PERSONALIZED GLYCEMIC RESPONSES TO FOOD AMONG INDIVIDUALS WITH TYPE 2 DIABETES IN INDIA: DEVELOPMENT OF A MACHINE LEARNING PREDICTION MODEL

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Objective

Emerging global evidence demonstrates marked inter-individual differences in post-prandial glucose response (PPGR) although no such data exists in India and prior studies have primarily evaluated PPGR variation in individuals without diabetes. This study sought to develop a machine learning model to predict individual PPGR responses to facilitate the prescription of personalized diets for individuals with type 2 diabetes.

Research Design and Methods

Adults with type 2 diabetes and a hemoglobin A1c (HbA1c) ≥7 were enrolled from 14 sites around India. Subjects wore a continuous glucose monitor and logged meals. PPGR was calculated for each meal, based on the incremental area under the curve, and a machine learning predictor of PPGR was developed using stochastic gradient boosting regression. Model calibration and discrimination was assessed using a Pearson product moment correlation and area under a receiver operating curve (AUC), respectively, and its performance was compared to models based only on meal carbohydrate and calorie content.

Results

The study included data from 488 patients (mean age 52.5 years, 36% female, mean duration of diabetes 6.4 years, mean hemoglobin A1c 8.16%). Mean PPGR to common foods varied substantially (e.g. PPGR for “aloo paratha with curd” ranged from 10 to 170 mg/dl*h). PPGR values predicted by the machine learning model were highly correlated with observed PPGR (r=0.69) and model calibration was substantially stronger than for a model based only on calorie (r=0.57) or carbohydrate (r=0.39) content. The machine learning model also demonstrated very strong discriminative ability (AUC 0.80).

Conclusions

A machine learning model built with nutritional content, health habits, biometric information and common laboratory data produced highly accurate individualized predictions of PPGRs that substantially outperformed predictions based upon calorie and carbohydrate content. These results could be used to facilitate the delivery of personalized medical nutritional therapy as is widely recommended by type 2 diabetes practice guideline in India and globally.

Article activity feed