Abstract
Objective: Identifying children at high risk of developing obesity can offer a critical time to change the course of the disease before it establishes. Numerous studies have tried to achieve this; but practical limitations remain, including (i) relying on data not present in routinely available pediatric data (like prenatal data), (ii) focusing on a single age prediction (hence, not tested across ages), and (iii) not achieving good results or adequately validating those. Methods: A customized sequential deep learning model was built to predict the risk of childhood obesity, focusing especially on capturing the temporal patterns. The model was trained only on routinely collected EHRs, containing a list of features identified by a group of clinical experts, and sourced from 36,191 diverse children aged 0 to 10. The model was evaluated using extensive discrimination, calibration, and utility analysis; and was validated temporally, geographically, and across various subgroups. Results: Our results are mostly better (and never worse) than all previous studies, including those that focus on single-age predictions or link EHRs to external data. Specifically, the model consistently achieved an area under the curve (AUROC) of above 0.8 (with most cases around 0.9) for predicting obesity within the next 3 years for children 2 to 7. The validation results show the robustness of the model. Furthermore, the most influential predictors of the model match important risk factors of obesity. Conclusions: Our model is able to predict the risk of obesity for young children using only routinely collected EHR data, greatly facilitating its integration with the periodicity schedule. The model can serve as an objective screening tool to inform prevention efforts, especially by helping with very delicate interactions between providers and families in primary care settings.
Introduction
Childhood obesity is a major public health challenge, yet pediatricians often fail to identify at-risk children until they have already crossed the threshold into obesity. While several predictive models exist, they often suffer from significant "real-world" friction points: they may require data not typically found in standard Electronic Health Records (EHRs) (such as prenatal history or genetic markers), they are often rigid and designed for a single specific age, and they frequently lack rigorous validation across different geographic regions or demographic subgroups.
This research demonstrates that "routinely collected" EHR data is sufficient for reliable, long-term risk assessment. By utilizing sequential deep learning, the authors provide a flexible screening window that allows clinicians to identify at-risk children years before obesity is established.
Methodology
Dataset and Cohort
The study utilized a massive longitudinal dataset from Nemours Children’s Health, a large pediatric network in the US. The cohort included 36,191 diverse children (aged 0–10) with over 44 million total encounters between 2002 and 2019. The data spanned five states (Delaware, Florida, Maryland, New Jersey, and Pennsylvania), allowing for robust geographic validation. Children with conditions significantly impacting weight gain, such as Type 1 diabetes or cancer, were excluded to ensure the model targeted the general pediatric population.
Sequential Deep Learning Architecture
The researchers employed an Encoder-Decoder model designed to handle longitudinal data:
- Encoder: Uses Long Short-Term Memory (LSTM) cells to capture temporal patterns and trajectories in a child's health status over time.
- Decoder: A standard feed-forward network that generates the final risk probability.
- Flexible Window Design: Unlike "one-shot" models, this approach can take any amount of available data (from birth to age 7) and predict the risk of obesity for the subsequent 1, 2, or 3 years.
Feature Engineering and Social Determinants
Raw EHR data was transformed into 506 binary features, including quintile binning for numeric measurements and categorization of 138 diagnoses and 84 medication groups. Additionally, the model incorporated the Child Opportunity Index (COI) by geolocating patient addresses, integrating neighborhood-level socioeconomic data into the model without requiring manual entry from clinicians.
Results
High Discriminative Power
The model consistently achieved an AUROC (Area Under the Receiver Operating Characteristic) above 0.8, with many scenarios reaching 0.9. In a direct comparison, when predicting obesity at age 5 using data from ages 0–2, the model achieved an AUROC of 0.81, significantly outperforming the standard clinical baseline of 0.71.
Key Insights
- Temporal Patterns Matter: The model's success suggests that the trajectory of health data is more predictive than any single measurement.
- EHR Sufficiency: One of the most significant findings is that routinely collected data contains enough "proxy" signals to match or exceed the performance of models using high-effort data like parental BMI or infant dietary habits.
- Early Infancy Indicators: The "Weight-for-Length (WFL)% change" before age 2 was among the top 10 most influential predictors, reinforcing the importance of the first 1,000 days of life.
- Robustness and Fairness: Performance remained stable across different races, ethnicities, and insurance types. The model also held up under geographic and temporal validation (testing on different states and time periods).
Discussion and Conclusion
While the study is limited by its focus on a single healthcare system and potential sample bias due to missing early measurements, it represents a major step forward in pediatric preventative care. The team is currently working on a SMART on FHIR tool to integrate this model directly into clinical workflows.
By turning routinely collected data into real-time decision support, this model can serve as an objective screening tool. It helps facilitate delicate interactions between providers and families, allowing for early intervention and potentially changing the course of the disease before it establishes.
Summary updated