1Universitätsklinikum Würzburg Dpt. Clinical research & Epidemiology; Deutsches Zentrum für Herzinsuffizienz und Medizinische Klinik I Würzburg, Deutschland; 2Universität Würzburg Institut für Epidemiologie und Biometrie Würzburg, Deutschland; 3TomTec Imaging Systems GmbH Unterschleißheim, Deutschland; 4Universitätsklinikum Würzburg Deutsches Zentrum für Herzinsuffizienz Würzburg, Deutschland; 5Universitätsklinikum Würzburg Service Center Medical Informatics Würzburg, Deutschland; 6Universitätsklinikum Würzburg Medizinische Klinik und Poliklinik I Würzburg, Deutschland; 7Universitätsklinikum Würzburg Institut für Klinische Epidemiologie und Biometrie Würzburg, Deutschland
Background. We recently showed that population data-based machine-learning can improve automated echocardiographic quantification of cardiac structure and function. The respective gain in accuracy and precision strengthens the confidence into automated echocardiographic readings and carries large potential for applications in various settings. We here explored the value of automated reading of echocardiographic images in epidemiology.
Methods and Results. The population-based Characteristics and Course of Heart Failure Stages A-B and Determinants of Progression (STAAB) Cohort Study included a representative sample of the population of Würzburg, Germany (aged 55±12 years, 52% women) stratified for age and sex. Participants underwent detailed cardiovascular phenotyping including transthoracic echocardiography. We assessed risk factors for the development of heart failure (arterial hypertension, obesity, metabolic syndrome, diabetes mellitus and presence of cardiovascular disease) as well as the presence of structural heart disease (left ventricular [LV] dilation, LV hypertrophy, diastolic dysfunction, LV valve stenosis or LV valve regurgitation >mild). AVE (Automatisierte Vermessung der Echokardiographie) is a cooperation project between University and University Hospital of Wuerzburg and Tomtec Imaging Systems using the STAAB echocardiography images within a federated machine learning environment. We applied an automated detector to the echocardiography images of n=4,965 STAAB participants and quantified LV end-diastolic diameter (LVEDD). Wie examined the capability of machine measurements to solve more sophisticated problems. Exemplarily, we asked if the relationship between risk factors and LVEDD was different in men and women (statistical interaction of risk factors and sex).
We performed this analysis twice, using the original human measurements as well as the automated measurements, respectively.
Using the human measurements (figure 1A), we found a significant association between the presence of risk factors and LVEDD in both men (+0.47 mm; 95%CI +0.05 to +0.88 mm, p=0.028) and women (+0.89 mm; 95%CI +0.51 to +1.26 mm, p<0,001). There was a trend towards a more pronounced association in women, but no interaction for presence of risk factors and sex (all p=n.s.). Using the automated measurements (figure 1B), we also found an association between the presence of risk factors and LVEDD (men: +0.42 mm; 95%CI –0.00 to +0.85 mm, p=0.052; women: +1.07 mm; 95%CI +0.68 to +1.46 mm, p<0.001), but with a significant interaction for presence of risk factors and sex (all p<0.05).
Conclusions. The application of an automated detector to a large number of echocardiograms was feasible and led to a stronger result when compared to human measurements. Our results await extension to further echocardiographic parameters and confirmation in other cohorts, but suggest that automated echocardiographic reading might become a valuable tool in epidemiologic studies.