In Vivo Classification of Oral Lesions Using Electrical Impedance Spectroscopy

Abstract

Objective:

To evaluate a new non-invasive, handheld Electrical Impedance Spectroscopy (EIS) device for assessing oral lesions in real-life surgical scenarios.

Methods:

A custom-designed probe with a 33-electrode sensor array was used to collect impedance measurements across multiple frequencies (100 Hz – 100 kHz) from non-consecutive patients undergoing surgical resection of oral cancer. In vivo EIS measurements were recorded from lesion and healthy tissue surfaces before resection, with no clinical decisions based on impedance data.

Results:

The study included 26 participants (median [IQR] age, 64.3 [59 – 70] years; 11 (42%) female) with oral squamous cell carcinoma. Cancerous tissue was found to have significantly lower resistance and reactance than healthy tissue (p<0.0001). Tissue classification using the permittivity at 40 kHz showed the highest accuracy (88%) with an AUC of 0.88. Multiple impedance parameters achieved AUCs >0.85 for differentiating healthy from malignant tissue.

Conclusion & Significance:

The study indicates that EIS can effectively differentiate between healthy and cancerous oral mucosa through rapid, non-invasive intraoperative measurements. The data processing pipeline developed demonstrates success in maintaining high data quality amidst the external disturbances presented in intraoperative data collection.

I. INTRODUCTION

Surgical resection is the primary treatment for oral squamous cell carcinoma (OSCC). Intraoperatively, it is important to achieve negative margins to prevent the need for further intervention. Patients left with impacted positive margins following cancer resection have significantly increased risk of local recurrence [1], [2], lower rates of progression-free survival [2], [3], and require adjuvant treatments including additional surgery, radiotherapy, and chemotherapy. These additional interventions are associated with increased risk of morbidity and decreased quality of life [4], [5].

Current rates of positive surgical margins are reported to range from 1% to 44% and are highly dependent on surrounding anatomy and lesion location within the oral cavity, surgeon experience, and positive margin definition [2], [3]. Studies have specifically shown poor outcomes associated with cancerous tissue within 5 mm of final margins [6], [7]. Achieving a 5 mm clearance at the deep margin is particularly challenging because preoperative imaging becomes unreliable due to tissue deformation, and there is limited space for palpation [8], [9]. The relatively high rate of positive margins in OSCC resections demonstrate limitations in the effectiveness of the current available methods for intraoperative margin surveillance, including frozen section analysis [10], [11], [12].

Given the poor patient outcomes associated with positive margins, surgeons would greatly benefit from a more advanced intraoperative margin assessment technology. Many emerging approaches leverage optical contrast for tissue classification including fluorescence imaging and optical coherence tomography (OCT) [13], [14], [15]. Although these modalities offer high-resolution assessment of tissue surfaces, they are limited to sensing depths of only 1–2 mm, which may be insufficient for guiding surgeons in achieving clear margins, particularly at the deep margin. Additionally, their interpretation often requires specialized expertise, and their intraoperative deployment can be costly.

Tissue electrical properties, governed by intra- and extracellular composition, morphology, and cellular makeup, can differentiate benign from malignant tissues across various organs, including the cervix [16], [17], breast [18], [19], [20], skin [21], [22], and prostate [23], [24], [25]. Cell membranes behave as capacitors, blocking current flow through cells at low frequencies, while allowing current to penetrate cells at higher frequencies. Thus, at low frequencies, impedance measurements largely reflect the properties of the extracellular fluid, while at higher frequencies they are more sensitive to the intracellular properties [26], [27]. Normal tissue and benign and malignant lesions of the oral cavity and oropharynx have vastly different morphological structures [28] that impact the electrical impedance spectra recorded from the tissue (Fig. 1). It is possible to sense these electrical properties by placing electrodes on the tissue of interest; electrical impedance spectroscopy (EIS) is ideally suited for this task in the context of classifying oral lesions since there is direct access to the oral cavity for electrode placement.

Fig. 1.

Intraoperative Impedance Probe Study Overview. A) Handheld probe, B) Example measurement pattern [current driven between blue peripheral electrodes, voltage sensed between red central electrodes], C) Typical measurement protocol of lesion and contralateral healthy sites, D) Example histological variation between the OSCC and healthy oral mucosa, E) Probe in use.

Prior studies have established significant differences in the impedance of oral mucosa at various sites in the oral cavity [29], [30], as well as differences between healthy and malignant oral tissues [31], [32], [33], [34], [35]. Cheng, et al. [31] designed a probe with four concentric rings and validated its performance in phantoms and ex vivo tissue. Murdoch, et al. [32] reported an 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 benign tissue). They specifically relied on the impedance difference between the healthy control and lesion measurements within an individual subject for classification instead of exploring the potential of using the measured impedance from a single site only for classification. While their results were encouraging, this approach can be challenging to deploy in cases with mid-line lesions (i.e., no contralateral region), bilateral disease, or non-malignant changes in contralateral mucosa. Additionally, the four-electrode device used in their study was designed to gauge cervical anatomy, and posed challenges in achieving good electrode contact within the oral cavity; this ultimately resulted in a failure to record a spectrum in nearly 10% of subjects enrolled in the study. Finally, Sun, et. al. [33] and Ching, et. al. [34] explored the use of a 4-electrode bioimpedance probe in the context of tongue cancer detection, but did not explore the broader oral cavity, and Hu, et. al. [35] developed a multi-electrode probe for assessing tongue tissue, but only reported results in phantom studies.

Here we present a multi-electrode probe designed specifically for capturing intraoperative oral cavity impedances. The multi-channel design provides measurement redundancy, enabling robust intraoperative EIS data capture, the potential for electrical impedance tomography (EIT) to image tissue heterogeneity (such as at tumor margins), and the ability to compute probe geometry-independent conductivity and permittivity properties of healthy and malignant oral tissues. Previous EIT work demonstrated the ability to localize 1.5 mm muscle inclusions [36], and muscle-fat boundaries have been imaged through ex vivo bovine experiments [37].

This probe was deployed intraoperatively and, to the best of our knowledge, represents the first study reporting on the use of absolute impedance to accurately classify in vivo oral cancer lesions from within the broader oral cavity. Additionally, this work introduces a impedance-based data processing pipeline robust enough to handle the disturbances presented in real-life applications of impedance spectroscopy in surgery.

II. METHODS

A. Electrical Impedance Spectroscopy

In EIS a small electrical alternating current ($I$) of a particular frequency is driven between a pair of electrodes placed on the tissue surface, while the voltage difference ($V$) between a second pair of electrodes is simultaneously measured. Ohm’s Law,

$$Z=\frac{V}{I}=R+jX[\mathrm{\Omega }]$$ (1)

can be used to calculate the frequency dependent complex bioimpedance ($Z$) consisting of a real resistive component ($R$) and an imaginary reactive component ($X$). The magnitude of impedance is defined as

$$|Z|=\sqrt{R^2+X^2}[\mathrm{\Omega }].$$ (2)

The phase of impedance is expressed as

$${\varphi }_Z={\mathrm{t}\mathrm{a}\mathrm{n}}^{-1}\frac{X}{R}[\mathit{radians}].$$ (3)

B. Electrical Impedance Acquisition System and Probe

A custom-designed electrical impedance probe (Fig. 1A) connected to an impedance analyzer was used for in vivo data collection. The probe interfaced to a commercial 32-channel impedance data acquisition system (EIT32, Sciospec Scientific Instruments GmbH, Bennewitz, Germany) programmed to inject electrical current at 0.1 mA amplitude and measure the resulting voltages simultaneously on all channels. The custom probe features an 11 mm diameter electrode array housed in a durable, resin-based 3D-printed casing (Form 3, Formlabs, MA, USA). Meter-long cables connect the probe to the impedance acquisition system, enabling the sterilized probe to remain in the sterile field during intraoperative measurements while the non-sterile impedance analyzer stays outside. The probe tip is angled 60° downward from the handle, optimizing access to most oral cavity sites. The 12 cm long handle enables measurements of lesions in the deepest regions (e.g., the tonsils).

The electrode array, previously described and validated in [36], [37], was designed for tetrapolar impedance sensing with an inner 5×5 grid of 600 μm diameter voltage-sensing electrodes encircled by eight larger current drive electrodes. The center voltage-sensing electrode was not used in this study due to the limited number of impedance-analyzer channels (N=32). Measurements can be impacted by a buildup of charges at the electrode-tissue interface of polarized current-carrying electrodes, creating a capacitive effect, especially at lower signal frequencies. To combat this, tetrapolar patterns are used, where the current is injected between two electrodes (denoted II) and the voltage recorded with a different pair of electrodes (denoted VV). The larger surface area peripheral electrodes exhibit lower contact impedance, thus enabling higher signal-to-noise ratios (SNR) at lower frequencies. The smaller voltage-sensing electrodes have high-input impedance amplifiers on the back-end, preventing current flow and limiting contact impedance affects.

The eight current drive electrodes enable twenty-eight unique II pairs with 276 uniquely paired VV measurements amongst the inner twenty-four electrodes. This results in a total of 7,728 possible IIVV combinations. Impedance is sampled from all IIVV combinations at each of 31 logarithmically spaced frequencies ranging from 100 Hz – 100 kHz. Each spectrum takes approximately three seconds to collect using a 0.5 Hz frame rate.

C. Study Population and Demographics

The protocol for this non-randomized, investigator-initiated study (NCT05430477) was approved by the Institutional Review Board at Dartmouth-Hitchcock Medical Center (IRB #02001317). Written informed consent was obtained from all patients before participation in the study. This was a prospective, single-center feasibility study. Non-consecutive adult patients (aged ≥18 years) undergoing primary surgery of the oral cavity or oropharynx for biopsy-proven squamous cell carcinoma from October 2022 to March 2023 were prospectively enrolled in the study. Patients who had undergone biopsy in the past 30 days or who had cardiac implantable electronic devices were excluded from the study.

D. Data Acquisition Protocol

Patients were treated according to institutional standard of care for surgical treatment and specimen histopathological assessment. Impedance devices were sterilized using low temperature hydrogen peroxide gas plasma (STERRAD) prior to use in surgery. A set of in vivo EIS measurements were acquired intraoperatively from the center of the oral lesion, followed by a recording at a healthy contralateral site, if available. The face of the probe was placed in contact with the oral tissue and held stationary for approximately ten seconds while three consecutive impedance spectra were recorded in a “burst” measurement. The surgeon then picked up the probe and this process was repeated two more times in “repeat test” measurements, resulting in nine impedance spectra recordings from each tissue sample. The impedance of a saline solution of known conductivity (0.2 S/m) was measured after in vivo tissue measurements to serve as a calibration data set. The average impedance values from a smaller subset of geometrically identical current drive patterns with high sensitivity to impedance changes were used for the clinical analysis (Fig. 1B) presented here.

E. Data Preprocessing

Several real-world challenges can limit acquisition of high-quality intraoperative impedance measurements. In practice, the largest of these is maintaining good electrode-tissue contact during data collection. The various sites and lesions within the oral cavity present differently textured tissue surfaces, accessibility is dependent on anatomic location of the target site, and probe motion during data acquisition can all impact electrode-tissue contact. In addition, probe failures due to improper use by the operating room staff can occur; in one case, the cable was partially cut by a surgical technician, fully disconnecting several electrodes between sample measurements.

To account for these scenarios, we developed a custom data cleaning framework, graphically described in Figure 2. Overall, the filter checks for poor or incomplete electrode contact in each burst measurement to provide high quality spectral data averaged across bursts and repeat tests. Measurement redundancy captured by the multielectrode probe enables multiple IIVV patterns to be filtered out while still providing high quality, robust data. Figure 2 shows how the developed filter retrieves the valid data from particularly challenging cases; the one shown here had the cable cut during measurements, which disconnected a current drive electrode and several voltage pickup electrodes. Since one outer electrode was disconnected, all current drive patterns involving that electrode, either as a source or sink, were impacted (Fig. 2–1a). The filter involves 6 steps.

Fig. 2.

Data Processing Workflow. Example shown highlights poor electrode contact occasionally observed in practice. 1a) Voltage filtering, 1b) Removed saturated IIVV patterns, 2) Impedance filtering through comparison to simulated impedances, 3) Average across 3 data files, 4) Average across repeat tests, 5) Extracting IIVV patterns, 6) Comparing tissue impedance spectra.

Step 1: Voltage Filtering

First, the filter detects disconnections or poor electrode contact by analyzing the average measured voltage of each of the 28 current drive patterns at 1 kHz, flagging those exceeding a 1 V threshold (Fig. 2–1a). The average measured voltage of the drive patterns at this frequency, ${\mu }_{il}$ of size $N_{ii}\times N_b$, is computed as:

$${\mu }_{il}={\frac{1}{N_e}\sum _{j=1}^{N_e}{v}_{ijkl}|}_{k=11}$$ (4)

where $v_{ijkl}$ is the matrix of complex voltage data with dimensions $N_{ii}\times N_e\times N_f\times N_b,N_{ii}=28$ is the number of current drive patterns, $N_e=32$ is the number of electrodes, $N_f=31$ is the number of logarithmically-spaced frequencies, and $N_b=3$ is the number of burst measurements. Note the 11^th frequency corresponds to 1 kHz.

The 1 V threshold corresponding to saturated voltage patterns from disconnected outer-drive electrodes in clinical cases were confirmed experimentally. Electrode flagging is performed at 1 kHz and extended to all frequencies for consistency in impedance calculations. Voltage measurements at 1 kHz are used due to strong agreement with simulations, reduced contact impedance effects, and minimal signal drop-off from parasitic capacitance. All voltage data associated with bad current drive electrodes ($i$) and a given burst ($l$) are set to NaN, i.e. across all the electrodes and the frequencies (see right side of Fig. 2–1a). Bad patterns are set to NaN instead of being removed to maintain the order of the IIVV patterns in the matrix across all bursts and test measurements. Thus, when IIVV measurements are averaged across the bursts or tests, if a pattern is NaN in one spectral measurement, the mean of the two valid bursts can be used. Patterns filtered in all measurements remain NaN in the final matrix. Loss of electrode contact typically results in saturation artifacts in all three burst measurements but not in repositioned repeats. Removing one current drive electrode eliminates 1,932 of 7,728 patterns (25%), two electrodes remove 3,588 (46%), and three remove 4,968 (64%), underscoring the need for full electrode contact.

Step 2: Impedance Filtering

In the second step, impedances are computed from the IIVV patterns by indexing the II’s in $v_{ijkl}$, extracting the corresponding VV voltage differences and dividing by the injected current, forming the IIVVZ matrix. IIVV patterns with impedance measurement-to-simulation ratios deviating by more than 10x (ratios >10 or <0.1) are set to NaN (Fig. 2–2). The simulated impedances are calculated using a Finite Element Method (FEM)-based electromagnetic simulation, described as the EIT forward model in Borsic et al. [38]. Specifically, the forward problem was solved once as an offline pre-processing step for a homogeneous solution of 0.2 S/m conductivity. The resulting simulated impedances are saved and ultimately used to compute the measurement-to-simulation ratios.

Step 3: Burst Averaging

In the third step (Fig. 2–3), valid IIVV impedance measurements are averaged across the burst measurements, $Z_{mklq}$ and computed as:

$$Z_{mkl}^B=\frac{1}{N_b}\sum _{q=1}^{N_b}{Z}_{mklq}$$ (5)

where $Z_{mkl}^B$ is a matrix with dimensions $N_p\times N_f\times N_r,Z_{mklq}$ is a matrix with dimensions $N_p\times N_f\times N_r\times N_b$, and $N_p=7728$ is the number of IIVV patterns and $N_r=3$ is the number of repeat tests.

Step 4: Repeat Test Averaging

In the fourth step, the processed IIVVZ matrices from each of the three repeat tests are averaged to obtain a single IIVVZ matrix for the tissue sample, defined as:

$$Z_{mk}^R=\frac{1}{N_r}\sum _{l=1}^{N_r}{Z}_{mkl}^B$$ (6)

where $Z_{mk}^R$ is a matrix with dimensions $N_p\times N_f$, combining the valid burst measurements $Z_{mkl}^B$ for each of the sample’s repeat tests (Fig. 2–4).

Step 5: Pattern Extraction

IIVV pattern selection occurs with the average IIVVZ matrix for each test, with different applications using different IIVV pattern subsets (Fig. 2–5). For EIS, using all IIVV patterns results in high impedance variability due to the different geometries of each pattern. To ensure consistency, sixteen geometrically standardized IIVV patterns were selected: outer electrodes positioned opposite each other (blue) and adjacent voltage pickup electrodes (red) (Fig. 1-A and 2-5). With long distances between the voltage pickup electrodes, these patterns measure greater differences in induced voltages and are sensitive to both shallow and deep tissues. The valid IIVV patterns from these 16 “long across” patterns are averaged to obtain a single impedance spectrum for each tissue sample (Fig. 2-6), defined as:

$$LA=\left\{\mathit{index}\mathit{set}\right\}\in \{1\dots 7728\}$$ (7)

$$Z_k^{LA}=\frac{1}{N_{LA}}\sum _{m=1}^{N_{LA}}{Z}_{mk}^R$$ (8)

where $Z_k^{LA}$, a 1 × 31 vector, represents the final impedance spectra used in data analysis and sample classification, with LA defined as the valid set of 16 “long across” IIVV patterns.

Note that while not explored here, other geometrically similar IIVV patterns could be used to interrogate different tissue depths. For example, more closely-space voltage electrode pairs would be more sensitive to shallower tissues only. Likewise, for EIT applications, all valid IIVV patterns from the repeat test measurement matrices could be used for data processing. By maintaining the 7728 × 31 IIVVZ with the use of NaNs, the geometry information from the order of the IIVV patterns is preserved for image reconstruction.

Step 6: Compare Sample Impedances

Data from each sample (tissue and saline) is processed through this pipeline, resulting in a single impedance spectrum ($Z_k^{LA}$) per site measured, typically one on a lesion site, one on a healthy control site, and one of saline (Fig. 2-6). In the following sections, the notation for $Z_k^{LA}$ is simplified as Z and represents the impedance spectra for each measured sample.

F. Electrical Impedance Parameters

From each filtered impedance spectra (i.e., Fig 2.6), several impedance-based parameters were extracted for tissue classification. Tissue-dependent differences were explored across resistance, R, reactance, X, as well as the ratios of the resistance and reactance, X/R, and the impedance magnitude, $\left|Z\right|$, at each discrete frequency. In addition, ratios between the magnitude of the impedance spanning a 10x frequency difference (e.g. $\left|Z\right|@100\mathrm{H}\mathrm{z}/\left|Z\right|@1\mathrm{k}\mathrm{H}\mathrm{z}$) were computed to capture the impedance changes over frequency.

Each subject’s healthy and lesion impedance spectra were also fit to the well-established Cole model [39], [40], which models the 31-frequency spectra as an equivalent circuit with four tissue-specific parameters:

$$Z=R_{\mathrm{\infty }}+\frac{\mathrm{\Delta }R}{1+(j\omega \tau {)}^{\alpha }}$$ (9)

defined by a high-frequency resistance (${\mathrm{R}}_{\mathrm{\infty }}$), a change in resistance ($\mathrm{\Delta }\mathrm{R}$), a characteristic frequency (${\mathrm{f}}_{\mathrm{c}}=1/2\pi \tau$), and an alpha term, $\left(\alpha \right)$. Parameters were constrained to values that are meaningful for the biophysical representation of the equivalent circuit of the tissue: ${\mathrm{R}}_{\mathrm{\infty }}$ was constrained to positive integers and $\alpha$ was constrained to values between 0 and 1.

Finally, we fit the impedance magnitude and phase data to a logarithmic quadratic function:

$$y=\beta +\phi \mathrm{*}\mathrm{l}\mathrm{o}\mathrm{g}(x)+\gamma \mathrm{*}\mathrm{l}\mathrm{o}\mathrm{g}{(x)}^2$$ (10)

where $\beta ,\phi$, and $\gamma$ are coefficients of the model and referred to as the term they precede. For ease of reference, ${\beta }_Z,\mathrm{l}\mathrm{o}\mathrm{g}(Z)$ and $\mathrm{l}\mathrm{o}\mathrm{g}(Z{)}^2$ will be used to denote the coefficients when impedance magnitude data is the dependent variable, $x$, and ${\beta }_{\varphi },\mathrm{l}\mathrm{o}\mathrm{g}(\varphi )$ and $\mathrm{l}\mathrm{o}\mathrm{g}(\varphi {)}^2$ when phase is used. This approach represents an alternative spectral parameterization that was empirically observed to match the spectral data well.

In addition to these long across pattern parameters, a single constant complex conductivity was estimated at each frequency from all IIVV patterns using an iterative algorithm that models the electrical field distribution with an FEM-based mesh of the probe’s geometry [36]. This more comprehensive approach utilizes all 7728 measurement patterns rather than just the subset of 16 long across patterns employed to extract the above parameters, thereby incorporating the maximum available information to enhance the accuracy and robustness of the estimates. While direct conversion to resistivity through use of the saline calibration represents a possible computationally efficient approach, the full FEM-based optimization more accurately accounts for noise within the conductivity estimate, making it a more robust method despite the substantial computational requirements. The complex conductivity is defined as:

$${\sigma }^{\mathrm{*}}=\sigma +j\omega {\epsilon }_0{\epsilon }_r,$$ (11)

with the real conductivity component, $\sigma$, and an imaginary component consisting of the angular frequency, $\omega$, free space permittivity, ${\epsilon }_0$, and relative permittivity, ${\epsilon }_r$. Real and imaginary components of conductivity are reported in $\mathrm{S}/\mathrm{m}$ with $\epsilon$ representing $\omega {\epsilon }_0{\epsilon }_r$. This best fit method assumes a homogeneous conductivity in the domain unlike typical EIT reconstructions that estimate complex conductivity at each voxel of a mesh. The technique utilized is described as “constant EIT” in Murphy et al. [41].

G. Statistical Analysis

Mann-Whitney U tests were used to evaluate statistical differences between each of the unpaired parameters for healthy vs lesion measurements as not all frequencies had normally distributed data. The median and interquartile range (IQR) of both tissue types were computed for each parameter. Receiver operator characteristic (ROC) curve analysis was then applied to the frequencies of impedance-based metrics and spectral parameters with highest significance from the hypothesis tests. The ROC curves were used to estimate the area under the curve (AUC), sensitivity (SN), specificity (SP), and accuracy for each parameter, with Youden’s index utilized to select the optimal point from the ROC curve [42]. Positive likelihood ratio (+LR) and negative likelihood ratio (−LR) were used to measure the diagnostic accuracy of each impedance parameter. The variability of test measurements for each sample was quantified as the standard deviation of the impedances recorded during 1) the three-impedance spectra “burst” measurement and 2) each of the three repeat tests, where the probe was lifted from the tissue between measurements. Test-retest reliability of impedance measurements was assessed using a two-way mixed effects intraclass correlation coefficient (ICC) model for absolute agreement. ICC was calculated across all frequency points for both burst and test protocols using a mixed-effects ANOVA framework with subjects as random effects and measurement occasions as fixed effects. Overall ICC values were obtained by averaging across all valid measurement combinations to provide a comprehensive measure of measurement consistency.

III. RESULTS

A. Description of Data

Twenty-six patients undergoing OSCC resection were enrolled in the study with median age 64.3 years, IQR 59–70, comprising of 42% females (11). Two patients did not have healthy measurements, and one patient did not have lesion measurements. Additionally, two healthy measurements were excluded due to poor data quality, as determined by the filtering process. Furthermore, four lesion measurements were excluded because they did not correspond to cancerous tissue; one was benign and three were dysplasia. The final analysis included data from twenty-five patients, with twenty-two healthy samples and twenty-one lesion samples. On average, the filter removed 17% of the IIVV patterns for the healthy samples, 5.1% of the lesion samples, and 2.9% of the reference saline samples. The average number of patterns used for analysis was 14.2 for the healthy samples, 15.8 for the lesion samples, and 15.8 for the saline samples (out of 16).

B. Impedance Results

Cancerous tissue exhibited significantly lower reactance and resistance compared to healthy tissue (Table I and Figs. 3A–B). Specifically, cancerous tissue had significantly lower resistance (p < 0.05) at frequencies from 100 Hz to 6 kHz and 50 kHz to 100 kHz (see Fig. 3A), significantly lower reactance magnitude from 100 Hz to 40 kHz (Fig. 3B), and a significantly lower impedance magnitude from 100 Hz to 16 kHz. No statistically significant differences were observed between cancer and healthy tissue for resistance ranging from 8 kHz to 40 kHz and for reactance from 50 kHz to 100 kHz. The X/R ratio was significantly different between healthy and cancerous tissue across all frequencies (p < 0.005). Table I provides the median and IQR of the impedance parameters at the frequency with the highest AUC.

TABLE I.

IMPEDANCE-BASED METRICS FOR HEALTHY AND MALIGNANT TISSUE

Parameters	Healthy	Cancer	p-value
R 160Hz	3.9 [1.8–5.9] kΩ	1.2 [0.79–1.7] kΩ	0.0001
X 1kHz	−1.2 −[0.4–1.9] kΩ	−0.1 −[0.09–0.4] kΩ	<0.0001
|Z| 160Hz	3.9 [1.8–6.0] kΩ	1.2 [0.79–1.7] kΩ	0.0001
ϕZ 4kHz	−40 −[22–46]°	−12 −[11–17]°	<0.0001
X/R 16kHz	0.5 [0.31–0.69]	0.17 [0.13–0.27]	<0.0001
fratio 2.5/25kHz	3.1 [2.1–4.0]	1.4 [1.3–1.7]	<0.0001
ΔR	4.5 [2.3–7.6] kΩ	1.3 [0.9–2.1] kΩ	0.001
fc	1.4 [0.8–3.4] kHz	10.7 [2.3–19] kHz	0.0004
α	0.64 [0.59–0.72]	0.54 [0.41–0.61]	0.06
R ∞	0 [0–61] Ω	0 [0–0] Ω	0.70
log(Z)	0.25 [0.11–0.43]	0.13 [0.05–0.23]	0.17
log(Z)2	7.7 [7.1–8.6]	6.6 [6.4–7.1]	0.0003
βZ	−0.03 −[0.02–0.05]	−0.02 −[0.01–0.02]	0.003
log(ϕ)	0.76 [0.6–0.9]	0.12 [−0.2–0.4]	0.0002
log(ϕ)2	−0.86 −[0.2–1.6]	0.26 [−0.9–1.6]	0.03
βϕ	−0.03 −[0.01–0.04]	0.02 [0.00–0.03]	0.0002
σ 100 Hz	0.03 [0.03–0.09] S/m	0.14 [0.08–0.20] S/m	0.0001
ε 40kHz	0.11 [0.09–0.13] S/m	0.05 [0.04–0.06] S/m	<0.0001

Median [IQR] reported for each metric. R: Resistance, X: Reactance, |Z|: Impedance magnitude, ϕ_Z: Phase, X/R: Reactance to Resistance ratio, f_ratio: Frequency ratio, ΔR: Change in resistance, f_c: Characteristic frequency, α: alpha term, R_∞: Resistance at high frequency, log(Z): Coefficient of the log of impedance magnitude, log(Z)²: Coefficient of log of impedance magnitude squared, β_z: y-intercept of logarithmic impedance model log(ϕ):Coefficient of the log of phase, log(ϕ)²:Coefficient of log of phase squared, β_ϕ:y-intercept of logarithmic phase model, σ: Best fit conductivity, ε: Best fit permittivity.

Fig. 3.

Resistance and reactance of healthy vs cancer over frequency. A) Resistance measurements shown at 11 of the 31 frequencies, B) Reactance measurements at same frequencies. Center lines indicate median, triangles indicate mean, boxes show the IQR, whiskers shown the range, circles indicate outliers, ***p<0.001, **p<0.005, *p<0.01.

The resistance and reactance spectra generally captured the dispersion characteristics of the tissue (Fig. 1 in supplementary information) including the low and high frequency plateaus along with the transition region defined by the characteristic frequency. The reactance vs. resistance curves yielded the semicircle relationship typically observed in tissues exhibiting Cole-like features (Fig. 2 in supplementary information). All healthy and malignant tissue samples were fit to the Cole model, yielding median root mean square error (RMSE) values of 113 Ω and 76 Ω, respectively. A discontinuity in the recorded tissue spectra was observed between 1 kHz and 1.5 kHz across all tissue measurements. This jump in reactance values has been reproduced measuring discrete loads and is due to a limitation in the hardware system. To improve model fitting, the data point at 1.2 kHz was removed from all samples. Additionally, six spectra were truncated to adequately fit the Cole model (Fig. 2 in supplementary information). The estimated value for $R_{\mathrm{\infty }}$ was zero for many of the tissue samples; this most likely arises from the parasitic capacitances associated with the meter-long cable and inherent in the data acquisition system, decreasing the high frequency resistance. Statistical analysis revealed significant differences in the $\mathrm{\Delta }\mathrm{R},{\mathrm{f}}_{\mathrm{c}}$, and $\alpha$ parameters (p-values: 0.001, 0.0004, and 0.06, respectively), while no significant differences were observed for the $R_{\mathrm{\infty }}$ parameter.

The quadratic logarithmic model was selected as it closely matched the observed impedance trends in both healthy and malignant tissue samples (see supplementary information Fig. 3). The logarithmic model of the impedance magnitude data produced average RMSE values of 97 Ω for healthy samples and 14 Ω for cancerous samples. Significant differences were found in the ${\beta }_Z$ and $\mathrm{l}\mathrm{o}\mathrm{g}(Z{)}^2$ terms (p-values: 0.003 and 0.0003, respectively), whereas no significant differences were observed for $\mathrm{l}\mathrm{o}\mathrm{g}(Z)(p=0.17)$. The phase-based model exhibited statistically significant difference for all parameters, $\mathrm{l}\mathrm{o}\mathrm{g}(\varphi )$, $\mathrm{l}\mathrm{o}\mathrm{g}(\varphi {)}^2$, and ${\beta }_{\varphi }$ (p-values: 0.0002, 0.03, and 0.0002, respectively).

Burst measurements had lower relative standard deviations compared to test-retest measurements (supplementary information Fig. 4). Median relative standard deviations for healthy tissue were 3.0% for burst and 7.2% for test-retest measurements. For lesions, the median was 4.1% for burst and 7.8% for test-retest. Saline measurements had minimal variation (0.62% for burst and 1.1% for test). Outliers, defined as being >1.5*IQR, were found in all groups except healthy test-retest measurements, with one outlier exceeding 30% likely due to probe placement variation. Outliers were not removed from the data, just indicated in plots. Intraclass correlation coefficient analysis demonstrated excellent test-retest reliability across impedance measurements. The reported ICC values represent averages across all thirty-one frequency points analyzed. For burst measurements, ICC values were 0.993 for healthy tissue, 0.996 for cancerous tissue, and 0.975 for saline measurements. Test measurements yielded ICC values of 0.943 for healthy tissue, 0.927 for cancerous tissue, and 0.825 for saline measurements. All tissue ICC values exceeded 0.9, indicating excellent measurement reliability according to established interpretive guidelines [43]. The relatively lower ICC values observed in saline measurements can be attributed to the narrow dynamic range of impedance values in this medium, where small measurement variations have a proportionally greater impact on reliability metrics. Despite this limitation, the saline test ICC of 0.83 still demonstrates good repeatability. Representative ICC scatter plots at 10 kHz for all measurement combinations across healthy, cancerous, and saline conditions for both burst and test protocols are presented in Supplemental Fig. 5 and 6.

The best fit conductivity at all 31 frequencies was found for each saline measurement taken during the clinical measurement procedure (Fig. 7 in supplementary information). Given the purely resistive nature of saline solutions, the estimated conductivity should be purely resistive. However, in practice a drop-off in magnitude occurs at higher frequencies as a phase shift is introduced. This is due to nonidealities in the hardware system, such as parasitic capacitances and resistances from extended cables. A calibration factor was applied to the saline complex conductivities to correct for hardware nonidealities. The calibration factor from the saline conductivity for each case was then applied to the respective tissue conductivities.

Dropoff in estimated conductivity in saline solutions occurred past 10 kHz as a result of parasitic capacitances decreasing the signal amplitude. Added resistances from cables and connectors resulted in the best fit to underestimate conductivities. The mean calibration factor, $\kappa$, was $1.5-j0.02$ at 100 Hz and 0.64 – j1.0 at 100 kHz (where $\kappa =Z_{\mathit{sim}}/Z_{\mathit{meas}}$), demonstrating the increasing phase shift in the data. The median calibrated conductivity of healthy tissue ranged from $0.04+j0.00\mathrm{S}/\mathrm{m}[0.03+j0.00-0.09+j0.01]$ at 100 Hz to $0.33+j0.08[0.28+j0.06–0.42+j0.13]$ at 100 kHz. The median calibrated conductivity for malignant tissue is $0.14+j0.01\mathrm{S}/\mathrm{m}[0.08+j0.00-0.20+j0.01]$ at 100 Hz and $0.32+j0.04\mathrm{S}/\mathrm{m}[0.28+j0.00-0.41+j0.06]$ at 100 kHz. Tissue conductivities across all frequencies are shown in Fig 4 and in Table 1 in supplementary information. There were statistically significant differences in the conductivity of healthy vs malignant tissue from 100 Hz – 25 kHz and significant differences in permittivity from 1.6 kHz – 100 kHz (p <0.05).

Fig. 4.

Median conductivity of healthy (blue) and cancerous tissue (orange) with IQR shown in the surrounding band.

C. Classification Results

Table II displays the AUC, sensitivity (SN), specificity (SP), positive likelihood ratio (+LR), negative likelihood ratio (−LR), and accuracy (Acc) for the impedance parameters. The highest AUC (0.88) was achieved by classifying tissue using permittivity ($\epsilon$) at 40 kHz with many more parameters also achieving AUCs > 0.85. Likewise, $\epsilon$ at 40 kHz was the best performing tissue classifier with highest combined SN (0.86) and SP (0.90) and an 88% accuracy. The X/R, ${\mathrm{f}}_{\text{ratio}},{\mathrm{f}}_{\mathrm{c}},{\beta }_{\varphi }$ and conductivity parameters all had high sensitivities $>0.86$. Over half of the variables tested achieved accuracies $>80\mathrm{\%}$; however, ${\mathrm{f}}_{\mathrm{c}},R_{\mathrm{\infty }}$, and $\mathrm{l}\mathrm{o}\mathrm{g}(\mathrm{Z})$ variables performed poorest, none of which exceeded 65% accuracy. The normalized distributions and ROC curves of the five parameters with the highest accuracies ($\epsilon ,{\beta }_{\varphi },\mathrm{l}\mathrm{o}\mathrm{g}(\varphi ),\mathrm{l}\mathrm{o}\mathrm{g}(Z{)}^2$, and ${\mathrm{f}}_{\text{ratio}}$) are shown in Fig. 5.

TABLE II.

CLASSIFICATION RESULTS OF METRICS

Parameters	AUC	SN	SP	+LR	−LR	Acc
R 160Hz	0.85	0.73	0.90	7.6	0.30	0.81
X 1kHz	0.86	0.73	0.90	7.6	0.30	0.81
|Z| 160Hz	0.85	0.73	0.90	7.6	0.30	0.81
ϕZ 4kHz	0.86	0.67	0.88	5.6	0.38	0.79
X/R 16kHz	0.87	0.95	0.71	3.3	0.06	0.84
fratio 2.5/25kHz	0.86	0.86	0.81	4.5	0.17	0.84
ΔR	0.82	0.73	0.90	7.6	0.30	0.81
fc	0.77	0.86	0.14	1.0	0.95	0.51
α	0.74	0.82	0.62	2.1	0.29	0.72
R ∞	0.54	0.32	0.81	1.7	0.84	0.56
log(Z)	0.62	0.5	0.81	2.6	0.62	0.65
log(Z)2	0.82	0.95	0.71	3.3	0.06	0.84
βZ	0.76	0.22	0.76	0.92	1.0	0.77
log(ϕ)	0.83	0.82	0.90	0.90	1.9	0.86
log(ϕ)2	0.70	1	0	n/a	0	0.51
βϕ	0.83	0.90	0.82	1.1	0.55	0.86
σ 100 Hz	0.85	0.86	0.73	3.1	0.20	0.79
ε 40kHz	0.88	0.86	0.90	9.1	0.15	0.88

Median [IQR] reported for each metric. R: Resistance, X: Reactance, |Z|: Impedance magnitude, ϕ_Z: Phase, X/R: Reactance to Resistance ratio, f_ratio: Frequency ratio, ΔR: Change in resistance, f_c: Characteristic frequency, α: alpha term, R_∞: Resistance at high frequency, log(Z): Coefficient of the log of impedance magnitude, log(Z)²: Coefficient of log of impedance magnitude squared, β_Z: y-intercept of logarithmic impedance model log(ϕ):Coefficient of the log of phase, log(ϕ)²:Coefficient of log of phase squared,β_ϕ:y-intercept of logarithmic phase model, σ: Best fit conductivity, ε: Best fit permittivity.

Fig. 5.

A) Top performing normalized metrics, center lines indicate median, triangles indicate mean, boxes show the IQR, whiskers shown the range, circles indicate outliers, ***p<0.001, **p<0.005, p<0.01. B) ROC of top performing parameters.

IV. DISCUSSION

This study demonstrates the potential of using a newly designed non-invasive multi-electrode impedance probe to collect in vivo electrical impedance measurements of oral lesions. The primary objectives of this study were to design and validate this novel device. The handheld probe is easy to manufacture and assemble with a 3D printer and readily available connector components. A data processing workflow was established to extract high quality data from measurements acquired in an uncontrolled setting like the operating room. The devices are highly robust, yielding consistent impedance measurements after over twenty hydrogen peroxide gas plasma sterilization cycles. As expected, nonidealities in the hardware were observed in the decreased magnitude of the impedance measurements at high frequencies, which resulted in decreased statistically significant differences between resistance and reactance of healthy and malignant tissue at high frequencies. The observed phase shift results from the presence of parasitic capacitances that degrade SNR quality at frequencies > 10 kHz. However, utilizing a calibration factor can correct the phase shift for more accurate conductivity estimates and unimpacted frequencies can be used for tissue classification.

To the best of our knowledge, this paper is the first to report conductivity values for in vivo oral tissue and cancer. Oral cancer has significantly higher conductivity than healthy oral tissue ($0.04+j0.00$ vs $0.14+j0.01$ S/m at 100 Hz), resulting in significantly lower impedance measurements for the malignant tissue.

Overall, the results show that OSCC can be differentiated from healthy mucosa using probe-sampled EIS measurements. Electrical impedance of tissue is influenced by cellular morphology, tissue architecture, extracellular pathways, and the cellular microenvironment. As squamous cell carcinoma develops, abnormal cells in the epithelial lining begin to invade the basement membrane into underlying connective tissue (lamina propria). These invading cells may create more extracellular pathways for current to flow, decreasing overall resistance of the tissue. Previously published ex vivo and in vivo studies have reported similar trends with larger impedances associated with healthy tissue as compared to oral cancer [31], [32], [35]. The seven top parameters analyzed here demonstrated equivalent or higher AUCs (0.88 vs 0.78) and better discrimination of cancerous oral lesions from healthy oral mucosa as compared to those reported in Murdoch, et al [32]. However, Murdoch et al. performed classification on high-risk vs low-risk lesions using impedance difference between the lesion and healthy measurement. Our multielectrode device with the developed filter provides robust impedance measurements that can better handle the challenges of in vivo oral cavity data collection, as seen by our 4% invalid data rate compared to their nearly 10% rate. The high ICC values indicate strong measurement reproducibility and minimal variability between repeat measurements, further supporting the reliability of the data collection methodology.

In this study, we also introduce a new model for parameterizing an impedance spectrum, the quadratic logarithmic model. This model fits impedance data with lower RMSE than the Cole model and can include high frequency data.

Our findings indicate that the permittivity is a particularly robust classification parameter, exhibiting a high AUC value of 0.88, along with high sensitivity (0.86), specificity (0.9), and accuracy (0.88). The quadratic logarithmic model using the phase of the impedance yields two parameters ($\mathrm{l}\mathrm{o}\mathrm{g}(\varphi )$ and ${\beta }_{\varphi }$) with 86% accuracy in classifying cancerous tissue. X/R ratio (AUC = 0.87, SN = 0.95, SP = 0.71, Acc = 0.84) and the ${\mathrm{f}}_{\text{ratio}}$ at 2.5/25 kHz (AUC = 0.86, SN = 0.86, SP = 0.81, Acc = 0.84) parameters also reported high classification performance. In our investigation, the highest positive likelihood ratio obtained was 9.1, utilizing the tissue permittivity. Notably, the parameters exhibiting the most favorable −LRs were the X/R and the log(Z)² metric, yielding −LRs of 0.06. The low −LR values validate the device’s capacity to accurately discern the absence of cancer and underscore its promising clinical utility as a rapid, non-invasive, and non-ionizing tool for clinical assessment of oral lesions.

Interestingly, our EIS approach shows promising performance compared to OCT, which recent meta-analyses indicate has a sensitivity of 0.91, specificity of 0.91, and AUC of 0.95 for head and neck cancer margin assessment [44]. While the permittivity at 40 kHz parameter (SN=0.86, SP=0.90, AUC=0.88) has slightly lower overall AUC than OCT, it can be used in the oral cavity as opposed to on the ex vivo specimen, making it valuable for future margin assessment applications. A further advantage of this device is that it achieves a greater depth of penetration than OCT (3–5 mm ([45], [46]) vs 1–2mm ([15]), where [45] demonstrated a ~4.5 mm depth of sensitivity for a probe with similar maximum electrode separation and [46] specifically found this probe was able to detect changes in conductivity 3–4 mm deep experimentally, and 5–6 mm deep in simulation.

When all patterns are incorporated, the best-fit complex-conductivity was found to be highly accurate (${\mathit{\epsilon }}_{40kHz}\mathrm{A}\mathrm{c}\mathrm{c}=0.88$); however, it is not time efficient, taking 5 minutes to compute the complex conductivity at a single frequency per spectrum. Computation time could be decreased through use of a higher speed CPU, parallel computing, or precomputation of the required Jacobian. Conversely, processing the samples with all other long across impedance parameters takes only ~0.5 seconds and the same laptop. The processing time of this code can also be improved to achieve faster classification if needed.

A limitation of this study is the relatively small sample size. Due to the moderate number of patients in this study, impedance was measured from a variety of sub-sites within the oral cavity and oropharynx having different tissue structures which may impact baseline impedance. Unlike Murdoch et al [32], who grouped measurements by tissue type (buccal, gingival, tongue), our analysis did not separate by tissue site due to the limited sample size. However, our preliminary analysis showed no statistically significant differences between tissue types, and it represents an added benefit that our approach can achieve high discrimination AUCs without requiring site-specific analysis within the oral cavity. A larger sample size would enable different lesion locations and non-malignant lesions to be analyzed separately for a more detailed evaluation of tissue properties. Subjects continue to be recruited, and in future studies the larger sample size could be used for multi-featured classification. The surgeon aimed to take measurements from the center of the tumors to avoid the edges where dysplastic morphologies are often present, but for small lesions that may not have been possible. Other factors, such as tissue dehydration and pooled blood around the lesion were mitigated with saline rinses before measurements. Despite many of the uncontrolled variables within the intraoperative setting, the initial clinical results are encouraging, demonstrating this technology’s potential in clinically evaluating tissues accessed from the oral cavity. Finally, this study aimed to explore the differences between healthy and malignant oral tissues and did not explicitly evaluate surgical margins. Given the positive findings reported here, evaluation of impedance-based assessment of surgical margin status could be considered as part of a future investigation.

V. CONCLUSION

This study presents a novel electrical impedance sensing probe capable of intraoperatively discriminating between healthy and malignant oral tissues, demonstrating the strong potential of impedance-based classification for oral cancer assessment. In addition to achieving high sensitivity, specificity, and AUC in tissue classification, this work introduces a new quadratic logarithmic model based on impedance magnitude and phase, which effectively captures tissue-specific electrical properties and supports robust classification performance. Furthermore, we developed a comprehensive data processing and filtering pipeline that successfully removes poor-quality measurements caused by external intraoperative disturbances, resulting in a highly reliable system for real-time data acquisition and analysis. Notably, this is the first study to report in vivo conductivity and permittivity values of oral cancer tissue and to achieve successful classification using absolute impedance values—a key advancement enabled by the robust sensing and processing framework established here. By demonstrating high sensitivity to cancerous tissue, this work represents a critical step toward the overarching goal of using impedance spectroscopy for intraoperative positive margin detection. These findings lay the groundwork for future studies with larger sample sizes to confirm these results, explore machine learning classification approaches, and ultimately evaluate surgical margins with this device to improve patient outcomes.