EchoMD Automated Ejection Fraction Software

The EchoMD Automated Ejection Fraction Software employs machine learning, specifically a convolutional neural network, to process transthoracic cardiac ultrasound images and provide automated estimation of left ventricular ejection fraction for adult cardiac evaluations. This system aims to streamline clinical workflows by automatically selecting image clips and calculating LVEF without requiring endocardial tracing, and its performance has been validated with a defined accuracy against established clinical methods.

The EchoMD Automated Ejection Fraction Software is designed to assist clinicians in cardiac evaluation by providing automated estimation of left ventricular ejection fraction (LVEF) in adult patients. The software processes previously acquired transthoracic cardiac ultrasound images, specifically apical four-chamber and apical two-chamber views, using a personal computer or a compatible DICOM-compliant PACS system. Its methodology involves machine learning algorithms that automatically select relevant image clips and calculate LVEF, streamlining the workflow for clinicians. The software operates as a microservice, integrating between the DICOM source and a PACS destination, where the resulting EF percentage is displayed, and also applies machine learning to assess image quality, offering user feedback. Scientifically, the EchoMD software leverages a convolutional neural network (CNN) to analyze echocardiographic images and estimate LVEF without generating endocardial traces, a distinct approach from other systems. Performance was validated through a formal retrospective, non-interventional study using over 300 previously-acquired patient studies, diverse in gender, ejection fraction values, body mass index, and acquired from multiple ultrasound scanner platforms, with test datasets strictly segregated from algorithm training data. The software demonstrated its accuracy by achieving a root mean square deviation of 8.290% (p-value 0.00052) when its LVEF measurements were compared to those derived from the biplane method of disks, thereby meeting its primary endpoint. An additional study with different cardiologists on a subset of these studies further supported the generalizability of the software.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the convolutional neural network's performance in LVEF estimation without endocardial tracing compare to algorithms that explicitly generate and use endocardial traces, particularly in cases with poor image quality or atypical ventricular geometry?; What specific features or criteria does the machine learning algorithm use for automated image clip selection, and how robust is this selection process across varying levels of operator skill in image acquisition?; Given the Root Mean Square Deviation of 8.290%, what is the clinical significance of this deviation range for LVEF, and how does it compare to inter-observer variability among expert cardiologists using traditional methods?; How does the software's automated image quality assessment influence the LVEF calculation, and are there thresholds below which the software will not provide an estimate or will flag it as potentially unreliable?; What were the characteristics of the algorithm training datasets, and how does their composition (e.g., patient populations, image quality spectrum) potentially influence the software's generalizability to new, unseen patient data?