Abstract
Childhood obesity is a major public health challenge. Early prediction and identification of the children at a high risk of developing childhood obesity may help in engaging earlier and more effective interventions to prevent and manage obesity. Most existing predictive tools for childhood obesity primarily rely on traditional regression-type methods using only a few hand-picked features and without exploiting longitudinal patterns of children data. Deep learning methods allow the use of high-dimensional longitudinal datasets. In this paper, we present a deep learning model designed for predicting future obesity patterns from generally available items on children medical history. To do this, we use a large unaugmented electronic health records dataset from a large pediatric health system. We adopt a general LSTM network architecture which are known to better represent the longitudinal data. We train our proposed model on both dynamic and static EHR data. Our model is used to predict obesity for ages between 2-20 years. We compared the performance of our LSTM model with other machine learning methods that aggregate over sequential data and ignore temporality. To add interpretability, we have additionally included an attention layer to calculate the attention scores for the timestamps and rank features of each timestamp.
Introduction
Childhood obesity is a complex public health crisis with long-term implications for physical and mental health. Despite its significance, most existing predictive tools are surprisingly simplistic, often relying on traditional regression models that utilize only a few hand-picked variables. Crucially, these methods frequently ignore the temporal trajectory of a child's health, which contains vital information about weight velocity and developmental patterns.
Electronic Health Record (EHR) data presents a rich but challenging resource for predictive modeling. It is high-dimensional, highly sparse (as most patients only have a small subset of thousands of possible clinical codes), and irregularly sampled. This study addresses these challenges by utilizing deep learning to capture longitudinal patterns in "unaugmented" EHR data—information already present in hospital databases without requiring specialized lifestyle or genetic studies.
Methods
The researchers proposed a hybrid neural network architecture designed to integrate both dynamic (time-series) and static (demographic) data. The core of the temporal modeling is a Long Short-Term Memory (LSTM) network, which is well-suited for processing sequences of clinical visits. To handle the irregular timing of medical encounters, visits were aggregated into 30-day "bins" over a two-year observation window.
Key components of the architecture include:
- Feature Embedding: An embedding layer reduces the massive space of medical codes (conditions, medications, procedures, and labs) into a dense, manageable format.
- Interpretability via Attention: To move beyond "black box" predictions, a softmax attention layer was added to calculate "attention scores" for each timestamp. This allows the model to identify which specific visits were most critical to the final risk score. Embedding weights were further used to rank individual features within those visits.
- Static Integration: A separate feed-forward network processes non-changing data, such as sex and race, merging its output with the LSTM’s temporal representation before the final prediction.
- Transfer Learning: To combat data scarcity in older age groups, general models were trained on the entire dataset before being fine-tuned for 48 specific age-based sub-cohorts.
Dataset
The study utilized a massive dataset from the Nemours Children Health System, part of the PEDSnet network. The dataset spans from 2002 to the present and includes 1,850 variables standardized using terminologies like SNOMED-CT, RxNorm, and LOINC.
As shown in the table above, the dataset includes over 68,000 unique patients with an average of 51 visits per patient, providing a robust foundation for training deep longitudinal models.
Results and Discussion
The LSTM model consistently outperformed traditional baselines, such as Linear Regression and Random Forest, across almost all age groups and prediction windows. The model achieved an AUC of 0.80 for predicting obesity at age 5 (preschool), 0.93 at age 11 (prepubertal), and 0.92 at age 18 (post-pubertal) for a three-year prediction window.
Several key insights emerged from the analysis:
- Temporality is Critical: The study confirms that the pattern and velocity of weight gain are often more predictive than weight at a single point in time. LSTMs naturally capture this "momentum" better than static models.
- Interpretable Deep Learning: The attention mechanism successfully identified the most informative visits. While top-ranked features like BMI percentiles aligned with clinical knowledge, the model also flagged less obvious correlations, such as certain chromosomal anomalies and hearing loss, as relevant predictors.
- Early Intervention Potential: The model was particularly effective at identifying children transitioning from "non-obese" to "obese," highlighting early predictors like prematurity and chronic tonsillitis.
Conclusion
This research demonstrates that deep learning models with attention mechanisms can provide both high accuracy and clinical interpretability for childhood obesity prediction. While the current model focuses on a three-year horizon and treats medical codes as independent features, future work will explore longer-term predictions and feature grouping to reduce noise. Testing the model across diverse healthcare systems remains a necessary step to ensure generalizability and clinical utility in varied pediatric populations.
Summary updated