Artificial Intelligence–Driven Detection, Characterization, and Risk Stratification in Patients with Severe Tricuspid Regurgitation

Abstract

Abstract Background: Tricuspid regurgitation (TR) is prevalent and challenging to manage. There is a high mortality rate associated with isolated surgery and late referral for intervention. Additionally, there has been a recent expansion of transcatheter therapies. In Brazil, there is a lack of data and validation of prognostic tools. Objective: To investigate and define clinical, laboratory, and echocardiographic data from patients with severe TR based on variables collected through the training of an artificial intelligence (AI) model. Methods: This observational, single-center, non-interventional, retrospective study was based on 71,911 echocardiographic reports performed between 2021 and 2024. A natural language processing (NLP) model was developed in Python to identify cases of severe TR, extract variables, and perform automated calculation of the TRI-SCORE and EuroSCORE II, which were subsequently reviewed. A two-tailed p-value < 0.05 was considered statistically significant. Results: A total of 803 patients with severe TR were identified. The mean age was 68.8 ± 13.5 years, and 64.1% were female. Hypertension (71.7%) and atrial fibrillation (69.4%) were the most prevalent comorbidities. Secondary etiology predominated (96.8%), mainly associated with mitral valve disease (68.5%). Surgical risk was high (medians: EuroSCORE II = 13.4%; TRI-SCORE = 7). Mortality was 17.1%. In the secondary mortality analysis, multivariable regression identified independent predictors, such as high-risk TRI-SCORE (relative risk [RR] 2.19; 95% CI 1.14-4.21) and increased creatinine (RR 3.13; 95% CI 1.91-5.14). Conclusions: In the largest Brazilian cohort of severe TR, constructed with the aid of AI, the disease demonstrated high clinical and prognostic complexity. The application of NLP to large echocardiographic databases proved to be efficient and may support multicenter studies.