Development and preliminary evaluation of a machine learning model for predicting low birth weight using placental IVIM-MRI and maternal clinical characteristics

Abstract

Objective: To identify key placental intravoxel incoherent motion (IVIM) MRI parameters and maternal factors associated with low birth weight (LBW), and develop a prenatal predictive model for LBW risk assessment.

Methods: This retrospective study analyzed 113 term neonates (January 2023-December 2024), categorized as LBW or normal birth weight. Twenty-one antenatal metrics, including maternal characteristics and region-specific placental IVIM MRI parameters (perfusion fraction [f], true diffusion coefficient [D], pseudo-diffusion coefficient [D*]), were evaluated. Feature importance was ranked using Shapley Additive Explanations (SHAP) analysis in a Random Forest algorithm. Key predictors were used to build a multivariable logistic regression nomogram. Discrimination (ROC-AUC), calibration, and clinical utility (DCA) were assessed. Internal validation employed bootstrap resampling (1000 iterations).

Results: SHAP identified f values from maximal placental section (f_MPS), central zone (f_CPZ), and fetal side (f_FS) as top predictors. The nomogram showed good discrimination (AUC = 0.86, 95 % CI: 0.74-0.98). Bootstrap validation yielded an AUC of 0.82 (95 % CI: 0.61-0.98), with high sensitivity and specificity. The calibration curve showed good model fit. DCA demonstrated considerable clinical benefit.

Conclusion: Placental IVIM MRI f values from distinct placental regions are significant LBW predictors. The model provides accurate prenatal risk assessment, guiding early interventions to optimize perinatal outcomes.

Keywords: Intravoxel incoherent motion (IVIM) MRI; Low birth weight (LBW); Machine learning; Nomogram; Placenta.