Abstract P3089: Predicting Reductions In Hemoglobin A1c From Large Dietary Datasets Using Artificial Intelligence Pipeline With Interpretable Encoder-Decoder Deep Neural Networks

Introduction: Artificial intelligence algorithms can help understand and predict the complex interactions between dietary intake and health outcomes, especially from large datasets. Precision nutrition leverages machine learning techniques to analyze such patterns, but the methods and uses need deeper examination. Objective: Demonstrate the viability of fully unsupervised training of a β-Variational Autoencoder (β-VAE, a fully unsupervised and robust framework for capturing latent features from high-dimensional data) to predict improved hemoglobin A1c from complex population-level dietary data. Methods: Our β-VAE neural network architecture (Figure 1) was trained using 10 cycles of cross-sectional datasets from the US National Health and Nutrition Examination Survey from 1999 to 2018. Dietary data were collected with 24-hour recalls. A cumulative 102,957 participants and 5885 food/nutrient items were aggregated into 547 interpretable latent variables. Results: The trained network performance was evaluated through an aggregated multi-participant study case (n=1000) that explored the relationship between the target hemoglobin A1c change percentage and associated potential changes in participants' diet. Among the top five dietary changes required to achieve a 20% reduction in the latent space value corresponding to hemoglobin A1c were increasing coffee/tea and seafood intake and decreasing sweets, pizza, and salt intake (Figure 1). Conclusion: The prediction model showed a reasonable alignment with dietary recommendations despite being trained in a fully unsupervised way. β-VAE may be a first-in-class, efficient, and interpretable framework to analyze complex population-level dietary data for precision nutrition research.