Obesity Prediction with EHR Data: A deep learning approach with interpretable elements.

Area of Science:

• Public Health
• Pediatrics
• Data Science

Background:

• Childhood obesity presents a significant public health concern.
• Early identification of at-risk children is crucial for timely interventions.
• Existing predictive tools often lack the ability to analyze longitudinal data patterns.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting future childhood obesity.
• To leverage electronic health records (EHR) for predictive modeling.
• To improve upon traditional methods by incorporating longitudinal data analysis.

Main Methods:

• Utilized a large, unaugmented electronic health records dataset from a US pediatric health system.
• Developed a Long Short-Term Memory (LSTM) network architecture incorporating an attention layer.
• Trained the model using both static and dynamic EHR data to predict obesity 1-3 years in advance for ages 3-20.

Main Results:

• The LSTM model demonstrated superior performance in predicting childhood obesity across various age ranges compared to existing literature.
• The attention layer provided interpretability by calculating attention scores for timestamps and ranking features.
• The model effectively predicted future obesity patterns using readily available EHR data.

Conclusions:

• Deep learning, specifically LSTM networks with attention, offers a powerful approach for predicting childhood obesity.
• The proposed model shows promise for early identification and intervention strategies.
• This method enhances predictive accuracy by utilizing longitudinal EHR data, outperforming traditional approaches.