Application of the deep learning algorithm in nutrition research - using serum pyridoxal 5'-phosphate as an example.

BACKGROUND: Multivariable linear regression (MLR) models were previously used to predict serum pyridoxal 5'-phosphate (PLP) concentration, the active coenzyme form of vitamin B6, but with low predictability. We developed a deep learning algorithm (DLA) to predict serum PLP based on dietary intake, dietary supplements, and other potential predictors. METHODS: This cross-sectional analysis included 3778 participants aged ≥20 years in the National Health and Nutrition Examination Survey (NHANES) 2007-2010, with completed information on studied variables. Dietary intake and supplement use were assessed with two 24-hour dietary recalls. We included potential predictors for serum PLP concentration in the models, including dietary intake and supplement use, sociodemographic variables (age, sex, race-ethnicity, income, and education), lifestyle variables (smoking status and physical activity level), body mass index, medication use, blood pressure, blood lipids, glucose, and C-reactive protein. We used a 4-hidden-layer deep neural network to predict PLP concentration, with 3401 (90%) participants for training and 377 (10%) participants for test using random sampling. We obtained outputs after sending the features of the training set and conducting forward propagation. We then constructed a loss function based on the distances between outputs and labels and optimized it to find good parameters to fit the training set. We also developed a prediction model using MLR. RESULTS: After training for 105 steps with the Adam optimization method, the highest R2 was 0.47 for the DLA and 0.18 for the MLR model in the test dataset. Similar results were observed in the sensitivity analyses after we excluded supplement-users or included only variables identified by stepwise regression models. CONCLUSIONS: DLA achieved superior performance in predicting serum PLP concentration, relative to the traditional MLR model, using a nationally representative sample. As preliminary data analyses, the current study shed light on the use of DLA to understand a modifiable lifestyle factor.

A homotopy training algorithm for fully connected neural networks.

In this paper, we present a homotopy training algorithm (HTA) to solve optimization problems arising from fully connected neural networks with complicated structures. The HTA dynamically builds the neural network starting from a simplified version and ending with the fully connected network via adding layers and nodes adaptively. Therefore, the corresponding optimization problem is easy to solve at the beginning and connects to the original model via a continuous path guided by the HTA, which provides a high probability of obtaining a global minimum. By gradually increasing the complexity of the model along the continuous path, the HTA provides a rather good solution to the original loss function. This is confirmed by various numerical results including VGG models on CIFAR-10. For example, on the VGG13 model with batch normalization, HTA reduces the error rate by 11.86% on the test dataset compared with the traditional method. Moreover, the HTA also allows us to find the optimal structure for a fully connected neural network by building the neutral network adaptively.