®®®® SIIA Público

Título del libro:
Título del capítulo: Early Birth Weight Prediction: A Machine Learning Explainability Analysis

Autores UNAM:
ERIK MOLINO MINERO RE; FERNANDO ARAMBULA COSIO; NIDIYARE HEVIA MONTIEL;
Autores externos:

Idioma:

Año de publicación:
2025
Palabras clave:

Early birth weight prediction; Machine learning; SHAP values


Resumen:

Fetal weight is a crucial health indicator during gestation, influencing outcomes and potentially preventing fetal death. Traditional clinical methods, like the Hadlock or Shepard formulas, rely on fetal biometry via ultrasound but have high error rates, up to 25%. This study emphasizes the use of machine learning techniques for early birth weight estimation while focusing on model interpretability through Shapley additive explanations (SHAP) values. SHAP value-based explainability charts allow experts to understand how different features influence predictions, enhancing the transparency and accessibility of these models. The dataset used consists of 578 participants from the National Institute of Perinatology of Mexico, with 18 maternal-fetal variables captured during the first trimester of pregnancy. Data was split into training, testing, and validation sets (80%, 10%, 10%), and five-fold cross-validation was applied. While various supervised learning methods were explored, this study highlights the use of SHAP values to identify key features influencing fetal weight, ensuring local accuracy, fair feature attribution, and consistent insights. Given the uniqueness of the dataset, comparisons with existing models in the literature are not feasible, and the focus remains on model explainability rather than performance benchmarking. By providing clearer insights into machine learning models, this work aids medical professionals in applying AI-driven predictions with greater confidence and understanding.


Entidades citadas de la UNAM: