Artificial Intelligence (AI)-Facilitated Analysis of Single-Image Tissue Doppler Signal to Characterize Right Ventricular Dysfunction

BACKGROUND: Quantitative assessment of right ventricular (RV) function by transthoracic echocardiogram (TTE) commonly relies on tricuspid annular plane systolic excursion (TAPSE) and lateral tricuspid annulus peak systolic velocity (S'). However, full cardiac cycle data may provide additional information beyond these two systolic measures.

OBJECTIVE: We sought to (1) automate the estimation of systolic parameters (TAPSE and S') from tissue spectral Doppler imaging (Tissue Doppler Imaging [TDI]) and (2) integrate these tabular systolic parameters and the full-cycle functional signal to estimate RV systolic function.

METHODS: We identified 387 patients who underwent both TTE and cardiac magnetic resonance imaging (CMR) within 24 h. We developed and validated an automated algorithm to extract TAPSE and S' from raw TDI. We trained two classifier models for RV dysfunction (RVEF < 45%): (1) Tabular model (RVD TABULAR) using algorithmic measurement of TAPSE/S' and age/sex, and (2) Integrated model (RVD INTEGRATED), an attention-based neural network model using the entire digitized TDI waveforms in addition to tabular data.

RESULTS: In the TTE-CMR paired dataset, the proposed algorithm accurately estimated S' (mean error: -0.05 cm/s) and TAPSE (mean error: -0.97 mm). Tabular model RVD TABULAR achieved an AUROC of 0.71 and an AUPRC of 0.48 for predicting RVEF <45%, while the integrated model RVD INTEGRATED achieved significantly better performance (AUROC: 0.768; AUPRC: 0.56). In the external validation cohort with pulmonary hypertension (PH), the integrated model's prediction was significantly associated with event-free survival (p = 0.036).

CONCLUSIONS: We developed a fully automated pipeline that integrates digitized TDI waveforms with both parametric and non-parametric features to classify RVEF <45%. This approach can effectively risk-stratify patients with PH.