Deep Learning Based Prediction of Pulmonary Hypertension in Newborns Using Echocardiograms

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Pulmonary hypertension (PH) in newborns is a critical condition requiring early detection and severity classification for effective management.
• Echocardiography is the primary diagnostic tool, but manual assessment is time-consuming and requires specialized expertise.
• Existing automated PH detection methods are limited, particularly for pediatric populations and severity classification.

Purpose of the Study:

• To develop and validate an explainable, multi-view video-based deep learning approach for predicting and classifying the severity of PH in newborns.
• To address the need for automated, accurate, and efficient PH assessment in pediatric echocardiography.

Main Methods:

• Utilized spatio-temporal convolutional neural networks for PH prediction from individual echocardiogram views.
• Employed a majority voting strategy to aggregate predictions from multiple views.
• Applied explainability techniques (saliency maps) to validate model focus on relevant cardiac structures.
• Evaluated the model on a cohort of 270 newborns using 10-fold cross-validation and a held-out test set.

Main Results:

• Achieved a mean F1-score of 0.84 for severity prediction and 0.92 for binary PH detection via cross-validation.
• Demonstrated performance on the held-out test set with an F1-score of 0.63 for severity and 0.78 for binary detection.
• Saliency maps confirmed the model's focus on clinically significant cardiac regions.

Conclusions:

• The developed deep learning model offers a promising automated solution for predicting and classifying PH severity in newborns using echocardiograms.
• This explainable AI approach can potentially assist clinicians in early and accurate PH diagnosis, improving patient outcomes.
• This represents the first automated assessment of PH in newborns utilizing echocardiographic data.