Artificial neural network promising for predicting neonatal metabolic bone disease

Artificial neural network (ANN) may be an effective tool to predict neonatal metabolic bone disease (MBD) during prenatal and postnatal periods, according to a recent study.

MBD is often seen in babies born prematurely who lack fetal mineral accumulation. Hypophosphatemia, hyperphosphatemia, and skeletal demineralization are all hallmarks of MBD. Osteopenia, osteoporosis, rickets and pathological fractures can all also occur in MBD, increasing bone fragility and possibly affecting short- and long-term bone growth.

While infants at risk of MBD receive screening, there is a wide variety of diagnostic methods, and early recognition can be difficult due to the late onset of clinical symptoms and lack of specific biochemical markers. This has led to a need for predictive tools for MBD.

A flexible and accurate machine learning algorithm, the ANN, was able to predict various diseases. Researchers have developed serial ANN models for the risk of MBD to determine the most efficient and accurate model for prediction.

Pregnant Chinese women gave written consent to participate in a diagnostic study, conducted from January 1, 2012 to December 31, 2021. Recruitment occurred early in pregnancy, and follow-up continued until 1 month after delivery.

To be included, participants had to have a singleton pregnancy, complete clinical data during the antenatal, delivery and postpartum periods, and surviving infants with detailed values ​​of alkaline phosphatase. A peak serum alkaline phosphatase level above 500 U/L 72 hours after birth was used to determine MBD.

Data on maternal and neonatal characteristics were collected from electronic health records. These characteristics included demographic data and previous pregnancy history, nutritional conditions during pregnancy, complications and comorbidities, medication use during pregnancy, birth outcomes, and neonatal anomalies.

Five predictive models were built using an ANN, with a receiver operating characteristic curve used to evaluate the model performance.

There were 10,801 Chinese women who participated in the study, of which 65.8% were local residents, 98.1% were of Han ethnicity, and 9.3% had a uterine scan. Of the babies, 55.1% were male and 44.9% female. MBD occurred in 138 infants, of which only 6 were term infants. In comparison, 75.9% of control babies were term babies.

In putative predictors, the offspring of women with insufficient folic acid during pregnancy were 2.31 times more likely to develop MBD, compared to 3.26 times higher when born to women taking calcium supplements and 0.38 times lower when born to women taking taking iron supplements.

MBD risk factors include magnesium sulfate use during pregnancy and low birth weight infants, anemia, septicemia, or respiratory distress syndrome.

The area under the receiver operating characteristic curve (AUC) was calculated for ANN models. Model 1 showed the highest AUC, followed by model 5, model 4, model 3 and model 2. As model 1 had a net reclassification improvement of 0.205 compared to model 5, it showed an improved discriminative ability.

Results also indicated improved identification of neonates at high risk of MBD with fewer cases of misclassification in model 1. Overall model 1 showed the best performance for significant prenatal and postnatal factors and model 5 for postnatal factors, making them viable for use in clinical practice.

Reference

Jiang H, Guo J, Li J, Li C, Du W, Canavese F, et al. Artificial neural network modeling to predict neonatal metabolic bone disease in the prenatal and postnatal periods. JAMA Netw open. 2023;6(1):e2251849. doi:10.1001/jamanetworkopen.2022.51849