Microbiome-based correction for random errors in nutrient profiles derived from self-reported dietary assessments

Accurate dietary intake assessment is crucial for understanding the link between diet and chronic diseases. Epidemiological studies commonly rely on self-reported instruments like food frequency questionnaires (FFQs), 24-hour recalls, and diet records. However, these methods are susceptible to both random (day-to-day variations) and systematic (under/over-reporting, portion size inaccuracies) errors, leading to inaccurate nutrient profiles. Existing correction methods primarily target systematic errors in habitual dietary intake, leaving a gap in addressing random errors in single-day assessments. This study addresses this gap by developing a novel computational method to correct for random errors in self-reported dietary data. The approach leverages the gut microbiome's relationship with diet, recognizing that gut microbial composition is influenced by the types and quantities of nutrients consumed. This connection offers a potential pathway for improving the accuracy of dietary assessments and enhancing the reliability of nutritional epidemiology research.

Several methods exist to improve the accuracy of dietary intake measurements, including regression calibrations and cumulative averages from repeated assessments. However, these are primarily designed to correct errors in habitual dietary intake, not the random errors inherent in single-day assessments. In other fields like computer vision, deep learning has proven effective in signal reconstruction from corrupted data. Techniques like Noise2Noise, which trains on noisy images alone, have shown promise. These approaches inspire the development of a method to correct random errors in nutrient profiles using only the available noisy data (the self-reported dietary assessment) and additional information (the gut microbiome). The established connection between gut microbiota and dietary intake provides a compelling basis for this innovative approach.

The researchers developed METRIC (Microbiome-based nutrient profile corrector), a deep-learning method using a multi-layer perceptron (MLP) neural network. METRIC takes the assessed nutrient profile and gut microbial composition as input and aims to infer the true nutrient profile. Because true nutrient profiles are unavailable for training, the researchers use a novel training strategy inspired by Noise2Noise. Instead of using true nutrient profiles, they generate "corrupted" nutrient profiles by adding random noise (Gaussian, Uniform, etc.) to the assessed nutrient profiles. The network is then trained to map these corrupted profiles back to the original assessed profiles. This forces the network to learn to remove the added noise, essentially acting as a denoiser. A skip connection is incorporated to enhance performance by providing the corrupted nutrient profile directly to a later layer of the network. The model uses a centered log-ratio transformation on microbial abundances and log transformation on nutrient profiles. The Adam optimizer minimizes mean squared error during training. Performance is evaluated using Pearson correlation coefficients (ρ) between predicted and true nutrient concentrations. The study utilized three real-world datasets (MCTS, MLVS, WE-MACNUTR) and a synthetic dataset generated using the Microbial Consumer-Resource Model (MICRM). In the real-world datasets, the nutrient profiles from more reliable methods (ASA24, 7DDR, complete feeding) served as the "true" profiles, with random noise added to create the "assessed" profiles for training and testing. The researchers also explored the performance of METRIC without using microbial composition as input and assessed the sensitivity of METRIC to changes in microbial compositions.

METRIC effectively minimized simulated random errors in nutrient profiles across all datasets, particularly for nutrients heavily metabolized by gut bacteria. In the synthetic data (MICRM), as the standard deviation of added noise increased, METRIC's correction performance improved, demonstrated by an increase in the difference between the correlation of corrected versus assessed values with the true values (ρc - ρa). In real-world datasets (MCTS, MLVS, WE-MACNUTR), METRIC consistently improved the correlation between predicted and "true" nutrient profiles, especially for nutrients with initially lower correlations. Dietary fiber showed consistently strong correction across datasets. The correction performance was dependent on the level of noise in the assessed nutrient profiles, with less effective correction when the initial noise was very low. Importantly, METRIC still exhibited substantial correction performance even without including microbial composition data, indicating robustness. Analysis of sensitivity revealed microbial taxa associated with fiber degradation were strongly linked to improved fiber concentration predictions. Furthermore, an analysis examining the temporal alignment of microbiome and dietary data suggested a causal relationship, with performance decreasing as the time offset increased.

METRIC successfully addresses a critical need in nutritional epidemiology by providing a method to correct random errors in self-reported dietary assessments. The ability to leverage gut microbial composition to improve accuracy is a significant advancement, particularly for nutrients that are significantly altered by gut microbial metabolism. The performance of METRIC, even without microbiome data, suggests broader applicability. The strong correction observed for dietary fiber highlights the potential for improved accuracy in the assessment of this important dietary component. The study's findings contribute to more accurate dietary assessment, potentially improving the understanding of diet-disease relationships and facilitating personalized nutrition strategies.

METRIC offers a novel deep-learning approach to improve the accuracy of nutrient profiles derived from self-reported dietary assessments. Its ability to correct random errors, particularly in nutrients processed by gut microbes, and its robustness even without microbiome data, highlight its significant potential. Future work should focus on validating METRIC using objective markers of nutrient intake and exploring its applicability to different dietary assessment methods and populations.

METRIC is limited to correcting random errors with zero means and cannot address systematic biases. The model's performance relies on the temporal proximity of dietary and microbiome data, limiting its applicability to dietary assessments with longer recall periods. The validation of METRIC is currently indirect, lacking the ideal use of objective markers of nutrient intake which would require extensive resources and datasets that are challenging to obtain. Further research with objective markers is needed.