An automated machine learning-based approach to define outcome-driven thresholds for echocardiographic parameters of RV function
M Tolvaj, M Tokodi, M Illyes, E Surkova, D Muraru, Y Nagata, B K Lakatos, A Fabian, A Ferencz, M Tomaselli, M Takeuchi, B Merkely, L P Badano, A KovacsAbstract
Background
Prognostically relevant grading of right ventricular (RV) dysfunction remains challenging. Although recent guidelines provide severity grading for individual echocardiographic parameters, these thresholds often lead to discordant classifications across measures and offer limited guidance on their combined use. Furthermore, most recommended cut-offs are largely consensus-based and lack robust prognostic validation.
Purpose
We aimed to define new, outcome-driven thresholds for RV function parameters using a multiparametric, interpretable machine learning (ML) approach.
Methods
The outcome of interest was the composite of heart failure hospitalization or all-cause death. A three-step, survival tree-based ML approach was applied to derive thresholds and define risk groups. First, a base tree was constructed using three conventional echocardiographic parameters of RV function: tricuspid annular plane systolic excursion (TAPSE), fractional area change (FAC), and free-wall longitudinal strain (FWLS). Second, each leaf node of the base tree was further split using an automatically determined cut-off of 3D echocardiography-derived right ventricular ejection fraction (RVEF), if significant separation in event-free survival was observed between the resulting subgroups. Finally, hierarchical clustering was applied to merge leaf nodes with similar risk profiles, resulting in three final risk groups. For model training and internal testing, a multicenter cohort of 1,119 patients was used, split in a 70:30 ratio. External validation was performed in an independent cohort of 412 patients with secondary tricuspid regurgitation (TR).
Results
The applied automated approach yielded an interpretable decision tree (Figure 1). Interestingly, it did not use TAPSE during base tree generation, and only two leaf nodes of the base tree could be split further, suggesting that assessing RVEF by 3D echocardiography might not improve risk stratification beyond conventional parameters in some patients. The survival curves of the three resulting risk groups separated similarly well in the training set, internal test set, and external validation cohort (Figure 2). Cox analysis showed progressively increasing risk across groups in the external TR cohort, with the moderate- and high-risk groups having hazard ratios (HRs) of 1.948 (95% confidence interval [CI]: 1.024–3.708, p=0.042) and 4.841 (95% CI: 2.345–9.996, p<0.001) compared to the low-risk group, respectively, and the high-risk group having a HR of 2.485 (95% CI: 1.440–4.288, p=0.001) compared to the moderate-risk group.
Conclusions
Using a survival tree-based ML framework, we defined a combination of outcome-driven thresholds for echocardiographic parameters of RV function. The resulting interpretable decision tree demonstrated consistent performance in internal testing and external validation, supporting the clinical relevance of outcome-guided, multiparametric risk stratification.Survival treeKaplan-Meier curves