Abstract
Electrical impedance spectroscopy classifiers often use Cole–Cole or Nyquist descriptors, yet among 20 representative studies none separated Cole fit residuals from dispersion parameters, and only three reported patient-separated evaluation. We compared resistance–reactance spectra, Cole–Cole descriptors, local and global Nyquist-trajectory descriptors, and hybrids for generalization to unseen biological units on human bladder, porcine colon–rectum, and five human head-and-neck tasks. Validation held out patients or specimens, which served as inferential units; comparisons of selected candidates are exploratory. No representation was uniformly best, and raw spectra remained a strong baseline. In bladder tissue, which a single dispersion described well, Cole descriptors increased balanced accuracy from 86.1% to 97.3%, mostly through the dispersion parameters. In mucous tissue, which a single dispersion described poorly, a smaller gain (+4.4 points) was concentrated in the fit-residual descriptors. Fit adequacy indicated where Cole-derived information lay. Neither contrast survived multiplicity correction, although cluster-bootstrap intervals excluded zero. Under nested regularization tuning, the bladder contrast remained substantial (+15.1 points) and the mucous contrast decreased (+3.8 points); both intervals included zero. A unit-level audit found one bladder patient’s labels contradicting acquisition metadata, an inconsistency that had reduced all configurations to near chance and that patient referencing would not reveal; excluding this patient gave similar results. Most head-and-neck patients contributed one class per task, so these tasks mainly tested between-patient discrimination; most of their spectra failed a Kramers–Kronig check, leaving Cole descriptors descriptive only; lymph-node and muscle tasks remained at chance. A representation-selection workflow and a ten-item reporting checklist condense the findings.

This publication has 49 references indexed in Scilit: