Optimizing In Vivo Oral Lesion Classification from Electrical Impedance Spectroscopy Using Data-driven Approaches
Sophie A. Lloyd, Jacob P. Thönes, Safina S. Suratwala, Noor Zaghlula, Liang Lu, Joseph Paydarfar, Ethan K. Murphy, Sascha Spors, Ryan J. Halter · 2026
Read Optimizing In Vivo Oral Lesion Classification from Electrical Impedance Spectroscopy Using Data-driven Approaches by Sophie A. Lloyd, Jacob P. Thönes, Safina S. Suratwala, Noor Zaghlula, Liang Lu, Joseph Paydarfar, Ethan K. Murphy, Sascha Spors, Ryan J. Halter (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-05729.txt.
Introduction
Finding oral cancer early saves lives, and a handheld probe that reads tissue electrically could help. Sophie A. Lloyd and eight co-authors build a machine-learning pipeline for electrical impedance spectroscopy of oral lesions. Measurements came from 104 patients undergoing oral cancer surgery or biopsy, taken in vivo at the bedside.
The team evaluates three tasks: separating healthy tissue from cancer, classifying lesion types, and grading severity. Their data-driven approach combines multiple impedance features rather than relying on any single measurement. The goal is a real-time, non-invasive aid for clinicians deciding which lesions need urgent attention.
Catalog record
| CREATOR | Sophie A. Lloyd, Jacob P. Thönes, Safina S. Suratwala, Noor Zaghlula, Liang Lu, Joseph Paydarfar, Ethan K. Murphy, Sascha Spors, Ryan J. Halter |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.05729
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Optimizing In Vivo Oral Lesion Classification from Electrical Impedance Spectroscopy Using Data-driven Approaches Sophie A. Lloyd, Jacob P. Thönes, Safina S. Suratwala, Noor Zaghlula, Liang Lu, Joseph Paydarfar, Ethan K. Murphy, Sascha Spors, Ryan J. Halter Abstract Oral cancer is a significant global health burden, and early detection remains a critical clinical need. Electrical impedance spectroscopy (EIS) offers a promising non-invasive approach for real-time tissue characterization, but classification frameworks that jointly leverage multiple impedance features for in vivo oral lesion discrimination remain underdeveloped. This paper presents a machine-learning (ML) pipeline to optimize classification of in vivo oral pathology from EIS data collected using a handheld, bedside device. Impedance measurements were acquired from 104 patients undergoing oral cancer resection or biopsy. Three classification tasks were evaluated: (1) healthy vs. cancer, (2) multi-class lesion-type discrimination (cancer, high-grade dysplasia, non-malignant), and (3) multi-class discrimination between the three lesion pathologies and healthy tissue. For each task, signal frequencies were independently ranked and reduced using PCA, and different current injection/voltage measurement (IIVV) pattern geometries were tested. Classification performance was assessed through leave-one-patient-group-out cross-validation to ensure robustness on unseen patients. Input data dimensionality was reduced by up to 99% across all tasks while improving diagnostic accuracy over baseline models trained on the full dataset. A logistic regression model achieved the highest binary classification accuracy of 80% with an AUC of 0.90, while multi-class scenarios maintained AUCs above 0.82. All top-performing models utilized the significantly reduced IIVV set as input. The proposed pipeline advances EIS-based cancer detection by providing a robust, computationally efficient, and clinically practical framework for early diagnosis of oral cancer lesions, with a methodology readily generalizable to other EIS devices and applications. I Introduction Oral squamous cell carcinoma (OSCC) is a major global health problem: in 2025 , there were an estimated 60\,000 new cases and 13\,000 deaths in the United States [ 28 ] . Late-stage diagnosis is common for oral cancer and requires more aggressive treatment, leading to higher mortality, morbidity, and health-care costs. Early detection substantially improves outcomes: reported five‑year survival rates increase from about 58\text{\,}\mathrm{\%} for late-stage disease to roughly 89\text{\,}\mathrm{\%} when diagnosed early [ 16 ] . Despite this, frontline screening and diagnosis of early-stage OSCC and premalignant lesions remain inadequate. The clinical standard for diagnosis is an invasive tissue biopsy with histopathological analysis. While histopathology is highly accurate for distinguishing malignant from non-malignant tissue [ 29 ] , grading of dysplasia shows substantial inter-observer variability [ 20 , 30 ] . Biopsies are also invasive and resource intensive, requiring specialized equipment and trained personnel. Many dental clinics do not have ready access to histopathology services, and only about 25\text{\,}\mathrm{\%} of dentists perform oral biopsies, commonly citing costs, distance to laboratories, and lack of training [ 4 ] . These limitations reduce opportunities for timely diagnosis and early intervention. Non-invasive diagnostic technologies are therefore needed to enable earlier detection of OSCC and improve risk stratification of pre-malignant lesions across diverse clinical settings. Optical modalities, such as fluorescence [ 33 ] , autofluorescence [ 26 ] , and optical coherence tomography [ 2 ] have shown promise for tissue classification in oral lesions, but they require expensive equipment and expert interpretation, which restrict adoption in many primary care and dental environments. Electrical impedance spectroscopy (EIS) is a non-invasive, non‑ionizing technique that can reveal changes in tissue architecture and cell morphology associated with disease progression. EIS has been shown to discriminate healthy from cancerous tissue in multiple anatomical sites, including breast [ 14 ] , prostate [ 11 ] , and the oral cavity [ 25 , 23 ] . The oral cavity is particularly well-suited for EIS because lesions are readily accessible without surgical access. Nevertheless, clinical adoption of EIS for oral lesions has been limited. Two key barriers are (1) reliably distinguishing OSCC and high risk dysplastic lesions that require intervention from low risk lesions suitable for monitoring, and (2) providing dental clinicians with easy-to-interpret, actionable outputs so they can confidently refer or biopsy when indicated. Prior EIS investigations of oral lesions have reported promising classification performance, though few in vivo studies have been performed. Murdoch et al. [ 25 ] included 51 control subjects and 11 patients with oral potentially malignant disorder (OPMD) or oral cancer. They reported an area under the curve (AUC) of 0.78 for discriminating between in vivo high-risk lesions (OSCC and high-grade dysplasia) and low-risk lesions (low-grade dysplasia and healthy tissue). Sun, et. al. [ 31 ] and Ching, et. al. [ 5 ] explored the use of a 4-electrode bioimpedance probe in the context of tongue cancer detection with 12 tongue cancer and healthy subjects at six different frequencies. Those studies indicate potential for the use of EIS but are limited by relatively small numbers of OPMD samples. Recent work with this device using a larger electrode-array EIS device recorded in vivo measurements from 25 subjects and reported high performance (AUC=0.88) for discriminating malignant from healthy tissue, motivating further study with broader patient populations and tissue types [ 23 ] . While conventional EIS devices typically inject current at one electrode pair and measure voltage at another, our multi-electrode architecture enables simultaneous voltage measurements across numerous electrode configurations. This approach, inspired by electrical impedance tomography (EIT) system design but applied to EIS measurement, generates high-dimensional datasets that contain rich diagnostic information but also present analytical challenges distinct from traditional single-pair EIS. Conventional threshold-based analyses typically use only one impedance parameter at a time, measured at one frequency; thus, they are constrained in their ability to exploit the complex, multi-electrode, multi-frequency data from our device. Machine learning (ML) methods have become widely used in medical diagnostics for classification, detection, and segmentation. In oral cancer research, several studies have applied ML, particularly convolutional neural networks (CNNs), to a range of imaging modalities and reported encouraging performance. Examples include CNN-based classification of biopsy-derived epithelial layer images with test accuracy near 89\text{\,}\mathrm{\%} [ 10 ] , automated lesion classification from the spectral-spacial features of hyperspectral-image lesion imaging [ 17 ] , deep-learning combined with traditional classifiers on confocal laser endomicroscopy achieving accuracies around 88\text{\,}\mathrm{\%} [ 1 ] , and 3D CNNs that improve CT-based diagnosis by capturing inter-slice continuity [ 35 ] . However, relatively few studies have applied ML directly to EIS data for oral lesions. A recent study by Lin et. al. [ 22 ] applied a CNN directly to EIS data collected from the ventral tongue and floor of mouth ( 51 healthy controls and 11 patients with OPMD or oral cancer), reporting strong discrimination with an AUC of 0.907 for classifying high versus (vs) low risk lesions. Veil et. al. [ 34 ] investigated a patient-based classification approach using a Gaussian process classifier (GPC) to differentiate healthy from tumorous bladder tissue using EIS data collected from ex vivo samples of six patients. Additionally, their team explored methods to reduce the dimensionality of the feature space to improve model efficiency. However, the limited number of patients with cancer and OPMD restricts the robustness and generalizability of these models. Though these models show promising results, the current study does not include CNN or GPC models due to the high risk of overfitting when trained on