Study design

We propose an exploratory cross-sectional evaluation of hypoglossal nerve function and swallowing among OPC survivors along the cancer care continuum. The project will involve primary data collection of tongue function procedures among an active cohort (the MD Anderson Oropharynx Cohort, “MDA-OPC,” protocol #PA14–0947). Participant recruitment will occur from June 2024 to May 2026. Preliminary results of the primary objective are expected in 2027. PROs, survey responses (PSS-HN, MDASI-HN, MDADI), functional outcomes (IOPI, EIM), EMG recordings, and _HDSEMG recordings of consenting patients will be stored on the parent cohort study (MDA #PA14–0947). Analyses herein do not overlap with the work of the parent cohort.

Population

The population will include patients evaluated at the University of Texas MD Anderson Head and Neck Center with a suspected or confirmed pathologic diagnosis of squamous cell carcinoma (SCCA) of the oropharynx, tonsil, or base of the tongue, or of head and neck SCCA of unknown primary. Spanish-speaking participants and pregnant women may be enrolled. In this exploratory analysis, patients will be enrolled regardless of radiation dosage.

Inclusion criteria

• Patients evaluated at MD Anderson Cancer Center with pathologic diagnosis of SCCA of the oropharynx, tonsil, or base of the tongue, or SCCA of the head and neck of unknown primary who consent to EMG testing.
• If consented post-treatment, there is no evidence of disease

Exclusion criteria

• Primary surgery for OPC.
• Documented CN XII nerve injury by tumor, surgery, or other source. (Confounds radiation-induced CN XII injury.)
• History of prior head and neck cancer or central nervous system cancer.
• Functionally limiting cardiac, pulmonary, or neuromuscular disease.
• Current tracheostomy.
• History of surgery near the hypoglossal nerve path.
• Platelet count <50,000 cells/mL.
• Patient < 18 years of age.
• Cognitive impairment (unable to complete questionnaires).

Objectives

Data collection

Participants consenting to the tongue testing procedures on MDA protocol #PA14–0947 will participate in measurement at baseline and/or 3–6-, 18–24-, 60-month data collection time points including: a) intramuscular EMG with or without nerve conduction study (NCS), b) tongue surface electrical impedance myography (EIM) and _HDSEMG, c) a video recording of Lingual range of motion (L-ROM), and d) tongue strength. L-ROM and strength will be assessed using published, validated methods [21,22]. The IOPI digital manometer will measure maximum isometric lingual strength to assess maximum pressures generated by the tongue. Together, the proposed measures will characterize the nature of nerve activity, the composition of the tongue, and its function. Tongue testing procedures require a clinical appointment in the neurology center. They are expected to take about 30 minutes for EMG and up to 30 minutes, including breaks between tasks, for tongue function measures. During the visit, the study team will assign each subject an accession number that corresponds to all their deidentified data at every time point.

Needle electromyography and nerve conduction

Intramuscular EMG will assess insertional activity in the genioglossus muscle, innervated by CN XII, as a marker of denervation [23]. Needle EMG is reported to have an inter-operator reliability of 0.53 [24]. EMG recordings will be conducted by a cancer neurologist fellowship trained in clinical neurophysiology and denervation potentials graded per standard neurologist’s grading:

0 - None

1 - Persistent, single train of potentials in at least two areas

2 - Moderate number of potentials in three or more areas

3 - Many potentials in all areas

4 - Full interference patterns of potentials

EMG and NCS may also be performed in the trapezius muscle (innvervated by CN XI), as the region is easily accessible for non-invasive testing and represents a muscle within the irradiated field with lower cranial nerve innervation. Both quantitative and qualitative EMG may be assessed [25]. Needle EMG will not be conducted if platelets are < 50,000. The intramuscular EMG will require a clinical visit with the neurologist.

Tongue electrical impedance myography

This brief, non-invasive procedure will measure electrical conductivity in the tongue as a proxy for tongue composition. The patient’s experience is equivalent to placing a tongue depressor briefly on the tongue. The EIM system comprises a battery-powered recording unit called the user tongue electronic system (UTES) that couples to the user tongue array (UTA) depressor, custom-fabricated by Dr. Sanchez’s lab [15,16,26,27]. A smartphone is paired with the EIM system via Wi-Fi, and a custom software application allows the user to initiate the measurement, display the results, and save de-identified data.

The UTES transmits an imperceptible electrical current during a measurement through the UTA depressor. The voltage generated by the current flowing through the tongue is measured on all remaining surface electrodes. The alternating current frequency ranges from 8 kHz to 256 kHz. Across all frequencies, the current amplitude is limited to 0.1 mA, well within the safety requirements for applied current specified in the International Electrotechnical Commission (IEC) 60601−1 safety standard for medical devices. Primary EIM outcomes will include resistance, reactance, phase, conductivity, and permittivity at selected frequencies.

The UTA depressor will be for single-patient use. Each patient will have a dedicated depressor, which will be air-dried before being stored in an approved, labeled biospecimen bag between time points while not in use. The UTES will be stored with the clinical research staff when not in use and wiped with alcohol before and immediately after use. All data from the device will be automatically de-identified and sent to Dr. Sanchez’s lab for processing and analysis.

The underlying premise is that directional changes in disease tongue EIM signatures will provide new information on tongue health [16,17]. The system results from ongoing research efforts from collaborator Dr. Sanchez and has similar functionality and safety performance to clinical systems for impedance in other body regions. The system has been used to measure the tongue of seven healthy subjects [16]. We have conducted previous work that demonstrates the preliminary feasibility of EIM to distinguish neuropathic and healthy tongues [17]. It is approved as an investigational device for human use under IRB approval at Beth Israel Deaconess Medical Center, Boston, MA (Clinicaltrials.gov Identifier: NCT02118805). The University of Utah has submitted a provisional patent disclosure for tongue technology.

High-density surface electromyography

_HDSEMG will be measured similarly to the EIM (as described above using the UTA/UTES system). The recording system will be a commercial TMSi SAGA, 32-channel _HDSEMG instrument (SAGA32, Twente Medical Systems; Oldensaal, The Netherlands). This investigational device will comply with all FDA regulations under 21 CFR 812. The investigators will comply with monitoring and record keeping, and the device will not be marketed or promoted. The surface electrodes in the tongue depressor will allow us to measure the distribution of electrical activity over a confined tongue area. The electrode grid consists of 32 electroless nickel-gold contacts of 2 mm diameter. The inter-electrode distance is 2.5 mm. The electrode array connects via a clamshell to a commercial _HDSEMG amplifier, which records signals at a 4 kHz sampling rate. The unipolar channels have RMS noise < 1.5 µV, input impedance > 1 GΩ, and an analog bandwidth of 0–800 Hz.

Each patient will clean and prepare their tongue for the exam by gently wiping it with gauze and drinking a small amount of water. The arrays will be prepped with an alcohol pad and allowed to dry before tasks are performed. The patient will lie supine on a procedure table, and the electromyographer will begin the left tongue relaxation task. The electromyographer (CA) will evaluate needle EMG for spontaneous activity, and then the depressor will be placed on the left tongue at rest for 30 seconds to evaluate spontaneous _HDSEMG activity simultaneously with needle EMG. For the voluntary MUP potential analysis, the patient will protrude the tongue at approximately 20% of maximum force, and the electromyographer will evaluate EMG during the voluntary contraction. Then the tongue depressor will be placed on the left tongue to record 30 seconds of voluntary muscle activity with EMG and _HDSEMG simultaneously. The electromyographer will then coach the patient to perform an isometric force ramp, increasing tongue force against the roof of the mouth from zero to maximum over about 10 seconds. These tasks are then repeated in the right tongue. If the patient swallows or speaks during the task, the measurement will be repeated. Table 1 summarizes the total _HDSEMG data collected. The duration of discharges and the frequency of potentials will be measured at all electrode locations to confirm _HDSEMG data stability and to reveal tongue fasciculations and recruitment patterns.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Surface electromyography measurement duration breakdown. Measurements will be recorded in the left tongue and right tongue.

https://doi.org/10.1371/journal.pone.0347891.t001

Needle EMG and _HDSEMG procedures will be performed during the same visit. An electromyographer (CA) will evaluate hypoglossal needle EMG submentally to provide a clinical assessment of CN XII health. The patient will then perform the tasks we proposed while the needle is still inserted and _HDSEMG is collected simultaneously. Finally, the patient will repeat our proposed tasks with only the _HDSEMG electrodes on the tongue.

Tongue strength and range of motion

Brief, non-invasive testing will measure tongue function, including strength and range of motion, using published methods. A video recording of L-ROM and strength will be assessed using published, validated methods. The IOPI digital manometer will measure maximum isometric lingual strength to assess the maximum pressures generated by the tongue (inter-operator reliability of 0.75) [28,29].

Clinician graded lymphedema and fibrosis

Clinician-graded soft tissue palpation sores in the bilateral submental, anterior, and lateral neck will be recorded using the Head and Neck External Lymphedema and Fibrosis (HN-LEF), which has a reported inter-operator reliability of 0.91 [30,31]. This tool uses ordinal scoring to evaluate the type (A-D) and severity grade (1–5) of soft tissue changes.

Data management and ethics

The data collected for this study are part of a larger cohort effort to collect comprehensive outcomes data from consenting patients in the Oropharynx program at the MD Anderson Cancer Center in Houston, Texas (MDA-OPC). The collection is approved under MD Anderson IRB #PA14–0947 for MD Anderson Oropharynx Cohort patients who consent to tongue testing procedures (optional procedure #2 on the parent consent form). Patients scheduled for appointments in the Head and Neck Center will be identified by either an attending physician or an appropriately trained designee. During clinic appointments, potential subjects will be informed of their eligibility and asked if they would be interested in participating in research. In the privacy of the exam room, physicians or designees with appropriate training will discuss all aspects of the study with potential subjects and answer any questions. Subjects will be given a copy of the Informed Consent and further instructed about the study and the elements of the consent document. The subject will be given as much time as needed to review and discuss the study and consent, and the research nurse, research data coordinator, attending physician, or designee will answer any remaining questions. Non-English-speaking patients will consent to use an MD Anderson Translator and a verbal translation preparatory sheet, in accordance with MD Anderson policies for Informed Consent translation, in that patient’s primary language. Patients may withdraw from the study at any time without any penalty. The MD Anderson Institutional Review Board classifies this study as minimal risk and does not monitor this study for adverse events. The PI is responsible for monitoring adverse events. Risk will be evaluated by the electromyographer upon review of the patient’s medical history and medication.

Since 2015, 1,750 OPC patients have enrolled with detailed, interdisciplinary characterization of exposures, diagnostic/staging, treatment, disease control, and longitudinal collection of validated clinician- and patient-reported survivorship outcomes from diagnosis through 5-year survival. All longitudinal MDA-OPC clinical data is securely stored in REDCap® and DMI, an institutionally secured Data Management Services platform. Our DMI-based MDA-OPC database contains 14 data forms, 1182 imported variables, 439 manually entered variables, and 18 customized reports, with nightly updates from integrated data sources. Anyone interested in analyzing the collected data can submit a request to the database oversight committee (oropharynxprogrampro@mdanderson.org). Functional outcomes, PROs, and electrodiagnostic data will be de-identified and provided to the Sanchez Research Lab for analysis under MD Anderson IRB #PA14–0947. Code and models will be stored in a public GitHub repository within the group “SanchezResearchLab/hdsemg_pipeline.” Because the software is not complete, the source code will be made publicly available when the analysis is complete.

Outcomes definition

Data analysis and machine learning plan

Clinical measures will be used to confirm the diagnosis and grade of neuropathy. We expect a pilot sample size of 36 patients with needle EMG and _HDSEMG to provide ample statistical power to develop a machine-learning classification algorithm, since acquiring 2 minutes and 20 seconds of EMG data at 10 kHz will yield 50.4 million values (36 pts x 140 sec x 10 kHz). For 32-channel _HDSEMG at 4 kHz (the threshold for the SAGA device), that is 645.12 million values (36 patients x 140 sec x 4 kHz x 32 EMG simultaneous channels). Still, we acknowledge our limitations in determining the optimal sample sizes, as we do not yet have _HDSEMG data in OPC patients. If the initial analysis shows insufficient performance, we will assess EMG and _HDSEMG subsets to further increase the training and test dataset sizes. Fig 1 depicts the general data processing pipeline that will translate _HDSEMG measures into CN XII neuropathy grade. Signals will be preprocessed using MATLAB R2023a software (MathWorks; Natick, MA). Experimental _HDSEMG will be denoised using a bandpass filter from 20–500 Hz. The signal quality assessment for each task consists of the following. First, a rolling mean absolute value of each channel (window size 400 samples) will represent the apparent effort exerted by the patient.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Tongue _HDSEMG data processing pipeline.

Tongue _HDSEMG recordings from three patient tasks will be decomposed via DWT, and time-frequency features will be selected using linear discriminant analysis. Voluntary contraction tasks will be decomposed into motor unit spike train estimates using MUEdit, and a CNN will be trained on the resulting MU estimates. A final multitask neural network will make a final CN XII neuropathy prediction from all the data. Abbreviations: _HDSEMG, High-density surface electromyography; MU, Motor unit; CNN, Convolutional neural network; WL, Waveform length; ZC, Zero crossings; SSC, Sign slope changes; RMS, Root-mean-square amplitude; MAV, Mean absolute value; IAV, Instantaneous absolute value; AMB, Amplitude-modulated bandwidth; FMB, Frequency-modulated bandwidth; SMPSD, Spectral moment of power spectral density; MFD, Mean first derivative of instantaneous frequency; MIF, mean instantaneous frequency; CN XII, Cranial nerve XII.

https://doi.org/10.1371/journal.pone.0347891.g001

• Relaxation: None are rejected because motion artifacts appear similar to fasciculations on _HDSEMG.
• Protrusion: Protrusion task channels will be rejected if the range of the apparent effort curve is larger than 2 standard deviations of the signal mean absolute value.
• Isometric: Isometric task channels will be rejected if the Pearson’s Correlation of the apparent effort curve shows a non-positive slope or a positive slope with r-squared < 0.2.

All accepted signals are divided into 2-second segments for feature extraction. The Discrete Wavelet Transform (DWT) will be used to extract time series features (waveform length, zero crossings, slope signal changes, root mean square, autoregressive coefficient, mean absolute value, and integrated mean absolute value) and time-frequency features (mean instantaneous frequency, mean first derivative, amplitude-modulated bandwidth, frequency-modulated bandwidth, and spectral moment of power spectral density) [32]. We will perform Mann—Whitney U-tests with α = 0.05 to determine whether these features are relevant to neuropathy through the following comparisons: reduced recruitment vs. no reduced recruitment; polyphasia vs. no polyphasia; increased amplitude vs. no increased amplitude; and increased duration vs. no increased duration. To further demonstrate the clinical relevance of _HDSEMG features, we will use Pearson’s Correlation Coefficients and multiple regression analysis to express correlations between _HDSEMG features and tongue functional outcomes (L-ROM and IPOI pressure).

The full dataset will be partitioned patient-wise into 70% for training, 20% for development, and 10% for validation. Features from the training set that discriminate neuropathy severity will be selected using linear discriminant analysis. Missing data may occur during time-frequency feature extraction, as some calculations can trigger divide-by-zero errors, especially for low-frequency components. If this occurs, such examples will be rejected as artifacts. If any class label is represented by less than 10% of the data, recording-wise mean imputation will be performed to augment the number of unique examples in the underrepresented classes.

We intend to train an initial decision tree and support vector machine (SVM), followed by ensemble methods using both trees and SVMs, where possible: Random Forest Classifier, Gradient Boost, and Bagging. We will also train a neural network classifier with a softmax activation function for multiclass classification. Hyperparameter optimization will be performed using 5-fold cross-validation and verified using the development set. To account for potential class imbalance, all ensembled learning methods will be trained using class-weighted sampling to ensure that individual classifiers within the ensemble see class-balanced subsets of the data. Machine learning will be performed using SciKit Learn [33].

The 32-channel _HDSEMG signals are a convolutive mixture on an _HDSEMG recording and can be extracted using convolutive blind source separation (BSS). Statistical motor unit decomposition software (MUEdit, Avrillon et al.) [34] will be used to extract estimated spike trains from each suite of voluntary contraction _HDSEMG tasks. Spike train estimates include valuable MU features such as the number of MUs and discharge rate. We plan to construct shimmer plots for each of the 32 spatial recordings by aligning the identified MU estimate peaks [34]. These aligned MU examples will train a convolutional neural network (CNN), using MU peak morphology to predict neuropathy grade. The CNN will be built with PyTorch and trained with help from Optuna for hyperparameter optimization [35,36].

For the secondary objective, we will estimate the prevalence of neuropathy at three time points (pre-treatment, 3–6 months after treatment, and 18 months to 5 years after treatment), along with exact 95% binomial confidence intervals (CIs). Logistic regression models will be fit with XII neuropathy as a binary response variable adjusted for sampling weights and clinical covariates. For hypothesis 1.1, we will model time as a binary variable (0 = pre-RT and 1 = post-RT), pooling the two post-RT groups (n = 24) versus the pre-RT group (n = 12). Time will be included as a continuous variable (i.e., months pre- or post-RT). We expect an exploratory sample of 36 patients who consent to tongue testing, with equal allocation across 12 patients in one of three groups based on the timing of EMG collection relative to their cancer treatment: pre-treatment, early post-treatment, and late post-treatment. Twelve patients per group ensure a maximum half-width of the 95% CI of 0.29. After treatment, we expect the prevalence of neuropathy to be approximately 40%, based on our published data that 10% of survivors develop clinically detectable neuropathy [9,37]. Unpublished data from the NIH intramural laryngeal study section found that 90% of chronic radiation-associated dysphagia cases have EMG-detectable denervation [38]. The investigators’ data show that 100% of patients with DIGEST grade ≥2 dysphagia who are ≥ 2 years disease-free HNC survivors have EMG detectable neuropathy as early as 2 years post-RT (38% of whom have subclinical neuropathy). Logistic regression on a binary independent variable (pre- vs. post-RT) with a sample size of 36 (n = 12 pre- and n = 24 post-RT) achieves 80% power (α = 0.10) to detect a difference in the probability of neuropathy from 5% pre-RT to 42% post-RT.

Evaluation criteria

The performance of machine learning classifiers will be divided into two classification settings: binary (CN XII neuropathy vs. no confirmed CN XII neuropathy) and multiclass (neuropathy grade). In binary classifiers, we aim for AUC ≥ 0.70 to support _HDSEMG as an inexpensive, non-invasive quantitative metric for screening OPC survivors for CN XII neuropathy. Narayanaswami et al. report that electromyographers with 10–20 years of experience have a diagnostic sensitivity of 0.77 and a specificity of 0.71 for neuropathy. [24] This will serve as our benchmark for reliable performance for all binary model development. For the 10% validation set, we will measure model performance simply by the percentage of correctly diagnosed cases of neuropathy. As a secondary validation outcome, we will compare the _HDSEMG neuropathy severity grade with the gold-standard needle EMG grades.

Multiclass classification will use micro- and macro-F1 scores to evaluate the validity of a refined CN XII neuropathy grade prediction. Saliency mapping for the CNN will reveal which signal regions contribute to each neuropathy grade prediction. Machine learning models will be evaluated using SHapley Additive exPlanations (SHAP) to investigate which features contribute to the model predictions [39].

Patient and public involvement

Patients and the public were not involved in the design. However, the investigators have recently engaged patient research partners for a program of research on delayed oropharyngeal cancer sequelae and expect to involve participants in the implementation and dissemination of the research plan, including consultation on recruitment/data collection strategies and on integrating these technologies into clinical trials.

Dissemination strategy

Research data collection began in June 2024. Data collection will continue through May 2027, and we are targeting December 2027 for completion of data analysis. We will perform preliminary analyses and report progress in the annual RPPRs associated with this study’s NIH funding. We expect to submit our complete findings to a suitable journal by March 2028. We will disseminate the findings of our analysis to the scientific and medical community through publication in a peer-reviewed journal and presentation at relevant conferences for head and neck surgery, electrodiagnostics, biomedical engineering, and machine learning.