Artificial Intelligence Helps Otolaryngologists Give Excellent Patient Care

By the end of 2023, otolaryngologists in Europe may be using an artificial intelligence (AI) system to diagnose laryngeal cancer.

The Zeno AI system, which analyzes images in real time during laryngoscopy, is currently under evaluation by European regulatory authorities and may be commercially available in Europe as early as mid-2023, according to Marius Wellenstein, chief operations officer and co-founder of WSK Medical in Amsterdam, Netherlands, the company that developed Zeno AI.

“This is a real-time closed system that is able to detect and distinguish cancer or benign lesions in the larynx,” said Wellenstein, noting that the system may allow otolaryngologists who do not specialize in head and neck cancer to evaluate laryngeal lesions more effectively.

In testing, Zeno AI correctly localized and classified vocal cord carcinoma in 71% and 78% of the cases in the validation and test set, respectively, and 70% to 82% of benign vocal cord lesions, according to a clinical evaluation report provided by WSK Medical. The system is currently in use at two European academic medical centers.

Is this the future of otolaryngology? Will humans someday rely on computers to detect, diagnose, and direct the treatment of head and neck cancer, dysphagia, otitis media, balance disorders, sinusitis, and other disorders of the ears, nose, and throat?

Yes. And probably not.

Some studies have already found that “a physician with the assistance of an AI algorithm is able to perform better than a physician without the AI algorithm,” said Andrés M. Bur, MD, an associate professor of otolaryngology–head and neck surgery and director of robotic and minimally invasive head and neck surgery at the University of Kansas in Lawrence. “AI has the potential to reduce the likelihood of missing something or making an error.”

If we can develop machine learning models that are highly accurate, patients who are at low risk may not need to undergo any neck surgery. —Andrés M. Bur, MD

But AI is not intended to be a standalone diagnostic modality or disease management tool.

“AI’s role is not to replace physicians; it’s to augment our abilities,” said Aaron Moberly, MD, an associate professor in both the department of otolaryngology–head and neck surgery and the department of hearing and speech sciences at Vanderbilt University Medical Center in Nashville, Tenn.

Computers can sift through vast amounts of data far faster than humans can, so computer-driven systems can spot patterns and connections that may be difficult for humans to perceive. Human physicians, though, will remain an integral part of healthcare for the foreseeable future. For while many researchers are investigating potential applications of artificial intelligence in otolaryngology, no AI-based tools have yet been approved for clinical use in otolaryngology in the United States, in part because the field lacks large, labeled datasets that can be used to train computer algorithms. In addition, researchers, physicians, regulatory agencies, and the public aren’t yet confident that AI systems can be used safely for clinical care.

The field of artificial intelligence is moving quickly, however, and dozens, if not hundreds, of AI tools are currently under development and investigation. Matthew G. Crowson, MD, an assistant professor in otolaryngology–head and neck surgery at Harvard Medical School in Boston who is researching the use of AI in otology, believes that “we’re probably going to see AI processes baked into existing clinical workflows within the next two to three years.”

With this coming expansion, it’s important to take a look at what technologies are being developed and how AI may help you in your clinical practice.

Detecting, Diagnosing and Guiding Head and Neck Cancer Treatment

One of the most promising applications of AI in otolaryngology is the diagnosis of laryngeal cancer based on video laryngoscopy and histopathology slides, according to a 2022 review of literature (Otolaryngol Head Neck Surg. 2022 July 5; doi: 10.1177/01945998221110839).

Zeno AI is one such system, but certainly not the only one. Dr. Bur also helped develop an AI system to detect and identify laryngeal lesions on flexible laryngoscopy. The system performed well in testing, according to Dr. Bur. “We found very encouraging results with it being able to differentiate benign and malignant tumors,” he said. Dr. Bur and his team have submitted a paper detailing their work; at press time, it was pending peer review.

Researchers and clinicians hope that AI technology will lead to more effective and earlier detection of laryngeal squamous cell carcinoma—and that it will eventually increase the accessibility of laryngeal cancer screening and expertguided treatment for populations who currently lack access to specialist care.

It’s currently difficult to determine with great accuracy which oral lesions may progress to cancer, and which will not. Repeatedly performing biopsies on lesions is inconvenient and uncomfortable for patients, and is “not a very efficient way over the long term to arrive at a diagnosis,” Dr. Crowson said. He and his team created and tested an AI system that uses demographic, clinical, and pathological data to predict malignant transformation with greater than 70% accuracy (Laryngoscope. [published online ahead of print July 9, 2022]. doi:10.1002/lary.30285).

Future systems that include additional data points, such as genetic information, may be even more accurate. Such systems may eventually allow physicians to focus their clinical attention on patients who are most likely to require (and benefit from) further medical care, while patients at low risk of progression may be able to safely avoid repeated biopsies and frequent follow-ups.

Artificial intelligence systems may also help clarify which patients are at increased risk of lymph node metastasis and are therefore good candidates for elective neck dissection. Dr. Bur led a multi-institutional team that developed and validated an AI system that could predict occult nodal metastasis with a significantly higher degree of accuracy as compared with tumor depth thresholds that are commonly used in clinical practice. The system had a sensitivity of 91.7%, specificity of 72.6%, positive predictive value of 39.3%, and negative predictive value of 97.8% (JAMA Netw Open. 2022;5:e227226. doi:10.1001/jamanetworkopen. 2022.7226).

Jonathan Garneau, MD, assistant professor of otolaryngology–head and neck surgery at the University of Virginia School of Medicine in Charlottesville, is investigating whether artificial intelligence can use computerized tomography (CT) images to predict the risk of metastasis. Computers can detect patterns and features that the human eye cannot detect, so AI may help clinicians extract more meaningful information from a CT scan. This additional information may eventually allow physicians to more effectively determine which patients require surgery, and which can be treated with other modalities.

“If we can develop machine learning models that are highly accurate, patients who are at low risk may not need to undergo any neck surgery,” Dr. Bur added.

Diagnosing Infection and Predicting Aspiration with Vocal Data

During the COVID-19 pandemic, researchers wondered if it might be possible to use cough sounds to detect COVID-19 disease. Scientists at the Massachusetts Institute of Technology in Cambridge developed an AI model that attempted to distinguish people with asymptomatic COVID- 19 from healthy individuals through forced-cough recordings submitted via computers and cell phones (Chu J. Artificial intelligence model detects asymptomatic Covid-19 infections through cellphone-recorded coughs. MIT News. October 29, 2020), and in April 2022, Pfizer submitted a $75 million offer to purchase an Australian digital health company that created a smartphone-based app that can analyze the sound of a cough and diagnose respiratory diseases, including asthma, pneumonia, croup, and COVID-19 (Schuster-Bruce C. A health firm says it has developed an app that can detect COVID-19 when an infected person coughs into their phone. Insider. April 19, 2022). Pfizer finalized its acquisition of the company in a $116 million deal announced at the end of September 2022.

Anaïs Rameau, MD, a laryngologist, chief of dysphagia at the Sean Parker Institute for the Voice, and assistant professor and director of new technologies in the department of otolaryngology–head and neck surgery at Weill Cornell Medical College in New York City, has received the prestigious Paul B. Beeson Emerging Leaders Career Development Award in Aging from the National Institute on Aging to create an AI-based clinical decision support tool to improve the detection of aspiration risk at the bedside, using a combination of voice and cough sound and clinical and demographic data. “We’re looking at different aspects of their voice, such as changes in voice quality after drinking water, to determine whether that can help us distinguish who’s at risk for aspiration,” Dr. Rameau said, noting that speech–language pathologists aren’t always available to provide instrumental swallowing studies in the outpatient setting, such as nursing homes. The clinical decision support tool will be app-based and will be driven by AI algorithms to process complex multi-modal data and help improve the reliability of the bedside swallow screen.

We’re looking at different aspects of their voice, such as changes in voice quality after drinking water, to determine whether that can help us distinguish who’s at risk for aspiration. —Anaïs Rameau, MD

At present, her (and others’) ability to create clinically useful AI tools is limited by the availability of high-quality diverse datasets. Although clinicians and healthcare systems have amassed tremendous amounts of data in recent years, including vocal data and images, most of it is not currently labeled, collated, or accessible. That’s why Dr. Rameau is working with Yael Bensoussan, MD, director of the University of South Florida’s Health Voice Center in Tampa, and others on a National Institutes of Health (NIH)-funded research project to collect voice data and develop an AI-ready bioacoustics dataset. The project is part of the NIH Bridge to AI program, a $100 million investment in the future of medicine (Acosta CM. Artificial intelligence could soon diagnose illness based on the sound of your voice. NPR Illinois. October 10, 2022).

Diagnosing Otitis Media, Vestibular Disorders, and More

Researchers in otology are exploring the use of AI to diagnose otitis media, diagnose and manage vestibular disorders, optimize hearing aid technology, and predict sensorineural hearing outcomes (Otolaryngol Head Neck Surg. 2020;163:1123–1133).

“Diagnostics for ear imagery is a hot area for artificial intelligence in otology,” Dr. Crowson said. “Here at Mass Eye and Ear, we’ve developed an integrated algorithm into a handheld device that clips onto a smartphone to assist in the diagnosis of pediatric ear infections. It takes pictures of the child’s eardrum, and then the algorithm provides a diagnosis.”

The algorithm developed by Dr. Crowson’s team has a mean prediction accuracy rate of 80.8%. In a validation survey, 39 clinicians analyzed a sample set of 32 endoscopic images and achieved an average diagnostic accuracy of 65.%. The AI model achieved an accuracy of 95.5% (Otolaryngol Head Neck Surg. [published online ahead of print August 16, 2022]. doi:10.1177/01945998221119156). The diagnostic accuracy of otolaryngology generalists and subspecialists was similar to the model (79.2% for generalists and 81.8% for subspecialists, compared to 80.8% mean prediction accuracy for the AI model); however, the AI tool achieved greater accuracy than pediatricians (60.1%) and family or internal medicine physicians (59.1%), which suggests that AI-augmented tools may be particularly useful in urgent care and general medicine clinics.

AI-enhanced tools may eventually reduce rates of misdiagnosis of otitis media, which could also decrease unnecessary antibiotic prescriptions and surgical procedures. AI tools that can be used with smartphones may make it possible to “deliver high quality diagnostic interpretation anywhere in the world,” Dr. Crowson said.

Significant obstacles remain, though. Otolaryngology doesn’t yet have standardized guidelines for an otitis media classification system, or large, labeled image datasets, said Dr. Moberly, who is also working to develop AI software to aid in ear pathology diagnosis. Additionally, clinicians and patients are leery of AI diagnosis. “Right now, there’s no real check on the accuracy of my otoscopic exams and diagnosis,” Dr. Moberly said. “I think some people are hesitant to accept something that’s going to serve as a check on their ability to diagnose.”

Patients perceive “more accurate diagnosis” as a primary strength of AI, in part due to its ability to draw upon more data than humans, according to a 2021 systematic review (Lancet Digit Health. 2021;3:e599–e611). But they also say “less accurate diagnosis” is a primary weakness of AI and express concerns about computer-based tools missing diagnoses due to a lack of context or human experience. A study published in 2022 found that 31% to 40.5% of surveyed respondents would be “somewhat” or “very” uncomfortable receiving a diagnosis from an AI algorithm that was accurate 90% of the time but incapable of explaining its rationale (JAMA Netw Open. 2022:5:e2210309).

The so-called “black box” issue—the inability to determine how or why AI algorithms arrive at conclusions—is a major limitation to the adoption of AI in clinical practice. “Many of these machine learning models don’t provide a lot of transparency. We don’t really understand how they’re making decisions internally,” Dr. Bur said. He believes that additional testing and prospective trials demonstrating safety and efficacy are necessary before AI tools are introduced into clinical practice.

Developers working on an AI support algorithm to assist with vestibular disorder diagnosis performed a small randomized controlled trial to compare diagnoses generated by primary care physicians and neuro-otology specialists. One group of primary care providers diagnosed patients with persistent dizziness via traditional methods; another also used the AI support tool. Neuro-otology experts then assessed the same patients. Researchers found that the AI-assisted primary care diagnoses agreed with the expert diagnosis in 54% of cases, compared to just 41% of cases evaluated by primary care providers without AI assistance (J Neurol. 2022;269:2584–2598).

Diagnosing Sinusitis

Scientists have been studying the potential of AI to detect and diagnose sinusitis based on radiographs since at least 2019 (Quant Imaging Med Surg. 2019;9:942–951). A 2021 study noted that an AI algorithm showed diagnostic performance comparable to radiologists in diagnosing frontal, ethmoid, and maxillary sinusitis (Diagnostics (Basel). 2021;11:250. doi:10.3390/diagnostics11020250). To date, however, no AI rhinological applications have been validated in a real-world setting, leading the authors of a 2022 systematic review to conclude that “AI applications remain more theoretical than useful in day-to-day clinics” (Eur Arch Otorhinolaryngol. October 19, 2022; doi:10.1007/s00405-022-07701-3).

Additional research will almost surely reveal new uses—and limitations— of AI in otolaryngology. Physicians studying AI applications recommend caution. “When it comes to medical decision-making, these models have the potential to cause significant patient harm,” Dr. Bur said. “I think it’s imperative that physicians work carefully to protect our patients.”

---

Jennifer Fink is a freelance medical writer based in Wisconsin.