M-TabNet: A Transformer-Based Multi-Encoder for Early Neonatal Birth Weight Prediction Using Multimodal Data.

Area of Science:

• Maternal-fetal medicine
• Artificial intelligence in healthcare
• Neonatal health

Background:

• Birth weight (BW) is crucial for neonatal health; low birth weight (LBW) increases mortality/morbidity.
• Early BW prediction aids in preventing impaired fetal growth, but current methods like ultrasonography have limitations.
• Existing models often overlook nutritional and genetic factors, focusing primarily on physiological and lifestyle aspects.

Purpose of the Study:

• To develop an accurate early (<12 weeks gestation) birth weight prediction model.
• To integrate diverse maternal data, including physiological, lifestyle, nutritional, and genetic factors.
• To address limitations of existing models and enhance clinical utility for risk stratification.

Main Methods:

• An attention-based transformer model with a multi-encoder architecture was employed.
• The model integrated diverse maternal data: physiological, lifestyle, nutritional, and genetic.
• Feature importance and SHAP analysis were used for model interpretability.

Main Results:

• The model achieved high predictive accuracy with a Mean Absolute Error (MAE) of 122 grams and R² of 0.94 on an in-house dataset.
• Independent validation demonstrated generalizability with MAE of 105 grams and R² of 0.95.
• Classification into low and normal BW categories yielded 97.55% sensitivity and 94.48% specificity.

Conclusions:

• Advanced deep learning models can significantly improve early birth weight prediction.
• The developed model offers a robust, interpretable, and personalized tool for identifying high-risk pregnancies.
• This approach facilitates timely intervention and optimization of neonatal outcomes.