Deep learning-based strain echocardiography adds prognostic value in heart failure with reduced ejection fraction
Daniela Melichova, John Nyberg, Thuy Mi Nguyen, Ivar M Salte, Øyvind O Salvesen, Havard Dalen, Lasse Lovstakken, Harald Brunvand, Kristina H Haugaa, Thor Edvardsen, Andreas Østvik, Bjørnar GrenneAbstract
Aims
Echocardiographic global longitudinal strain (GLS) and mechanical dispersion (MD) improve risk prediction in heart failure. To facilitate clinical implementation, we aimed to develop a deep learning-based method for fast, fully automated GLS and MD and validate its ability to predict ventricular arrhythmia (VA) and all-cause mortality.
Methods and results
This prospective observational study included 568 patients (73 ± 13 years, 67% male) with first diagnosis of heart failure and reduced ejection fraction (HFrEF). A deep learning-based method was designed to measure echocardiographic GLS and MD. Median follow-up was 4.8 years (interquartile range: 2.9, 6.1 years). VA occurred in 49 patients (9%), while 118 patients (21%) died. Optimal cut-off values for GLS and MD were 9% and 70 ms. Patients with both impaired GLS and increased MD were at high risk of VA (HR 6.98 [95% CI: 3.23–15.1]) compared to those with either impaired GLS or increased MD alone (HR 2.84 [95% CI: 1.38–5.86], P < 0.001). GLS and MD were independently associated with VA and provided potential incremental prognostic value beyond ejection fraction. Feasibility of deep learning measurements was high (96%), with processing time of 6.4 ± 2.5 seconds per echocardiogram.
Conclusions
A novel deep learning-based method for fully automated GLS and MD was feasible, fast, and predicted VA in patients with HFrEF. Combining GLS and MD provided potential incremental prognostic value beyond GLS, MD, or EF alone. This approach may enhance workflow and facilitate arrhythmic risk stratification. Further studies are needed to determine whether it can improve selection of patients for ICD therapy.