AI Case Study
DayTwo improves blood sugar management through personalised diet recommendations based on gut microbiome analysis using machine learning
DayTwo is an app which provides users with their optimal diet based on their individual gut microbiome analysis. Ostensibly this is used to manage blood sugar levels and can help avoid developing prediabetes and type II diabetes, but can be used for a personalised approach to healthier eating in general. This is done through analysing biomaterial as well as eating and sleeping habits using a model developed with machine learning.
Pharmaceuticals And Biotech
DayTwo is an app which, according to its website, uses "a scoring system to rate thousands of different foods and food combinations based on your biometrics, gut microbiome analysis, lifestyle factors and health questionnaire – which yield a unique nutrition profile that enables blood-sugar balance." After sending in stool and blood test results, users receive personalised diet advise, including a list of foods to indulge in and those to avoid. The analysis is based on research conducted by the Weizmann Institute of Science.
From the research paper: "Elevated postprandial blood glucose levels constitute a global epidemic and a major risk factor for prediabetes and type II diabetes, but existing dietary methods for controlling them have limited efficacy."
Results from the research paper indicated that "[d]ietary interventions based on our predictor showed significant improvements in multiple aspects of glucose metabolism, including lower PPGRs and lower fluctuations in blood glucose levels within a short 1-week intervention period. Our predictor that integrates the above person-specific factors predicts the held-out PPGRs of individuals with a significantly higher correlation."
The research paper "employed a two-phase approach. In the first,
discovery phase, the algorithm was developed on the main
cohort of 800 participants, and performance was evaluated using
a standard leave-one-out cross validation scheme, whereby
PPGRs of each participant were predicted using a model trained
on the data of all other participants. In the second, validation
phase, an independent cohort of 100 participants was recruited
and profiled, and their PPGRs were predicted using the model
trained only on the main cohort. Given non-linear relationships between PPGRs and the different factors, we devised a model based on gradient boosting regression. This model predicts PPGRs using the sum of thousands of different decision trees. Trees are
inferred sequentially, with each tree trained on the residual of
all previous trees and making a small contribution to the overall
prediction. The features within each tree are selected by an inference procedure from a pool of 137 features representing
meal content (e.g., energy, macronutrients, micronutrients);
daily activity (e.g., meals, exercises, sleep times); blood parameters (e.g., HbA1c%, HDL cholesterol); CGM-derived features; questionnaires; and microbiome features (16S rRNA and metagenomic RAs, KEGG pathway and module RAs and bacterial
growth dynamics - PTRs)".
The research paper "continuously monitored glucose levels during an entire week in a cohort of 800 healthy and prediabetic individuals and also measured blood parameters, anthropometrics, physical activity, and selfreported lifestyle behaviors, as well as gut microbiota composition and function. With a total of 10,000,000 Calories logged, our data provide a global view into the cohort’s dietary habits, showing the fraction that each food source contributes to the cohort’s overall energy intake... We further validated our model on an independent cohort of 100 individuals that we recruited separately. Data from this additional cohort were not available to us while developing the algorithm. Participants in this cohort underwent the same profiling as in the main 800-person cohort".