Predicting worsening and mortality in heart hailure: a novel deep learning approach using echocardiographic data outperforms established biomarkers

Heart failure (HF) is a heterogeneous clinical syndrome with uncertain prognosis, particularly in advanced stages. This project aimed to develop and validate a robust deep learning (DL) approach for personalized risk assessment in HF, focusing on the prediction of worsening of HF and mortality.

Echocardiography from participants of the MyoVasc study (n=3,289), a prospective study on HF, were preprocessed using EchoCLIP. The processed videos were used for the development and validation of a novel feedforward neural network, with model performance evaluated on a hold-out test set from the same cohort. A neural network was used to predict all-cause mortality, worsening of HF, and the transition from asymptomatic (Stage B) to symptomatic (Stage C/D) HF over a two years follow-up period. HF stages were determined according to the Universal Definition of Heart Failure. Worsening of HF was defined as a composite of HF-related hospitalization or cardiac death in HF C/D or progression from asymptomatic to symptomatic HF. Transition from asymptomatic to symptomatic HF was defined as having HF stage B at baseline and HF stage C/D at the two-year follow-up examination. Mortality and worsening of HF were estimated using multivariable Cox proportional hazards models, adjusted for age and sex with DL model results, NT-proBNP, left ventricular ejection fraction (LVEF), or E to e prime ratio (E/e’) as predictors in separate models. C-indices were compared using a non-parametric statistical test to evaluate whether the DL model results demonstrated superior prognostic performance compared to established clinical markers.

A total of 7,441 videos from 2,765 participants was analysed. Follow-up time for all-cause mortality (626 events) was 10 years, for worsening of HF (317 events) was 4 years, and for the transition from stage B to C/D (167 events) was 2 years. The DL model demonstrated high performance across all three tasks, with an area under the curve of 0.83 for all-cause mortality, 0.80 for worsening of HF, and 0.76 for the transition from asymptomatic to symptomatic HF. It correctly identified 84.6% of mortality events, 79.2% of worsening of HF events, and 76.5% of transition events. The DL model had equally good performance across all HF phenotypes (HFpEF, HFmrEF, and HFrEF). The DL model outperformed NT-proBNP, LVEF and E/e’ in predicting all-cause mortality (p<0.0001). For the prediction of worsening of HF, the DL model results outperformed NT-proBNP (p=0.04) and E/e’ (p=0.0002), but were equal to LVEF.

The novel DL model presented in this work was able to identify individuals at risk of worsening of HF and death and outperformed gold-standard HF markers, suggesting an advancement in HF risk assessment by this approach.