Deep Learning Model for Predicting Neurodevelopmental Outcome in Very Preterm Infants Using Cerebral Ultrasound.

Area of Science:

• Neonatal neurology
• Medical imaging analysis
• Machine learning in healthcare

Background:

• Very preterm infants (VPIs) face risks of neurodevelopmental impairment (NDI).
• Accurate prediction of NDI is crucial for timely intervention.
• Cranial ultrasound (CUS) and clinical data are potential predictors.

Purpose of the Study:

• To develop and validate deep learning (DL) models for predicting NDI in VPIs.
• To integrate CUS images and clinical variables for enhanced predictive accuracy.
• To assess the performance of DL models compared to traditional methods.

Main Methods:

• Retrospective cohort study of VPIs (22-30 weeks' gestation).
• Development of DL models using elastic net (EN) and convolutional neural networks (CNN).
• Utilized CUS images and clinical data at hospital discharge and 3 time points.

Main Results:

• A CNN model combining CUS and clinical variables showed superior performance (PR-AUC: 0.75) compared to clinical predictors alone (PR-AUC: 0.60).
• The anterior coronal plane at 6 weeks yielded the highest predictive accuracy (PR-AUC: 0.81).
• Models were validated using PR-AUC and ROC-AUC metrics.

Conclusions:

• A DL prognostic model integrating CUS and clinical data effectively predicts NDI in VPIs.
• Early and accurate NDI risk identification facilitates targeted interventions.
• This predictive model can improve functional outcomes for VPIs.