 3.3% and 4.6%, respectively). Left ventricular ejection fraction was also more influential in interventionally treated as compared to surgically or conservatively managed patients (RI 8.8, 6.6 and 2.9%, respectively) Further, previous heart failure hospitalization was a relevant determinant of outcome in interventionally, but not in conservatively or surgically managed patients. 

Conclusion: Machine learning models can be used to individualize risk prediction for TR patients according to their treatment strategy. Further, risk factors for all-cause mortality show an expected large overlap between conservatively, surgically and interventionally treated patients. However, subtle differences exist between those three treatment approaches and ventricular function as well as previous heart failure hospitalization seem to play a crucial role in the management of interventionally treated patients with severe TR, while renal function and diuretic resistance seem to be more important in conservatively and surgically managed patients.