+233 50107 3889 admin@ghscientific.com

Blog Details

  • Home / News / ESTIMATING TRUE FOOD…
ESTIMATING TRUE FOOD CONSUMPTION WITH ARTIFICIAL INTELLIGENCE

ESTIMATING TRUE FOOD CONSUMPTION WITH ARTIFICIAL INTELLIGENCE

MinoHealth recently published a paper on a completed research project where we applied Artificial Intelligence to Dietetics/Nutrition. The problem tackled was  Food Consumption Estimation, people tend to wrongly estimate how much they consume and this is hurdle for Dietitians and Nutritionists to promote and ensure  food portion size control and nutritional ​management ​of ​diseases. Also, The increase of obesity prevalence has been partly attributed to the increased consumption of food portion size which is largely due to overestimation. So we developed FoodEstNet, which is an xGBoost(an implementation of Gradient Boosting) model  that’s trained to estimate how much of a food type a person consumes.

This Artificial Intelligence model/system,  FoodEstNet takes in as input a person’s Gender, Age, BMI, Socioeconomic status attributes, Food Type and Perceived Consumption then outputs their Actual Food Consumption. The Food Types supported are the 12 commonly consumed carbohydrate foods in Ghana (according to research work by Boateng, GP), which are  chocolate drink, white bread, boiled rice, Ga kenkey, granulated sugar, corn porridge, waakye, gari, boiled yam, boiled green plantain, banku ​and fufu​.

The dataset that we used to train and test FoodEstNet had 4,800 data samples and 12 columns. The 12 columns were; ​Sex, Age, Ethnicity, Marital Status, Education, Average Monthly Income, Mental & Physical Health, BMI, Religion, Food Type, Perceived Amount, Actual Amount.  Prior to choosing XGBoost, multiple Machine Learning algorithms were applied to the problem, including ​Linear Regression, Ridge, Lasso, Elastic Net, Bagging, Random Forest, Extra Trees, K Nearest Neighbors, Decision Tree, Gaussian Process, Support Vector Machine, Multilayer Perceptron(Neural Networks/Deep Learning) ​along with XGBoost. The XGBoost(with Early Stopping) model scored considerably higher than the other models. The model was trained with  ​Mean Absolute Error(MAE) ​loss function, and was tested with  Explained Variance Accuracy score and ​Root Mean Square Error( RMSE or RMSD).  FoodEstNet scored an Explained Variance Accuracy score of 63.62% ​ andRMSE of ​1.55709.  ​

With  larger datasets and additional research, we’re confident models could be developed with higher accuracy scores and performances that they could be used in the real world by Nutritionists, Dietitians and people in general alike to accurately estimate Food Consumption for food portion size control and nutritional ​management ​of ​diseases.

Link to full paper on BioRxiv: https://www.biorxiv.org/content/early/2018/01/19/250506

Leave a Reply

Your email address will not be published. Required fields are marked *

en_GBEnglish