782-P: Personalized, Machine Learning-Based Nutrition Reduces Diabetes Markers in Type 2 Diabetic Patients

Abstract

HbA1c% is the most commonly used metric for assessing glycemic status, but only represents an averaged, indirect measure. Continuous Glucose Monitors (CGM) provides additional, elaborated assessment metrics. A recent study, [Zeevi et al., 2015, Cell ] showed that glycemic responses to foods vary across individuals and can be predicted using a machine learning framework. Moreover, it showed that personally tailored diets based on this framework improves Postprandial Glycemic Response (PPGR) in one-week interventions. Here, we tested the ability of such framework to improve longer term measures of glycemic control. A cohort of 28 diabetic individuals, not using short term insulin, were tracked. Each participant was tested for initial HbA1c, and connected to a CGM for a period of 14 days. Then, participant's anthropometrics, lifestyle, clinical parameters, and gut microbiome composition, were fed into a machine learning algorithm built into a personalized mobile application. Using the application, participants could define meals by combining foods, and obtain instant scores indicating the predicted PPGR for each meal. Participants were instructed to limit consumption to highly scoring meals. After 4-20 months of using the application, participants were re-connected to CGM, and HbA1c% levels were measured again. Results: Significant improvements in multiple endpoints: Average HbA1c% dropped from 7.2% to 6.5% (p-value: 1.2e-8). Average %time-in-range [70,140] mg/dl increased from 69.1% to 79.6% (p-value: 0.005). Average %time-in-range [70,180] mg/dl increased from 89.6% to 94.2% (p-value: 0.002). Mean glucose levels decreased from 125.6 mg/dl to 114.6 mg/dl (p-value: 0.0002). These observations, coupled with the short-term benefits shown in recent work, imply that drugless, personalized, nutrition-based interventions, may be key to achieving significant improvements in glycemic control of type 2 diabetic patients. Disclosure Y. Ben Shlomo: Employee; Self; DayTwo. S. Azulay: Employee; Self; DayTwo. T. Raveh-Sadka: Employee; Self; DayTwo. Y. Cohen: Employee; Self; DayTwo. A. Hanemann: Employee; Self; DayTwo.