Aim: To improve accuracy of the automated echocardiographic quantification of cardiac structure and function by federated re-training of a machine learning algorithm (MLA).

Methods & Results: The population-based Characteristics and Course of Heart Failure Stages A-B and Determinants of Progression (STAAB) Cohort Study included individuals from the general population of Würzburg, Germany, aged 30-79 years, stratified for age and sex. From the total STAAB sample (n=4,965, 52% women, mean age 55±12 years) 98% underwent transthoracic echocardiography (Vivid S6 / E95, GE Healthcare). AVE (Automatisierte Vermessung der Echokardiographie) was performed as cooperation project between the University Hospital & University of Würzburg and Tomtec Imaging Systems, Unterschleißheim. The stored echocardiography images were imported into the TomtecArena® located at the Academic Core Lab Ultrasound-based Cardiovascular Imaging of the Comprehensive Heart Failure Center, Würzburg, Germany. Trained and internally certified personnel performed measurements according to a pre-specified protocol including left ventricular (LV) enddiastolic diameter (LVEDd), LV endsystolic diameter (LVEDs), interventricular septum (IVSd), LV posterior wall (LVPWd), LV septal and lateral relaxation velocity (e´), early (E) and late (A) mitral inflow velocity. Thus derived measurements constituted the human referent. For AVE, a random algorithm allocated n=3,226 participants to a training pool and re-training of the MLA was performed based on the respective human referent measurements. Both the original detector as the re-trained detector were then applied to the echocardiograms of n=563 participants allocated to the validation pool. Comparing the automated measurements of the original detector with the human referent (scenario A; Bland-Altman analysis), we observed a significant measurement difference for LVEDd, LVEDs, LVPWd, E, A, e´ lateral, and e´septal, respectively, whereas no such differences emerged for IVS and LVOT diameter (table). When contrasted with the original detector´s performance, the re-trained detector revealed superior accuracy as it arrived at significantly smaller mean differences in all but one parameter (IVS) with respect to the human referent (scenario B, table). Specifically, all differences between the re-trained detector and the human referent became non-significant, with the exception of LVEDd and LVPWd, respectively.