Skin layer classification by feedforward neural network in bioelectrical impedance spectroscopy

Kiagus Aufa Ibrahim,Marlin Ramadhan Baidillah,Ridwan Wicaksono,Masahiro Takei

DOI: https://doi.org/10.2478/joeb-2023-0004

2023-01-01

Journal of Electrical Bioimpedance

Abstract:Abstract Conductivity change in skin layers has been classified by source indicator o k ( k =1: Stratum corneum, k =2: Epidermis, k =3: Dermis, k =4: Fat, and k =5: Stratum corneum + Epidermis) trained from feedforward neural network (FNN) in bioelectrical impedance spectroscopy (BIS). In BIS studies, treating the skin as a bulk, limits the differentiation of conductivity changes in individual skin layers, however skin layer classification using FNN shows promise in accurately categorizing skin layers, which is essential for predicting source indicators o k and initiating skin dielectric characteristics diagnosis. The o k is trained by three main conceptual points which are (i) implementing FNN for predicting k in conductivity change, (ii) profiling four impedance inputs α ξ consisting of magnitude input α | z |, phase angle input α θ , resistance input α R , and reactance input α x for filtering nonessential input, and (iii) selecting low and high frequency pair ( f r l h ) $$(f_{r}^{lh})$$ by distribution of relaxation time (DRT) for eliminating parasitic noise effect. The training data set of FNN is generated to obtain the α ξ ∈ R 10×17×10 by 10,200 cases by simulation under configuration and measurement parameters. The trained skin layer classification is validated through experiments with porcine skin under various sodium chloride (NaCl) solutions C NaCl = {15, 20, 25, 30, 35}[mM] in the dermis layer. FNN successfully classified conductivity change in the dermis layer from experiment with accuracy of 90.6% for the bipolar set-up at f 6 l h = 10 & 100 [ kHz] $$f_{6}^{lh}=10\,\And 100\,{\rm{[kHz]}}$$ and with the same accuracy for the tetrapolar at f 8 l h = 35 & 100 [ kHz] $$f_{8}^{lh}=35\,\And 100\,{\rm{[kHz]}}$$ . The measurement noise and systematic error in the experimental results are minimized by the proposed method using the feature extraction based on α ξ at f r l h $$f_{r}^{lh}$$ .