Home > Cardiology > AHA 2021 > Other > Accurate ejection fraction assessment in paediatric patients via artificial intelligence

Accurate ejection fraction assessment in paediatric patients via artificial intelligence

Presented by
Dr Charitha Reddy, Stanford Children’s Health, USA
AHA 2021
A video-based, deep-learning model for the automated assessment of ejection fractions (EF) in paediatric cardiac patients neared human accuracy. Notably, the model delivered different results when trained on adults or paediatric patients, demonstrating the significant differences in EF assessment between these populations. Future efforts should be directed at clinical implementation and validation of the model in a broad paediatric population.

“Left ventricular function is assessed for diagnosis, screening, and treatment management in all of our paediatric patients,” said Dr Charitha Reddy (Stanford Children’s Health, CA, USA) [1]. She explained that human assessment of EF measurement is limited. “Since the guideline-recommended 3 separate cardiac cycles, that are needed to average the EF, are rarely measured in clinical practice due to time constraints, we need an alternative.” According to Dr Reddy, deep-learning models have been developed to assess EF in adults but not in paediatric patients. “EF evaluation is different in young patients, due to increased frame rates, increased heart rates, and a wider range of body surface area in these patients,” argued Dr Reddy.

The EchoNet-Dynamic, video-based, deep-learning model has been developed to assess EF in adult patients [2]. The current study investigated the assessment of EF in paediatric patients via this model. In total, 4,400 ECGs were collected from 1,923 cardiac patients under 18 years of age with structurally normal hearts. The training phase used 80% of the ECG, whereas 10% was utilised for validation, and 10% for testing. The input for the model consisted of apical 4 chamber (A4C) videos, parasternal short axis view (PSAX) videos, and a combination of both.

The EchoNet-Dynamic model showed excellent overlap with human measurement. The authors used a dice similarity coefficient to measure the similarity of 2 data sets, with values ranging from 0, no overlap, to 1, complete overlap. In this study, the values were 0.901 for A4C videos and 0.887 for PSAX videos. The model showed an R^2 of 0.78 in the prediction of EF for the combined input (A4C plus PSAX videos) in paediatric patients if the model used a paediatric dataset in the training phase. When the model used a paediatric dataset in assessing the EF in adult patients, it resulted in an R^2 of 0.33. This reflects the distinction between assessing paediatric patients and adult patients. Notably, the model could estimate an EF 5/6 area length based on the input of only A4C or PSAX video.

Dr Reddy argued that the model should be validated in a broader paediatric dataset. Moreover, to protect the privacy of patients when applying deep-learning models in clinical practice, the utilisation of edge servers and federated models should be addressed.


    1. He B, et al. Video-Based Deep Learning Model for Automated Assessment of Ejection Fraction in Pediatric Patients. CH.AOS.466, AHA 2021 Scientific Sessions, 13–15 November.
    2. Ouyang D, et al. Nature. 2020;580:252–256.


Copyright ©2021 Medicom Medical Publishers

Posted on