M-TabNet: A Multi-Encoder Transformer Model for Predicting Neonatal Birth Weight from Multimodal Data

TL;DR

Birth weight (BW) prediction in early gestation is challenged by ultrasound limitations. The authors introduce M-TabNet, a transformer-based multi-encoder that fuses physiological, nutritional, lifestyle, and genetic maternal data to predict BW before 12 weeks. On the Reus–Tarragona cohort, M-TabNet achieves $MAE = 122$ g and $R^2 = 0.94$, with high LBW classification performance ($\text{sensitivity}=97.55\%$, $\text{specificity}=94.48\%$); generalization to the IEEE dataset yields $MAE = 105$ g and $R^2 = 0.95$. SHAP and feature-importance analyses highlight maternal age, tobacco exposure, and vitamin B12 status as key drivers, supporting interpretable, personalized prenatal care and early risk stratification for at-risk pregnancies.

Abstract

Birth weight (BW) is a key indicator of neonatal health, with low birth weight (LBW) linked to increased mortality and morbidity. Early prediction of BW enables timely interventions; however, current methods like ultrasonography have limitations, including reduced accuracy before 20 weeks and operator dependent variability. Existing models often neglect nutritional and genetic influences, focusing mainly on physiological and lifestyle factors. This study presents an attention-based transformer model with a multi-encoder architecture for early (less than 12 weeks of gestation) BW prediction. Our model effectively integrates diverse maternal data such as physiological, lifestyle, nutritional, and genetic, addressing limitations seen in prior attention-based models such as TabNet. The model achieves a Mean Absolute Error (MAE) of 122 grams and an R-squared value of 0.94, demonstrating high predictive accuracy and interoperability with our in-house private dataset. Independent validation confirms generalizability (MAE: 105 grams, R-squared: 0.95) with the IEEE children dataset. To enhance clinical utility, predicted BW is classified into low and normal categories, achieving a sensitivity of 97.55% and a specificity of 94.48%, facilitating early risk stratification. Model interpretability is reinforced through feature importance and SHAP analyses, highlighting significant influences of maternal age, tobacco exposure, and vitamin B12 status, with genetic factors playing a secondary role. Our results emphasize the potential of advanced deep-learning models to improve early BW prediction, offering clinicians a robust, interpretable, and personalized tool for identifying pregnancies at risk and optimizing neonatal outcomes.

Figures (6)

• Figure 1: Distribution of BWs in the original data and after applying SMOGN to the Reus-Tarragona Birth Cohort and IEEE Child Birth Dataset.
• Figure 2: Proposed Transformer-based Multi-encoder Architecture for Early Neonatal BW Prediction. The architecture integrates multiple encoders to process maternal nutritional, lifestyle, genetic, and health insights, facilitating more accurate prediction of BW and ultimately improving neonatal outcomes.
• Figure 3: Model Performance Evaluation: (a) Predicted vs. Actual BW across the R Line: The scatter plot shows the relationship between predicted and actual BW, with the blue line representing the linear regression fit. The R² value of 0.94 indicates a strong correlation between the predictions and actual values. (b) MAE during 5-fold cross-validation: The plot illustrates the changes in MAE across each fold, with the orange line highlighting the variation in performance as the model is trained and evaluated on different subsets of the data.
• Figure 4: Feature importance scores for neonatal BW prediction, highlighting age, tobacco use, and sun exposure as dominant factors. Anemia and the MTHFD1 105TC variant show minimal influence.
• Figure 5: Sensitivity analysis of maternal factors on neonatal BW shows age, tobacco use, and folate as key influencers, while anemia and MTHFD1 105TC had minimal impact.