Comparing innovative artificial intelligence algorithms to assess echocardiographic videos for clinical modeling

To use multiple dynamic deep learning algorithms to develop predictive models with video-based echocardiographic images using sample size determination as a key variable to assess optimal performance metrics.

Our study compares performance of 3-dimensional convolutional neural networks, video vision transformers, and hybrid convolutional neural networks and long short-term memory models within both supervised learning and semi-supervised learning (SSL) domains using variable sample sizes.

For supervised learning, the ResNet3D model achieved the lowest mean absolute error (MAE) and root mean squared error (RMSE) across all training set sizes (200-, 400-, and 800-video datasets), with the best performance observed on the 800-video training set (MAE = 7.409, RMSE = 10.216). In the SSL setting, both the ResNet3D and ResNet+LSTM models benefited from the inclusion of unlabeled data, particularly with larger data sets.

Because SSL models use both labeled and unlabeled data sets, our findings are significant in showing that performance of certain predictive models using mixtures of unlabeled and labeled data is comparable to that of models using only labeled data with similar sample sizes, thus obviating the need for large sample sizes of labeled data.