Yiheng Shen, Euiji Choi, Samantha Kleinberg
Although CGM was the most important feature group, combining it with personal training data and timely microbiome samples led to the most accurate models in our analysis. These findings can help researchers understand the tradeoffs between the time and effort of data collection and how data types impact model performance.
BACKGROUND: Individual glycemic responses to foods vary and can be predicted using microbiome, activity, and dietary data. However, these data are expensive and invasive to collect, and it is not known how much each modality contributes to accuracy.
OBJECTIVES: We aim to quantify the contributions of dietary, sleep, continuous glucose monitor (CGM), and microbiome features for glycemic response prediction; understand how much personal data are required for training; and evaluate how microbiome sample timing impacts model accuracy.
METHODS: We used data from 8334 participants in the Human Phenotype Project cohort study who provided demographic, anthropometric, dietary, and CGM data. Participants self-reported meals in a dietary tracking application for a mean of 10.78 d, during which they wore CGMs. We trained CatBoost models to predict postprandial glycemic response (PPGR) using 2-h incremental area under the curve and peak 2-h postprandial glucose rise (Glumax). We conducted ablation studies with varied feature combinations to assess the contribution of each data modality. We used 3 train/test splits (split-by-meal, 5-d personal training, and split-by-person) to assess the impact of personal training data. Lastly, we evaluated accuracy as a function of microbiome sample timing (from before meal logs to ≤60 d after).
RESULTS: The model combining all features performed best, and CGM was the most informative feature. Models trained with more personal data had the best performance (PPGR split-by-meal R = 0.731; split-by-person R = 0.590), and personal training data had a larger effect on accuracy than microbiome. Microbiome features improved predictions most when collected within 7 d of meal logs and did not improve performance without personal training data or for samples collected >14 d after meal logs.
CONCLUSIONS: Although CGM was the most important feature group, combining it with personal training data and timely microbiome samples led to the most accurate models in our analysis. These findings can help researchers understand the tradeoffs between the time and effort of data collection and how data types impact model performance.