TLDR: A new AI model uses a multi-view Variational Autoencoder (VAE) to predict pulmonary hypertension (PH) in newborns from echocardiographic videos. By integrating information from multiple heart views, the model improves diagnostic accuracy and generalization compared to single-view or traditional methods, offering a more objective and reliable assessment for this critical condition.
Pulmonary hypertension (PH) in newborns is a serious condition where there’s high pressure in the lung arteries, which can strain the baby’s heart and lead to heart failure. Early and accurate diagnosis is vital for treatment, but it’s often challenging because newborns have tiny anatomical structures and their cardiovascular systems are rapidly changing after birth.
Traditionally, right heart catheterization (RHC) is considered the most accurate diagnostic method, as it provides precise measurements of the disease’s severity. However, RHC is an invasive procedure with risks like bleeding and infection, making it impractical for routine screening in newborns. Echocardiography, a non-invasive imaging technique, is a preferred alternative because it’s safer and widely available. Despite its advantages, assessing PH using echocardiography often relies on a doctor’s visual interpretation of heart images, which can be subjective and lead to inconsistent diagnoses, potentially delaying crucial interventions.
To address these challenges, researchers Lucas Erlacher, Samuel Ruipérez-Campillo, Holger Michel, Sven Wellmann, Thomas M. Sutter, Ece Ozkan, and Julia E. Vogt from ETH Zurich and University Children’s Hospital Regensburg (KUNO) have developed a new approach. Their work, detailed in their paper “PREDICTING PULMONARY HYPERTENSION IN NEWBORNS : A M ULTI -VIEW VAE APPROACH”, introduces a multi-view variational autoencoder (VAE) framework for predicting PH using echocardiographic videos.
A Variational Autoencoder (VAE) is a type of artificial intelligence model that excels at learning complex patterns from high-dimensional data, like medical images. In this study, the VAE framework is designed to process multiple echocardiographic views simultaneously. This is important because different views of the heart provide complementary information, much like looking at an object from various angles gives a more complete picture. Existing automated methods often focus on single views or are designed for adults, which doesn’t translate well to the unique physiology of newborns.
The researchers utilized a dataset of 936 echocardiography videos from 192 newborns, collected from a single medical center. These videos captured five standard heart views: PLAX, A4C, PSAX-P, PSAX-S, and PSAX-A. The severity of PH was categorized into three levels: none, mild, and moderate-to-severe, based on expert assessment. The team preprocessed the videos, applied data augmentation techniques to enhance model robustness, and used class balancing to handle the uneven distribution of PH severity in the dataset.
Their multi-view VAE model, specifically the Multi-modal Variational Mixture Prior Model (MMVM-VAE), was trained to learn shared and view-specific features across the different echocardiographic views. This approach allows the model to capture more comprehensive latent representations of the heart’s condition. After this feature extraction, a separate classifier network was used to predict PH severity.
The results showed that incorporating multiple echocardiographic views within their MMVM-VAE framework significantly improved classification performance compared to using single-view inputs. The multi-view MMVM-VAE model provided more stable results, especially for predicting PH severity, achieving higher balanced accuracy and a strong AUROC (Area Under the Receiver Operating Characteristic curve), which are key metrics for evaluating diagnostic models. This multi-view framework also helped the model generalize better to new, unseen data, which is a common challenge in medical AI due to variations in patient populations and imaging protocols.
While promising, the study acknowledges some limitations. The videos were not acquired simultaneously, meaning slight physiological variations might exist between views. Also, the model assumes most standard views are available, which might not always be the case in real clinical settings where some views could be missing or of poor quality. Furthermore, the data came from a single medical center, so future work will involve evaluating the method on larger, multi-center cohorts to ensure broader applicability.
Also Read:
- Predicting Lung Disease in Preterm Infants: A Deep Learning Approach Using Early Chest X-rays
- Enhancing Heart Motion Tracking in Ultrasound with Bias-Reduced AI
In conclusion, this research represents a significant step towards developing more reliable and less operator-dependent methods for diagnosing pulmonary hypertension in newborns. By effectively combining information from multiple echocardiographic views, the multi-view VAE approach offers improved accuracy and generalization, paving the way for earlier and more consistent PH evaluation in neonates.
