Optimizing In Vivo Oral Lesion Classification from Electrical Impedance Spectroscopy Using Data-driven Approaches
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910196117274624 |
|---|---|
| author | Lloyd, Sophie A. Thönes, Jacob P. Suratwala, Safina S. Zaghlula, Noor Lu, Liang Paydarfar, Joseph Murphy, Ethan K. Spors, Sascha Halter, Ryan J. |
| author_facet | Lloyd, Sophie A. Thönes, Jacob P. Suratwala, Safina S. Zaghlula, Noor Lu, Liang Paydarfar, Joseph Murphy, Ethan K. Spors, Sascha Halter, Ryan J. |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_05729 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Optimizing In Vivo Oral Lesion Classification from Electrical Impedance Spectroscopy Using Data-driven Approaches Lloyd, Sophie A. Thönes, Jacob P. Suratwala, Safina S. Zaghlula, Noor Lu, Liang Paydarfar, Joseph Murphy, Ethan K. Spors, Sascha Halter, Ryan J. Signal Processing 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. |
| title | Optimizing In Vivo Oral Lesion Classification from Electrical Impedance Spectroscopy Using Data-driven Approaches |
| topic | Signal Processing |
| url | https://arxiv.org/abs/2605.05729 |