Abstract
Childhood obesity is a critical public health issue with long-term health implications. Predicting obesity early allows for timely intervention, yet traditional models often fail to capture the complex temporal patterns in a child's medical history. This paper introduces a deep learning framework that utilizes longitudinal Electronic Health Record (EHR) data to predict obesity up to three years in advance. By integrating Long Short-Term Memory (LSTM) networks with an attention mechanism, the model achieves high predictive performance (AUC up to 0.93) and offers interpretability by identifying key clinical visits and medical features driving the risk.
Introduction
Predicting childhood obesity accurately is a complex task. Most existing clinical tools rely on traditional regression models that utilize a limited set of static features, such as birth weight or maternal BMI. These approaches suffer from two primary limitations: they ignore the temporal "story" of a child's growth trajectory by aggregating data into simple averages, and they often function as "black boxes" that lack the transparency required for clinical trust.
To address these gaps, this research proposes a model that leverages the full history of a child's medical visits while remaining transparent enough for clinical application.
Methods
The researchers developed a dual-stream architecture designed to handle the inherent complexity and irregularity of EHR data.
Data Preprocessing and Aggregation
EHR data is typically irregularly sampled, as clinical visits occur at random intervals. To standardize this input for a sequential model, the researchers aggregated clinical data into 30-day windows over a 2-year observation period.
Model Architecture
The proposed framework consists of two main processing streams:
- Dynamic Stream (LSTM): A Long Short-Term Memory network processes sequences of medical events, including conditions, medications, procedures, and lab results, captured within the 30-day buckets.
- Static Stream (Feed-Forward): Non-temporal demographic data (sex, race, ethnicity, and zip code) is processed through a separate feed-forward network and subsequently merged with the dynamic stream's output.
An Interpretability Layer using an attention mechanism was integrated to calculate "importance scores." This allows the model to rank specific timestamps (which visit was most influential) and specific medical features (which clinical code drove the risk) for each prediction. Additionally, transfer learning was employed to address data sparsity in older age groups by pre-training on the full dataset before fine-tuning on specific 5-year age cohorts.
Datasets
The study utilized a massive dataset from the Nemours Children Health System, part of the PEDSnet network. The dataset includes 44 million records for 68,003 unique patients, spanning from 2002 to the present. The raw data encompasses over 20,000 conditions, 10,000 procedures, and 7,000 lab variables. To manage high dimensionality, variables present in less than 2% of the population were filtered out, resulting in a final feature set of 1,737 variables.
Key Results
The model was evaluated on its ability to predict obesity 1, 2, and 3 years into the future across age ranges from 2 to 20 years.
- Predictive Accuracy: The LSTM-based model achieved an AUC of 0.80 for preschoolers (age 5), 0.93 for prepubertal children (age 11), and 0.92 for post-pubertal adolescents (age 18) for a 3-year prediction horizon.
- Performance Gains: The proposed model consistently outperformed Random Forest and Linear Regression baselines across nearly all age cohorts and prediction windows.
- Interpretability: The attention mechanism successfully identified critical features. In validation cases, the model correctly highlighted "BMI Percentile" and "Previous Obesity Label" from recent visits as the primary risk drivers.
Discussion and Future Work
The findings underscore that the pattern of weight gain is often more predictive than absolute weight at a single point in time. The model also identified non-obvious predictors, such as hearing loss and hyperactive behavior, which warrant further clinical investigation.
Despite its success, the study notes several limitations. The current prediction horizon is capped at three years, and the model treats every medical code as a distinct feature rather than grouping related clinical concepts. Future work will focus on external validation across different health systems and exploring semi-supervised learning techniques to incorporate data from patients with shorter medical histories.
Summary updated