Predicting Body Mass Index From Self-Declared Socio-Demographic, Psychological, and Behavioral Data With Artificial Neural Network

Erkan Tiyekli, Olgun Duran, Nimet Gök, Pemra C. Unalan, Hatice Kurdak

Keywords: Artificial neuronal network, body mass index, prediction, cognitive-behavioral data

Background:
Artificial neural network (ANN) is a powerful tool that can successfully predict significant patterns among big data for very different disciplines, including medicine. Obesity has also entered the field of interest of ANN applications due to its complex aetiology.

Research questions:
Can it be possible to create a reasonable model for predicting body mass index (BMI) with socio-demographic components and associated cognitive-behavioural factors using an ANN?

Method:
In this study, a feed-forward backpropagation algorithm is applied to create an ANN model and tested with mean squared error (MSE). The dataset was provided from a dissertation study on obesity prevalence and related factors on our university employees. Investigators agreed on a parameter set with principal component analysis. The categoric variables were transformed to binary codes by one-hot encoding, and min-max normalization was performed for continuous variables.

Results:
Of the 825 subjects 497 were women (60.24%), with age from 18 to 66 y (mean=38.37±9.53) and BMI from 16.90 to 44.10 kg/m2 (mean=25.95±4.36). In the model created, 22 parameters were selected as input data (socio-demographics, chronic illness, eating habits, eating speed, weight control methods, weighing frequency, response to an irresistible meal, emotional response in weight increase, emotional eating). The model predicted the BMI with high accuracy and low MSE (R2= 0.85, MSE=0.1).

Conclusions:
The BMI is reasonably an easy tool to calculate. However, this ANN that predicted the BMI successfully brings forward the importance of the underlying elements of obesity. Moreover, a better understanding of the model between obesity and the underlying paradigms could make it easier for primary healthcare professionals to guide their patients in obesity management. A web-based ANN calculator could also be shared with the patient to illustrate the importance of underlying components.

Points for discussion:
Which different healthy lifestyle parameters that can be obtained easily would be considered ANN input data to improve the impact of the information shared with patients?

Can a modified ANN model be developed from internet cookies or social media behaviour?

Can this model be converted into a healthcare professional-friendly data mining/web application?