Deep learning–based right ventricular free wall longitudinal strain predicts mortality in an unselected clinical cohort
V Holmstrom, A Azad, A Chernyshov, E Holte, H Dalen, L Lovstakken, A Ostvik, B GrenneAbstract
Background
Right ventricular (RV) free-wall longitudinal strain (FWLS) is a sensitive marker of RV systolic dysfunction and provides prognostic value across a range of cardiac diseases. However, currently available semi-automated methods for RV-FWLS analysis are time-consuming and operator-dependent, contributing to underutilization in routine clinical practice.
Purpose
To develop a fully automated deep learning (DL) method for RV-FWLS analysis, evaluate its agreement with conventional semi-automatic measurements, and determine its ability to predict the primary outcome of all-cause mortality in a large, unselected clinical cohort.
Methods
Using a combination of DL networks for anatomical key point detection and temporal tracking, a method for fully automated RV-FWLS measurements was developed and trained, first using LV image sequences (n=643) and then fine-tuned on RV-focused image sequences (n=701). Subsequently, the DL method was validated in an independent retrospective cohort of 489 consecutive clinical echocardiograms acquired in 2021. Participants were classified as having reached the primary endpoint if they were listed as deceased in the national death registry, as assessed in October 2025. Agreement and correlation with conventional semi-automated measurements (EchoPAC version 204, GE HealthCare) was assessed by Bland-Altman and Pearson correlation (r) statistics. Prognostic performance was assessed using Kaplan–Meier survival analysis and Cox proportional hazards regression adjusted for age.
Results
Mean age was 61 ± 16 years, and two-thirds were male. Comorbidities included hypertension (34%), heart failure (31%), coronary artery disease (45%), and valvular heart disease (31%). Conventional RV-FWLS was feasible in 407 patients (83%), and DL analysis was successful in every case with feasible conventional measurements. Agreement with conventional semi-automatic measurements was high, with negligible bias (-0.3, p = 0.07) and limits of agreement ranging from –6.4 to 5.9. Correlation was strong (r = 0.87). During a median of 50 months, 88 patients (22%) died. DL-FWLS demonstrated strong prognostic stratification for all-cause mortality, with progressively lower DL RV-FWLS values associated with stepwise reductions in survival (Figure 1). Age-adjusted hazard ratio (HR) per 1% worsening in RV-FWLS was significantly associated with increased mortality risk across 3 months (HR 1.18) to 4 years (HR 1.08) follow-up intervals (Figure 2).
Conclusion
Fully automated DL-based RV-FWLS was feasible, showed excellent agreement with conventional semi-automatic measurements, and demonstrated strong prognostic performance for all-cause mortality. These findings support our automated RV-FWLS as a practical and clinically relevant marker for risk assessment.Figure 1:Kaplan-Meier curvesFigure 2:Hazard ratio