Deep learning-based prediction of cardiac blood biomarkers from echocardiograms improves prognosis

Background This study aimed to develop a deep learning (DL) model that uses transthoracic echocardiography videos to estimate current and future cardiac biomarker concentrations and to assess their prognostic value for clinical outcome. Methods Echocardiography videos from participants of the MyoVasc study (n=3,289), a deeply phenotyped prospective HF cohort with biannual follow-up, were analysed by a neural network building on the preprocessing by the EchoCLIP DL model. The videos were used to train and validate neural networks (discovery step), which were subsequently tested on an independent data set. The cardiac biomarkers concentrations of NT-proBNP, Troponin I, and MR-proANP were predicted by DL at baseline (BL) and at 2-year follow-up (2yFU). The DL-based predictions of biomarkers were compared to laboratory-assay based measurements. Cox regression in the independent data set, adjusted for age and sex, was applied to evaluate the prognostic value for worsening of HF, three-point major adverse cardio- and cerebrovascular events (MACCE) and mortality. Worsening of HF was defined as a composite of HF-related hospitalizations, cardiac death and progression to clinically established HF stages. MACCE included myocardial infarction, stroke and cardiac death. Results A total of 5,439 videos from 2,046 individuals was analyzed. Follow up time for worsening of HF (n=43 events) and MACCE (n=41 events) was 4 years, and 10 years for all-cause death (n=87 events). DL model predictions strongly correlated with lab-measured biomarker concentrations at both BL and 2yFU for NT-proBNP (rBL=0.76, rFU2=0.67), MR-proANP (rBL=0.76, rFU2=0.63), and Troponin I (rBL=0.64, rFU2=0.60). Prediction of incident worsening of HF using DL-estimated NT-proBNP at 2yFU (Hazard ratio [HR]=3.56[2.26; 5.61], C-Index [C]=0.81) outperformed laboratory measurements at both BL (HR=1.98[1.43; 2.75], C=0.66) and 2yFU (HR=1.99[1.46; 2.72], C=0.73). Prediction of incident worsening of HF using DL-estimated Troponin I at 2yFU (HR=3.58[2.11;6.05], C=0.80) outperformed laboratory measurements (HR=1.97[1.45;2.68], C=0.75), whereas Troponin I at 2yFU from both DL-estimation and laboratory measurements was comparable in predicting mortality and MACCE. DL estimated concentrations of MR-proANP at BL (HR=4.36[2.06; 9.23], C=0.73) predicted incident MACCE, outperforming laboratory measured MR-proANP (HR=1.99[1.21; 3.25], C=0.69). Additionally, DL estimations of MR-proANP at 2yFU (HR=4.33[2.42; 7.75], C=0.81) predicted incident worsening of HF, outperforming laboratory measurements at both BL (HR=1.52[1.29; 1.79], C=0.65) and 2yFU (HR=2.33[1.44; 3.75], C=0.73). Conclusion Deep learning-based analysis of echocardiograms can reliably predict current and future cardiac biomarker concentrations, which outperform laboratory measurements in predicting clinical outcome in HF in this cohort, enhancing HF risk stratification and allowing future insights in HF disease progression.