Skin layer classification by feedforward neural network in bioelectrical impedance spectroscopy

Abstract

Conductivity change in skin layers has been classified by source indicator ok (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 ok and initiating skin dielectric characteristics diagnosis. The ok is trained by three main conceptual points which are (i) implementing FNN for predicting k in conductivity change, (ii) profiling four impedance inputs αξ consisting of magnitude input α|z|, phase angle input αθ, resistance input αR, and reactance input αx for filtering nonessential input, and (iii) selecting low and high frequency pair (frlh)$$(f_{r}^{lh})$$ by distribution of relaxation time (DRT) for eliminating parasitic noise effect. The training data set of FNN is generated to obtain the αξ ∈ R10×17×10 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 CNaCl = {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 f6lh=10 &100 [kHz]$$f_{6}^{lh}=10\,\And 100\,{\rm{[kHz]}}$$ and with the same accuracy for the tetrapolar at f8lh=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 αξ at frlh$$f_{r}^{lh}$$.

Keywords

• bioelectrical impedance spectroscopy
• conductivity change
• distribution of relaxation times
• feedforward neural network