Prediction of Visceral Fat in Qatari Population Using Machine Learning

The rapid increase in obesity prevalence globally is posing a major epidemiological challenge to today’s healthcare systems. Obesity is responsible for 5% of all deaths with a global economic impact of approximately $2 trillion annually. Timely identification and quantification of excessive body fat plays a key role in having successful intervention programs. Visceral fat that is stored in the abdominal cavity and surrounds internal organs, in contrast to subcutaneous fat that is stored underneath the skin, is a marker of obesity-related diseases. Visceral fat causes inflammatory immune response causing damage to tissue and organs and is best quantified using medical imaging devices such as Computed Tomography (CT), Magnetic Resonance Imaging (MRI), or Dual-energy X-ray Absorptiometry (DXA). However, these devices are expensive, not always available, and are not suitable for large-scale epidemiological studies and community-wide screening. In this study, several linear, nonlinear, and ensemble machine learning regression models are developed to estimate visceral fat mass by using easy to obtain anthropometric measurements. Our methods are evaluated on 1,000 Qatar Biobank study participants and achieve adjusted coefficient of determination (R²) of 0.843 using a two-level stacked ensemble learning model. Waist circumference is identified as the feature of most importance for the estimation of visceral fat mass. In addition to estimating visceral fat mass, machine learning classification models are developed to evaluate the participants metabolic risk profile using visceral fat area thresholds. A linear Support Vector Machine (SVM) classifier achieves a recall of 90.1%. Using these models, clinicians and healthcare practitioners in Qatar will be able to perform initial screening for visceral obesity in a rapid, inexpensive, and noninvasive manner.