https://doi.org/10.1007/s00392-025-02625-4
1Universitätsklinikum Würzburg Deutsches Zentrum für Herzinsuffizienz/DZHI Würzburg, Deutschland; 2TomTec Imaging Systems GmbH Clinical Innovations Unterschleißheim, Deutschland; 3Universitätsklinikum Würzburg Medizinische Klinik und Poliklinik I Würzburg, Deutschland
Background: Handheld echocardiography (HHE) devices are small and easy-to-use devices, suited for settings like emergency care or a general practitioner’s office, but for the trade-off of lower resolution and image quality. Application by non-cardiologists might further limit diagnostic accuracy of the HHE derived parameters.
Purpose: We recently showed that automated reading of high-resolution echocardiography images results in high measurement accuracy. Here, we aimed to assess the performance of an automated image selection tool (AIST), identifying appropriate images for the analysis, as well as of the automated analysis of left atrial (LA) volume and strain in HHE images compared to cart-based ultrasound machine (CBUM).
Methods and Results: PAVE (Pathology-oriented automated reading of echocardiography) is a cooperation project between University Hospital Wuerzburg (UKW) and TOMTEC Imaging Systems GmbH using an established federated machine-learning environment. We here report on n= 400 PAVE participants with prospectively CBUM and HHE images acquired within one session. An echocardiographer from the Academic CoreLab Ultrasound-based Cardiovascular Imaging at the Comprehensive Heart Failure Center Würzburg assessed the correctness of the AIST, which calculates the probability of containing an apical 4-chamber view for each of the available clips. In a second step, the clinician analyzed the HHE and CBUM images, respectively, using the AutoStrainLA in the Ultrasound Workspace® blinded to the other method´s results. Agreement of measurements was assessed using Bland-Altman analysis.
From the n=400 HHE, n=3 had no suitable 4-chamber view and were excluded from further analysis. One study (HHE) was excluded from measurements due to invalid annotation.
Table 1 summarizes the performance of the AIST. It significantly trims the study to a relevant subset of clips suitable for AutoStrainLA analysis (reducing the number of clips to 8% and 20% for CBUM and HHE respectively). Additionally, in most cases, the top-ranked AIST clip was also the clip preferred by the expert. Adjusting the AIST to propose only one image per study for AutoStrainLA analysis would have been enough to process 89.5% and 92.4% of the CBUM and HHE studies, respectively.
Table 1: Acceptance rate of AIST selected 4-chamber view
|
Method |
# of Studies |
Overall # Clips |
Overall # AIST-proposed Clips |
First clip accepted and measured |
Second clip accepted and measured |
Third-Fifth clip accepted and measured |
|
CBUM |
400 |
15654 |
1368 |
89.5% |
6.3% |
4.2% |
|
HHE |
396 |
5510 |
1088 |
92.4% |
6.3% |
1.3% |
The LA ejection fraction analysis showed some discrepancies between of both methods ([%] CBUM 44.9 ± 15.9, HHE 43.3± 14.7) as well as in LA axis length [cm] (CBUM 5.4 ±0.9, HHE 5.4 ± 0.9 ). This could be due to a change in the insonation angle between both methods.
Agreement for the LA reservoir stain for n=396 HHE and CBUM was -2% (-22. 4%; 18.4%) (Fig. 1), for LA-EF 1.5% (24.3%; -21.2%) (Fig. 2).
Conclusions: Application of AIST to HHE and CBUM was feasible. For each study, at least one suitable clip was identified and measured. Acceptance of the automated view selection was high and automated reading showed good agreement between both methods.