Blood-based screening for HPV-associated cancers.

Background: HPV-associated oropharyngeal cancer (HPV+OPSCC) is the most common HPV-associated cancer in the United States yet unlike cervical cancer lacks a screening test. HPV+OPSCCs are presumed to start developing 10-15 years prior to clinical diagnosis. Circulating tumor HPV DNA (ctHPVDNA) is a sensitive and specific biomarker for HPV+OPSCC. Taken together, blood-based screening for HPV+OPSCC may be feasible years prior to diagnosis. Methods: We developed an HPV whole genome sequencing assay, HPV-DeepSeek, with 99% sensitivity and specificity at clinical diagnosis. 28 plasma samples from HPV+OPSCC patients collected 1.3-10.8 years prior to diagnosis along with 1:1 age and gender-matched controls were run on HPV-DeepSeek and an HPV serology assay. Results: 22/28 (79%) of cases and 0/28 controls screened positive for HPV+OPSCC with 100% detection within four years of diagnosis and a maximum lead time of 7.8 years. We next applied a machine learning model classifying 27/28 cases (96%) with 100% detection within 10 years. Plasma-based PIK3CA gene mutations, viral genome integration events and HPV serology were used to orthogonally validate cancer detection with 68% (19/28) of the cohort having multiple cancer signals detected. Molecular fingerprinting of HPV genomes was performed across patients demonstrating that each viral genome was unique, ruling out contamination. In patients with tumor blocks from diagnosis (15/28), molecular fingerprinting was performed within patients confirming the same viral genome across time. Conclusions: We demonstrate accurate blood-based detection of HPV-associated cancers with lead times up to 10 years before clinical cancer diagnosis and in doing so, highlight the enormous potential of ctDNA-based cancer screening.

An External Validation Study for Automated Segmentation of Vestibular Schwannoma.

OBJECTIVE: To validate how an automated model for vestibular schwannoma (VS) segmentation developed on an external homogeneous dataset performs when applied to internal heterogeneous data. PATIENTS: The external dataset comprised 242 patients with previously untreated, sporadic unilateral VS undergoing Gamma Knife radiosurgery, with homogeneous magnetic resonance imaging (MRI) scans. The internal dataset comprised 10 patients from our institution, with heterogeneous MRI scans. INTERVENTIONS: An automated VS segmentation model was developed on the external dataset. The model was tested on the internal dataset. MAIN OUTCOME MEASURE: Dice score, which measures agreement between ground truth and predicted segmentations. RESULTS: When applied to the internal patient scans, the automated model achieved a mean Dice score of 61% across all 10 images. There were three tumors that were not detected. These tumors were 0.01 ml on average (SD = 0.00 ml). The mean Dice score for the seven tumors that were detected was 87% (SD = 14%). There was one outlier with Dice of 55%-on further review of this scan, it was discovered that hyperintense petrous bone had been included in the tumor segmentation. CONCLUSIONS: We show that an automated segmentation model developed using a restrictive set of siloed institutional data can be successfully adapted for data from different imaging systems and patient populations. This is an important step toward the validation of automated VS segmentation. However, there are significant shortcomings that likely reflect limitations of the data used to train the model. Further validation is needed to make automated segmentation for VS generalizable.

A Proof-of-Concept Computer Vision Approach for Measurement of Tympanic Membrane Perforations.

OBJECTIVE: Size, an important characteristic of a tympanic membrane perforation (TMP), is commonly assessed with gross estimation via visual inspection, a practice which is prone to inaccuracy. Herein, we demonstrate feasibility of a proof-of-concept computer vision model for estimating TMP size in a small set of perforations. METHODS: An open-source deep learning architecture was used to train a model to segment and calculate the area of a perforation and the visualized tympanic membrane (TM) in a set of endoscopic images of mostly anterior and relatively small TMPs. The model then computed relative TMP size by calculating the ratio of perforation area to TM area. Model performance on the test dataset was compared to ground-truth manual annotations. In a validation survey, otolaryngologists were tasked with estimating the size of TMPs from the test dataset. The primary outcome was the average absolute error of model size predictions and clinician estimates compared to sizes determined by ground-truth manual annotations. RESULTS: The model's average absolute error for size predictions was a 0.8% overestimation for all test perforations. Conversely, among the 38 survey respondents, the average clinician error was a 11.0% overestimation (95% CI, 5.2-16.7%, p = 0.003). CONCLUSIONS: In a small sample of TMPs, we demonstrated a computer vision approach for estimating TMP size is feasible. Further validation studies must be done with significantly larger and more heterogenous datasets. LEVEL OF EVIDENCE: N/A Laryngoscope, 2024.

Current Practices in Voice Data Collection and Limitations to Voice AI Research: A National Survey.

INTRODUCTION: Accuracy and validity of voice AI algorithms rely on substantial quality voice data. Although commensurable amounts of voice data are captured daily in voice centers across North America, there is no standardized protocol for acoustic data management, which limits the usability of these datasets for voice artificial intelligence (AI) research. OBJECTIVE: The aim was to capture current practices of voice data collection, storage, analysis, and perceived limitations to collaborative voice research. METHODS: A 30-question online survey was developed with expert guidance from the voicecollab.ai members, an international collaborative of voice AI researchers. The survey was disseminated via REDCap to an estimated 200 practitioners at North American voice centers. Survey questions assessed respondents' current practices in terms of acoustic data collection, storage, and retrieval as well as limitations to collaborative voice research. RESULTS: Seventy-two respondents completed the survey of which 81.7% were laryngologists and 18.3% were speech language pathologists (SLPs). Eighteen percent of respondents reported seeing 40%-60% and 55% reported seeing >60 patients with voice disorders weekly (conservative estimate of over 4000 patients/week). Only 28% of respondents reported utilizing standardized protocols for collection and storage of acoustic data. Although, 87% of respondents conduct voice research, only 38% of respondents report doing so on a multi-institutional level. Perceived limitations to conducting collaborative voice research include lack of standardized methodology for collection (30%) and lack of human resources to prepare and label voice data adequately (55%). CONCLUSION: To conduct large-scale multi-institutional voice research with AI, there is a pertinent need for standardization of acoustic data management, as well as an infrastructure for secure and efficient data sharing. LEVEL OF EVIDENCE: 5 Laryngoscope, 134:1333-1339, 2024.

Proof-of-Concept Computer Vision Model for Instrument and Anatomy Detection During Transcanal Endoscopic Ear Surgery.

High-definition video captured during transcanal endoscopic ear surgery (TEES) can serve as imaging data for computer vision algorithms. This report describes a proof-of-concept model for automated anatomy and instrument detection during TEES.

Towards Medical Billing Automation: NLP for Outpatient Clinician Note Classification.

Objectives: Our primary objective was to develop a natural language processing approach that accurately predicts outpatient Evaluation and Management (E/M) level of service (LoS) codes using clinicians' notes from a health system electronic health record. A secondary objective was to investigate the impact of clinic note de-identification on document classification performance. Methods: We used retrospective outpatient office clinic notes from four medical and surgical specialties. Classification models were fine-tuned on the clinic notes datasets and stratified by subspecialty. The success criteria for the classification tasks were the classification accuracy and F1-scores on internal test data. For the secondary objective, the dataset was de-identified using Named Entity Recognition (NER) to remove protected health information (PHI), and models were retrained. Results: The models demonstrated similar predictive performance across different specialties, except for internal medicine, which had the lowest classification accuracy across all model architectures. The models trained on the entire note corpus achieved an E/M LoS CPT code classification accuracy of 74.8% (CI 95: 74.1-75.6). However, the de-identified note corpus showed a markedly lower classification accuracy of 48.2% (CI 95: 47.7-48.6) compared to the model trained on the identified notes. Conclusion: The study demonstrates the potential of NLP-based document classifiers to accurately predict E/M LoS CPT codes using clinical notes from various medical and procedural specialties. The models' performance suggests that the classification task's complexity merits further investigation. The de-identification experiment demonstrated that de-identification may negatively impact classifier performance. Further research is needed to validate the performance of our NLP classifiers in different healthcare settings and patient populations and to investigate the potential implications of de-identification on model performance.

Machine Learning Analysis of Physical Activity Data to Classify Postural Dysfunction.

BACKGROUND: Machine learning (ML) analysis of biometric data in non-controlled environments is underexplored. OBJECTIVE: To evaluate whether ML analysis of physical activity data can be employed to classify whether individuals have postural dysfunction in middle-aged and older individuals. METHODS: A 1 week period of physical activity was measured by a waist-worn uni-axial accelerometer during the 2003-2004 National Health and Nutrition Examination Survey sampling period. Features of physical activity along with basic demographic information (42 variables) were paired with ML models to predict the success or failure of a standard 30 s modified Romberg test during which participants had their eyes closed and stood upon a 3-inch compliant surface. Model performance was evaluated by area under the receiver operating characteristic curve (AUC-ROC), balanced accuracy, and F1-score. RESULTS: The cohort was comprised of 1625 participants ≥40 years (median age 61, IQR 51-71). Approximately half (47%) were diagnosed with postural dysfunction having failed the binarized (pass/fail) scoring mechanism of the modified Romberg exam. Five ML models were trained on the classification task, achieving AUC values ranging from 0.67 to 0.73. The support vector machine (SVM) and a gradient-boosted model, XGBoost, achieved the highest AUC of 0.73 (SD 0.71-0.75). Age was the most important variable for SVM classification, followed by four features that evaluated accelerometer counts at various thresholds, including those delineating total, moderate, and moderate-vigorous activity. CONCLUSIONS: ML analysis of accelerometer-derived physical activity data to classify postural dysfunction in middle-aged and older individuals is feasible in real-world environments such as the home. LEVEL OF EVIDENCE: 3 Laryngoscope, 133:3529-3533, 2023.

Development of a Predictive Model for Persistent Dizziness Following Vestibular Schwannoma Surgery.

OBJECTIVE: In an era of vestibular schwannoma (VS) surgery where functional preservation is increasingly emphasized, persistent postoperative dizziness is a relatively understudied functional outcome. The primary objective was to develop a predictive model to identify patients at risk for developing persistent postoperative dizziness after VS resection. METHODS: Retrospective review of patients who underwent VS surgery at our institution with a minimum of 12 months of postoperative follow-up. Demographic, tumor-specific, preoperative, and immediate postoperative features were collected as predictors. The primary outcome was self-reported dizziness at 3-, 6-, and 12-month follow-up. Binary and multiclass machine learning classification models were developed using these features. RESULTS: A total of 1,137 cases were used for modeling. The median age was 67 years, and 54% were female. Median tumor size was 2 cm, and the most common approach was suboccipital (85%). Overall, 63% of patients did not report postoperative dizziness at any timepoint; 11% at 3-month follow-up; 9% at 6-months; and 17% at 12-months. Both binary and multiclass models achieved high performance with AUCs of 0.89 and 0.86 respectively. Features important to model predictions were preoperative headache, need for physical therapy on discharge, vitamin D deficiency, and systemic comorbidities. CONCLUSION: We demonstrate the feasibility of a machine learning approach to predict persistent dizziness following vestibular schwannoma surgery with high accuracy. These models could be used to provide quantitative estimates of risk, helping counsel patients on what to expect after surgery and manage patients proactively in the postoperative setting. LEVEL OF EVIDENCE: 4 Laryngoscope, 133:3534-3539, 2023.

Making Use of Artificial Intelligence-Generated Synthetic Tympanic Membrane Images.

This diagnostic study examines the application of generative artificial intelligence in clinical tool research and development.

Paediatric sleep apnea event prediction using nasal air pressure and machine learning.

Sleep-disordered breathing is an important health issue for children. The objective of this study was to develop a machine learning classifier model for the identification of sleep apnea events taken exclusively from nasal air pressure measurements acquired during overnight polysomnography for paediatric patients. A secondary objective of this study was to differentiate site of obstruction exclusively from hypopnea event data using the model. Computer vision classifiers were developed via transfer learning to either normal breathing while asleep, obstructive hypopnea, obstructive apnea or central apnea. A separate model was trained to identify site of obstruction as either adeno-tonsillar or tongue base. In addition, a survey of board-certified and board-eligible sleep physicians was completed to compare clinician versus model classification performance of sleep events, and indicated very good performance of our model relative to human raters. The nasal air pressure sample database available for modelling comprised 417 normal, 266 obstructive hypopnea, 122 obstructive apnea and 131 central apnea events derived from 28 paediatric patients. The four-way classifier achieved a mean prediction accuracy of 70.0% (95% confidence interval [67.1-72.9]). Clinician raters correctly identified sleep events from nasal air pressure tracings 53.8% of the time, whereas the local model was 77.5% accurate. The site of obstruction classifier achieved a mean prediction accuracy of 75.0% (95% confidence interval [68.7-81.3]). Machine learning applied to nasal air pressure tracings is feasible and may exceed the diagnostic performance of expert clinicians. Nasal air pressure tracings of obstructive hypopneas may "encode" information regarding the site of obstruction, which may only be discernable by machine learning.

Generation of synthetic tympanic membrane images: Development, human validation, and clinical implications of synthetic data.

Synthetic clinical images could augment real medical image datasets, a novel approach in otolaryngology-head and neck surgery (OHNS). Our objective was to develop a generative adversarial network (GAN) for tympanic membrane images and to validate the quality of synthetic images with human reviewers. Our model was developed using a state-of-the-art GAN architecture, StyleGAN2-ADA. The network was trained on intraoperative high-definition (HD) endoscopic images of tympanic membranes collected from pediatric patients undergoing myringotomy with possible tympanostomy tube placement. A human validation survey was administered to a cohort of OHNS and pediatrics trainees at our institution. The primary measure of model quality was the Frechet Inception Distance (FID), a metric comparing the distribution of generated images with the distribution of real images. The measures used for human reviewer validation were the sensitivity, specificity, and area under the curve (AUC) for humans' ability to discern synthetic from real images. Our dataset comprised 202 images. The best GAN was trained at 512x512 image resolution with a FID of 47.0. The progression of images through training showed stepwise "learning" of the anatomic features of a tympanic membrane. The validation survey was taken by 65 persons who reviewed 925 images. Human reviewers demonstrated a sensitivity of 66%, specificity of 73%, and AUC of 0.69 for the detection of synthetic images. In summary, we successfully developed a GAN to produce synthetic tympanic membrane images and validated this with human reviewers. These images could be used to bolster real datasets with various pathologies and develop more robust deep learning models such as those used for diagnostic predictions from otoscopic images. However, caution should be exercised with the use of synthetic data given issues regarding data diversity and performance validation. Any model trained using synthetic data will require robust external validation to ensure validity and generalizability.

Using Machine Learning to Predict Operating Room Case Duration: A Case Study in Otolaryngology.

OBJECTIVE: Optimizing operating room (OR) efficiency depends on accurate case duration estimates. Machine learning (ML) methods have been used to predict OR case durations in other subspecialties. We hypothesize that ML methods improve projected case lengths over existing non-ML techniques for otolaryngology-head and neck surgery cases. METHODS: Deidentified patient information from otolaryngology surgical cases at 1 academic institution were reviewed from 2016 to 2020. Variables collected included patient, surgeon, procedure, and facility data known preoperatively so as to capture all realistic contributors. Available case data were divided into a training and testing data set. Several ML algorithms were evaluated based on best performance of predicted case duration when compared to actual case duration. Performance of all models was compared by the average root mean squared error and mean absolute error (MAE). RESULTS: In total, 50,888 otolaryngology surgical cases were evaluated with an average case duration of 98.3 ± 86.9 minutes. Most cases were general otolaryngology (n = 16,620). Case features closely associated with OR duration included procedure performed, surgeon, subspecialty of case, and postoperative destination of the patient. The best-performing ML models were CatBoost and XGBoost, which reduced operative time MAE by 9.6 minutes and 8.5 minutes compared to current methods, respectively. DISCUSSION: The incorporation of other easily identifiable features beyond procedure performed and surgeon meaningfully improved our operative duration prediction accuracy. CatBoost provided the best-performing ML model. IMPLICATIONS FOR PRACTICE: ML algorithms to predict OR case time duration in otolaryngology can improve case duration accuracy and result in financial benefit.

Predicting Progression of Oral Lesions to Malignancy Using Machine Learning.

OBJECTIVE: To use large-scale electronic health record (EHR) data to develop machine learning models predicting malignant transformation of oral lesions. METHODS: A multi-institutional health system database was used to identify a retrospective cohort of patients with biopsied oral lesions. The primary outcome was malignant transformation. Chart review and automated system database queries were used to identify a range of demographic, clinical, and pathologic variables. Machine learning was used to develop predictive models for progression to malignancy. RESULTS: There were 2192 patients with a biopsied oral lesion, of whom 1232 had biopsy proven oral dysplasia. There was malignant transformation in 34% of patients in the oral lesions dataset, and in 54% of patients in the dysplasia subset. Multiple machine learning-based models were trained on the data in two experiments, (a) including all patients with biopsied oral lesions and (b) including only patients with biopsy-proven dysplasia. In the first experiment, the best machine learning models predicted malignant transformation among the biopsied oral lesions with an area under the curve (AUC) of 86%. In the second experiment, the random forest model predicted malignant transformation among lesions with dysplasia with an AUC of 0.75. The most influential features were dysplasia grade and the presence of multiple lesions, with smaller influences from other features including anemia, histopathologic description of atypia, and other prior cancer history. CONCLUSION: With diverse features from EHR data, machine learning approaches are feasible and allow for generation of models that predict which oral lesions are likely to progress to malignancy. LEVEL OF EVIDENCE: 3 Laryngoscope, 133:1156-1162, 2023.

"Human vs Machine" Validation of a Deep Learning Algorithm for Pediatric Middle Ear Infection Diagnosis.

OBJECTIVE: We compared the diagnostic performance of human clinicians with that of a neural network algorithm developed using a library of tympanic membrane images derived from children taken to the operating room with the intent of performing myringotomy and possible tube placement for recurrent acute otitis media (AOM) or otitis media with effusion (OME). STUDY DESIGN: Retrospective cohort study. SETTING: Tertiary academic medical center from 2018 to 2021. METHODS: A training set of 639 images of tympanic membranes representing normal, OME, and AOM was used to train a neural network as well as a proprietary commercial image classifier from Google. Model diagnostic prediction performance in differentiating normal vs nonpurulent vs purulent effusion was scored based on classification accuracy. A web-based survey was developed to test human clinicians' diagnostic accuracy on a novel image set, and this was compared head to head against our model. RESULTS: Our model achieved a mean prediction accuracy of 80.8% (95% CI, 77.0%-84.6%). The Google model achieved a prediction accuracy of 85.4%. In a validation survey of 39 clinicians analyzing a sample of 22 endoscopic ear images, the average diagnostic accuracy was 65.0%. On the same data set, our model achieved an accuracy of 95.5%. CONCLUSION: Our model outperformed certain groups of human clinicians in assessing images of tympanic membranes for effusions in children. Reduced diagnostic error rates using machine learning models may have implications in reducing rates of misdiagnosis, potentially leading to fewer missed diagnoses, unnecessary antibiotic prescriptions, and surgical procedures.

Development of a Predictive Model for Individualized Hearing Aid Benefit.

OBJECTIVES: To develop a model to predict individualized hearing aid benefit. To provide interpretations of model predictions on global and individual levels. METHODS: We compiled a data set of patients with hearing loss who trialed hearing aids and completed the Client Oriented Scale of Improvement (COSI) questionnaire, a validated patient-reported outcome measure of hearing aid benefit. Features included demographic, medical, and audiological measures. The outcome was the COSI score for change in listening ability with hearing aids, scaled from 1 to 5. Model development was performed using fivefold cross-validation repeated three times with hyperparameter tuning. Model performance was assessed using the root mean squared error (RMSE) of the COSI scores. Model interpretation was performed using Shapley Additive Explanations. RESULTS: The data set comprised 1,286 patients across 3,523 listening situations. The best performing model was random forest with an RMSE of 0.80, found to be significantly better than the next best model (eXtreme gradient boosting with RMSE of 0.85, p < 0.01). The most important features in predicting hearing aid benefit were shorter duration of hearing aid use, higher pure-tone average in the better hearing ear, and younger age. CONCLUSION: We have developed a predictive model for hearing aid benefit that can also provide individualized explanations of model predictions. Predictive modeling could be a useful tool in assessing a patient's candidacy and predicted benefit from hearing aids.

New Medical Device and Therapeutic Approvals in Otolaryngology: State of the Art Review of 2021.

Objective: To evaluate new medical devices and drugs pertinent to otolaryngology-head and neck surgery that were approved by the Food and Drug Administration (FDA) in 2021. Data Sources: Publicly available FDA device and drug approvals from ENT (ear, nose, and throat), anesthesia, neurosurgery, plastic surgery, and general surgery FDA committees. Review Methods: FDA device and therapeutic approvals were identified and reviewed by members of the American Academy of Otolaryngology-Head and Neck Surgery's Medical Devices and Drugs Committee. Two independent reviewers assessed the relevance of devices and drugs to otolaryngologists. Medical devices and drugs were then allocated to their respective subspecialty fields for critical review based on available scientific literature. Conclusions: The Medical Devices and Drugs Committee reviewed 1153 devices and 52 novel drugs that received FDA approval in 2021 (67 ENT, 106 anesthesia, 618 general surgery and plastic surgery, 362 neurosurgery). Twenty-three devices and 1 therapeutic agent relevant to otolaryngology were included in the state of the art review. Advances spanned all subspecialties, including over-the-counter hearing aid options in otology, expanding treatment options for rhinitis in rhinology, innovative laser-safe endotracheal tubes in laryngology, novel facial rejuvenation and implant technology in facial plastic surgery, and advances in noninvasive and surgical treatment options for obstructive sleep apnea. Implications for Practice: FDA approvals for new technology and pharmaceuticals present new opportunities across subspecialties in otolaryngology. Clinicians' nuanced understanding of the safety, advantages, and limitations of these innovations ensures ongoing progress in patient care.

A systematic review of federated learning applications for biomedical data.

OBJECTIVES: Federated learning (FL) allows multiple institutions to collaboratively develop a machine learning algorithm without sharing their data. Organizations instead share model parameters only, allowing them to benefit from a model built with a larger dataset while maintaining the privacy of their own data. We conducted a systematic review to evaluate the current state of FL in healthcare and discuss the limitations and promise of this technology. METHODS: We conducted a literature search using PRISMA guidelines. At least two reviewers assessed each study for eligibility and extracted a predetermined set of data. The quality of each study was determined using the TRIPOD guideline and PROBAST tool. RESULTS: 13 studies were included in the full systematic review. Most were in the field of oncology (6 of 13; 46.1%), followed by radiology (5 of 13; 38.5%). The majority evaluated imaging results, performed a binary classification prediction task via offline learning (n = 12; 92.3%), and used a centralized topology, aggregation server workflow (n = 10; 76.9%). Most studies were compliant with the major reporting requirements of the TRIPOD guidelines. In all, 6 of 13 (46.2%) of studies were judged at high risk of bias using the PROBAST tool and only 5 studies used publicly available data. CONCLUSION: Federated learning is a growing field in machine learning with many promising uses in healthcare. Few studies have been published to date. Our evaluation found that investigators can do more to address the risk of bias and increase transparency by adding steps for data homogeneity or sharing required metadata and code.

A Systematic Review of Clinical Vestibular Symptom Triage, Tools, and Algorithms.

OBJECTIVE: The evaluation of peripheral vestibular disorders in clinical practice is an especially difficult endeavor, particularly for the inexperienced clinician. The goal of this systematic review is thus to evaluate the design, approaches, and outcomes for clinical vestibular symptom triage and decision support tools reported in contemporary published literature. DATA SOURCES: A comprehensive search of existing literature in August 2020 was conducted using MEDLINE, CINAHL, and EMBASE using terms of desired diagnostic tools such as algorithm, protocol, and questionnaire as well as an exhaustive set of terms to encompass vestibular disorders. REVIEW METHODS: Study characteristics, tool metrics, and performance were extracted using a standardized form. Quality assessment was conducted using a modified version of the Quality of Diagnostic Accuracy Studies 2 (QUADAS-2) assessment tool. RESULTS: A total of 18 articles each reporting a novel tool for the evaluation of vestibular disorders were identified. Tools were organized into 3 discrete categories, including self-administered questionnaires, health care professional administered tools, and decision support systems. Most tools could differentiate between specific vestibular pathologies, with outcome measures including sensitivity, specificity, and accuracy. CONCLUSION: A multitude of tools have been published to aid with the evaluation of vertiginous patients. Our systematic review identified several low-evidence reports of triage and decision support tools for the evaluation of vestibular disorders.