Machine Learning-Based Thickness Estimation of Abdominal Fat and Muscle Using Simulated Radio Frequency Scattering Parameters

The estimation of three abdominal tissue thicknesses, namely subcutaneous fat (SAT), muscle tissue (MT), and visceral fat (VT), was performed using radio frequency (RF) signal measurement and machine learning algorithms. The scattering parameter data were collected from simulations of a planar multilayered model of the human abdomen, with parametric analysis on each layer's thickness. Various models, incorporating combinations of several frequency bands and 120 data points, were investigated. The results showed that polynomial regression (PR) and ridge polynomial regression (RiPR) using scattering parameters from frequency bands of 300 MHz, 400 MHz, 500 MHz, and 1 GHz performed the best among other models. PR (300 MHz, 400 MHz, 500 MHz, and 1 GHz) achieved RMSE values of 0.71 mm for SAT estimation, 1.2 mm for MT estimation, and 1.17 mm for VT estimation. RiPR (300 MHz, 400 MHz, 500 MHz, and 1 GHz) achieved RMSE values of 0.35 mm for SAT estimation, 1.57 mm for MT estimation, and 1.7 mm for VT estimation. However, these results indicated a high error percentage for VT estimation, as VT had a mean thickness of only 3 mm. The results also showed that the deeper the tissue, the more prone the estimation was to error. This is due to the nature of electromagnetic reflection and transmission within a multilayered medium, which affects the scattering parameter measurement. While the models perform well for upper tissue estimation, further improvements in the model, feature selection, and physical solutions are necessary to enhance the accuracy of deeper tissue estimation.