It has been well established that early intervention could have a profound impact on overweight and obesity later in life (24–26). The earlier the intervention the bigger the impact on the child and his family’s lifestyle (27). Many early intervention programs attempt to minimize later overweight and obesity (24, 28, 29). Some of the programs encourage and support breastfeeding over supplementation with formulas (30, 31). Assessing newborn at risk may initiate intervention at a critical period immediately post birth when nutrition habits are determined.
Since early intervention programs might incur high costs (32, 33), better identification of newborn at risk is crucial for allocating available resources (34). However, possible predicting models for identifying newborn at risk were not evaluated so far.
In this study, we assessed three possible models based on growth measurements – LGA, macrosomia and WHO weight-for-length growth percentiles.
Although these models’ AUC is modest (0.588–0.653), and only up to 17% of the children selected based on these models for intervention post birth will eventually become overweight (above 97th percentile at age 18–36 months), this model's prediction is 3-time better than random sampling. Furthermore, 50% of the children selected based on these models for intervention post birth will eventually be at risk for overweight (over 85th percentile at age 18–36 months). Thus, we believe that using these predicting models to detect newborns at risk for overweigh could be beneficial for better directing resources for intervention programs.
There weren't any significant differences using macrosomia or LGA as predicting models, however, using WHO percentiles at birth yield less effective model.
In this study we further attempted to build machine learning model based on many other parameters available in our database including demographic background and pregnancy and birth parameters to improve screening performance. However, this machine learning model did not significantly improve the prediction results.
Our study shows the importance of using macrosomia or LGA as predicting models for identifying neonates at risk for overweight and obesity. In this study our machine learning model did not improve the prediction results of the current simple models available.
The study has several notable strengths. Firstly, it benefits from a substantial cohort of over 200,000 children born within a specific time frame, providing a rich and diverse dataset that closely mirrors the real-world population under investigation. This large and representative dataset enhances the external validity of the study's findings, making them more applicable to the broader population. Secondly, the study offers a comprehensive assessment of multiple prediction models, including the LGA model, Macrosomia model, and WHO Child Growth Standards, as well as a machine learning model. This thorough approach ensures a well-rounded evaluation of available methods for predicting childhood overweight. Furthermore, the study underscores the critical importance of early intervention in addressing childhood overweight and obesity. It emphasizes the potential benefits of identifying at-risk newborns and infants for targeted interventions during the early stages of life, aligning with established principles of prevention and health promotion. Lastly, by comparing and contrasting various prediction models, the study provides valuable insights into which methods may be more effective in identifying newborns at risk. This information can guide healthcare organizations and providers in making informed decisions regarding the selection of appropriate screening tools.
On the other hand, the study is not without limitations. The study acknowledges the limited predictive power of the examined prediction models, whether existing or machine learning-based. Modest area under the curve (AUC) values suggest that a significant number of false positives and false negatives may occur, which could impact the effectiveness of early interventions.While the study discusses the potential benefits of identifying newborns at risk for overweight, it does not delve into the practical implications of resource allocation for intervention programs. Determining the cost-effectiveness and feasibility of implementing such interventions is crucial for healthcare decision-makers.Additionally, the study's findings are primarily applicable to the Israeli population. Although, the Israeli population is divers and multicultural, the direct relevance to other regions or countries with distinct healthcare systems, demographics, and healthcare practices may be limited.
Lastly, the machine learning model's performance did not significantly outperform existing methods, however our multivariable model lies in the absence of crucial parameters. Specifically, important variables such as the mother's weight prior to pregnancy, pregnancy weight gain, gestational diabetes, and smoking status were not available in our database. These factors have been established in the literature as potentially relevant for predicting childhood overweight and obesity. The absence of these variables in our analysis may limit the model's ability to comprehensively account for all pertinent risk factors and could impact the accuracy of predictions related to childhood overweight. (35–39)
Recognizing the significance of early detection and intervention, future research endeavors that incorporate essential parameters could potentially enhance the predictive capabilities of the current model.