Machine Learning: The Future of Endoscopy and Laryngology

Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical field.  In this episode, we will talk about the advancements of Artificial Intelligence and Machine Learning in some common respiratory conditions. 

Did you know that artificial intelligence can help evaluate, improve, and personalize treatment in obstructive sleep apnea?¹

Obstructive sleep apnea, or OSA, is a common medical condition with an estimated prevalence rate of 5 to 14 percent; however, some recent studies suggest it might be as high as 50%.¹

Untreated OSA is associated with adverse health outcomes including cardiovascular disease, metabolic conditions, cerebrovascular events, cognitive impairment, and even motor vehicle accidents. The first-line, gold standard treatment for OSA is nasal continuous positive airway pressure, commonly known as CPAP, which has been shown to be effective when used with appropriate adherence. Nevertheless, accurate diagnosis is still a barrier to efficient treatment, while the current treatment selection and pathway is also a major contributor to the disease process.²

Since OSA is a complex and heterogeneous disorder that varies in symptomatology, physiologic etiology, comorbidities, and outcomes, it is also well suited for a personalized approach to treatment, which can be designed to target the underlying OSA pathophysiology, including instability of ventilator control, insufficient dilator muscle response, and low arousal threshold. Artificial intelligence can help predict treatment success, and evaluate current treatment and compliance, leading to a personalized treatment approach.² 

Currently, the apnea–hypopnea index, commonly referred to as AHI, remains the most used metric of OSA severity, despite lacking the ability to fully reflect the complexity and genetic and biological underpinnings of the disorder. However, machine learning can be applied to identify phenotypes or previously unidentified patterns and complement the development of alternative metrics to quantify and describe OSA, rather than solely by this index. Artificial intelligence tools could also help in the identification of mechanisms responsible for OSA, which could enable personalized treatment.² 

Regarding CPAP treatment, a recent study by Zhu and colleagues identified genes in OSA and CPAP treatment by utilizing machine learning algorithms and a genetic database. This predictive model suggests that individuals at high risk for OSA showed extensive activation of immune cells and pathways, and higher expression of these genes, which decreased after treatment. The information provided by machine learning in this setting can improve the identification of people with high risk of OSA as well as an insight into the individual treatment benefit.²

Machine learning can also help to increase the understanding of OSA physiology and etiology, and even the location of upper airway collapse, to subsequently improve treatment selection and outcomes. The study that evaluated this approach used a model using linear discriminants to analyze audio signals from snore sounds. The model demonstrated to have fair accuracy in discerning tongue and non-tongue collapse with an overall accuracy of 81% and 64% accuracy for all sites of collapse classes.²

As previously mentioned, the main challenge with OSA treatment is poor adherence. Scioscia and collaborators developed a machine learning method to predict adherence, with a sensitivity of 68.6% and an area under the curve of 72.9%. Although the elucidation of factors that impact long-term CPAP adherence is complex, machine learning can identify patients with poor adherence, which can allow for further support or other treatment selection. This method has also been used to help build a CPAP compliance-monitoring system to improve the management of OSA patients.²

But despite the recent strides in treatment, OSA is still severely underdiagnosed and under-recognized in the community, with at least one study suggesting that 80 to 90% of middle-aged adults with moderate to severe sleep apnea remain undiagnosed. Traditionally the diagnosis of this condition has been done by overnight polysomnography, which is performed in a sleep laboratory. However, this requires significant investment in infrastructure and having certified personnel, and hence became an impractical solution to diagnose sleep apnea efficiently. Later, home sleep testing made it easier to screen more people for sleep apnea, but this approach is still associated with significant barriers and diagnosis delays.¹

Lately, there has been a significant surge in AI and neural network-based applications in healthcare diagnostics, including several predictive models that consider demographic and anthropometric data and the presence of comorbidities and symptoms to improve the accuracy of screening.¹

Applying AI neural networks for the clinically based prediction of OSA allows the inclusion of more variables than in linear regression or logistic regression techniques.¹

Here are a few examples. Kirby and colleagues trained a neural network using 23 clinical variables in 255 patients, and subsequently evaluated the predictive properties in a cohort of 150 additional patients. There was a 69% prevalence of OSA in the sample. Their model showed an accuracy of 91.3%, a sensitivity of 98.9%, and a specificity of 80%.¹

Another study by Mencar and team selected a mixture of respiratory signals and clinical variables to include 19 variables, with a communality index of 0.50 or higher out of the 32 initial features, to train classification models and regression models to evaluate the prediction of OSA severity represented either by class or by available apnea–hypopnea indices from polysomnography. The discriminating power between normal subjects and patients with obstructive sleep apnea syndrome has been examined considering different efficacy and efficiency parameters such as classification accuracy, precision, sensitivity, and area under the curve using cross-validation. The support vector machine classification model, trained with the first eight features, gave the best results in terms of classification accuracy, with 44.7%, and area under the curve of 65%. In terms of precision and/or recall, the random forest classification model trained with the first five features showed the best results.¹

On a similar note, Huang and colleagues developed a support vector model incorporating features from routinely collected parameters during clinical evaluation from 6,875 Chinese patients referred for suspected sleep apnea. AHI cutoffs of 5, 15, and 30 were used to stratify OSA severity. The modeling was achieved through fivefold cross-validation. For the prediction of AHI cutoff equal to or greater than 5 per hour two features were selected, while six were selected for the other two groups. The area under the curve in the three groups was 0.82, 0.80, and 0.78, respectively; while sensitivity was 74, 75, and 70 percent, with a specificity of 75, 69, and 70 percent, respectively. This model performed better than other traditional screening tools such as the Berlin questionnaire, and the NoSAS Score.¹ 

A.I. has also been applied to help in the accurate prediction of surgical success and to avoid unnecessary procedures as well as optimizing the surgical treatment outcomes of patients with OSA.¹

For instance, a study conducted by Liu and collaborators demonstrated how machine learning can be used to predict surgical candidates for pediatric OSA to help avoid unnecessary surgery in children. An unsupervised technique, the K-means clustering method, was used to stratify patients into two groups depending on physiological and neurophysiological symptoms. The study demonstrated that children with mild symptoms could avoid adenotonsillectomy and outlined an approach to more accurately predict who should be treated this way.²

Another study by Yang and team used several machine learning models, including logistic regression, tree-based models, support vector machine and neural networks. They aimed to determine the predicted success of uvulopalatal flap or palatal muscle resection surgery. The models were developed with 15 variables including demography, polysomnography data, Friendman stage and Drug-induced sleep endoscopy results, and nasal or palate surgery. Lasso logistic regression was the model with the highest accuracy.²

CÁPSULA
Contrast-enhanced MRI is recommended as the gold-standard imaging for nasopharyngeal cancer delineation. MRI-guided manual delineation of gross tumor volume is challenging and time-consuming due to its complex anatomical structure. Moreover, inter-observer variability in tumor delineation is not negligible. Therefore, automatic delineation based on deep learning models could be a desirable alternative to overcome these difficulties. The deep learning models can detect tumor invasion more accurately and assist junior oncologists in nasopharynx gross tumor volume delineation.⁴

Recently, Wang and collaborators proposed a deep learning model-based framework to automatically draw the contours of the nasopharynx tumor on MRI images. The team’s model showed promising preliminary results with potential applications in the future to assist radiation oncologists in saving time and help in clinical decision-making.⁴
FIN DE CÁPSULA

Welcome back. Moving on to other topics, we will now discuss the clinical utility and efficacy of artificial intelligence in the assessment of laryngeal lesions based on laryngoscopy imaging studies, and its accuracy, sensitivity, and specificity in doing so.
Introducing AI into the diagnostic process in the case of medical imaging is thought to contribute to better precision, replicability, and efficiency in making diagnoses.³

Potentially, 6 to 22 percent of premalignant lesions will develop into malignancies, and the transformation rate depends on the severity of the precancerous lesions. The aim of modern diagnostics is the proper assessment of lesions with the lowest possible invasiveness of examination. Beforehand, however, it is necessary to distinguish malignant and potentially malignant lesions from benign ones.³

There are several tools currently used in the diagnosis of laryngeal lesions at different stages of advancement, including indirect and direct laryngoscopy, ultrasound, computer tomography, and MRI. AI may help accelerate diagnosis and improve its accuracy, while having a beneficial impact on efficiency in clinical practice. Also, the fact that some subclasses of AI allow models to learn how to use gathered information and make decisions independently is very promising.³

Zurek and collaborators performed a systematic review which included 11 papers evaluating the diagnostic accuracy of AI in laryngeal endoscopy. All the participant studies were retrospective and used AI to assess images of laryngeal lesions. The character of the lesions based on vascular patterns, shape, and/or color were assessed with neural networks. Six of the papers evaluated endoscopic images in white light and five used the narrow-band imaging method.³

In the first group of studies, less than 2500 were analyzed with AI, and the accuracy of the AI ranged from 0.806 to 0.997. There was also a tendency of increasing accuracy with the quantity of applied pictures; however, in each of these studies advanced and different pre-processing methods for images were applied, including Gaussian smoothing, the investigation of texture-based global descriptors, the calculation of first-order statistics, specular reflection removal, and region of interest detection.³

For the second group, with a number of images exceeding 2500, an evident trend of increasing accuracy was recognized, from 0.88 to 0.94. This tendency could not yet be confirmed statistically, because only four studies on such a scale had been performed so far. Nonetheless, it is worth noting that for those studies using many analyzed images, the pre-processing methods were quite simple, compared to the first group, and included only choosing images and detecting regions of interest. This comparison identifies two directions for future research, in which there is an awareness of the limitations of data preparation, which should be unified and verified as to not influence the accuracy score.³

Now, in terms of differentiation between benign and malignant lesions, AI also performed very well, with a pooled sensitivity of 0.91 and a pooled specificity of 0.94. This high specificity indicates the ability of AI to discriminate patients with benign lesions from those with malignancies, which can be of high value in clinical practice.³

When comparing the results in narrow-band imaging with those in white light endoscopy, the AI’s accuracy and assessment were comparable regardless of the technology used. The sensitivity and specificity of AI for both methods were 0.89 and 0.95 for white light, and 0.93 and 0.94 for narrow-band imaging, respectively.³

All in all, AI showed high accuracy, sensitivity, and specificity in assessing images of laryngeal lesions. Such accuracy indicates the great utility of this technology in laryngology and provides potential opportunities to introduce it into diagnostic standards. The regression model of accuracy for the seven included studies shows a statistically significant trend between the accuracy of AI diagnoses and the number of images. This means that the key element to improving the quality of AI models in the assessment of laryngeal lesions is the increase in the number of images used while maintaining a high-quality pre-processing method.³

Depending on the study, healthy tissue was differentiated from malignant lesions, such as cancer or severe dysplasia, but also from benign lesions, such as nodules, polyps, Reinke’s edemas, granulomas, or vocal fold palsies. These results indicate the potential of AI in helping young doctors learn the correct diagnosis of laryngeal lesions and, in particular, how to differentiate benign from malignant lesions.³

The meta-analysis’ results indicate that AI is a valuable tool for assessing laryngeal lesions and that the effectiveness of neural networks does not differ from the professional assessment. It would be particularly useful to introduce AI in facilities that do not use narrow-band imaging, because the sensitivity of the network assessments in white light was higher than that of professionals, while the specificity was similar.³

AI is regarded as a very valuable tool in various fields, including healthcare, where it has been deemed suitable for repetitive analytic tasks, complex calculations, and complex forecasts. Such tasks are not uncommon in rhinology, some examples include procedures such as nasal cytology smear analysis, nasal airflow computational fluid dynamics modeling, and radiomics-based oncological risk stratification. However, several intrinsic technical issues make AI applications in rhinology challenging and embryonic at best, and only speculative at the moment due to the complexities of using data in real-world scenarios.⁵

To that end, the team of researchers led by Bulfamante, showed that although several AI rhinological applications have been developed recently, none have been validated in a real-world setting. Rhinological AI applications appear generally restricted to specific tasks, distinctly regulated by the input homogeneity required by AI models and the oversimplifications required to provide answers. Therefore, inputs are often numerically compiled from a prior set of variables. Likewise, graphical information undergoes heavy preprocessing before AI submission. For example, only three reviewed articles by the team used three-dimensional native volume information to allow segmentation of sinonasal structures; eight studies used native bidimensional images, and the rest used some form of data manipulation.⁵

Likewise, theoretically simple analyses such as locating the sinuses in a CT volume remain challenging for AI and only volume estimates have been performed on three-dimensional models. Narrow categorization of answers is required at the output level; therefore, nearly half of the reviewed models used dichotomous outputs, while the remaining used predefined categorical answers or continuous numerical scales.⁵

Finally, we must mention that the application of AI in endoscopic evaluation is currently the subject of intense research, especially in digestive tract endoscopy. The objective is to enhance its performance and resolve limitations related to experience and uncertainty, and therefore implement it in modern instrument systems for the automatic detection of pathologies. The topic of laryngeal endoscopy is still at its initial stage; however, a significant increase in research has been observed in the last two years and the subject will certainly be intensively explored. At this early stage, it is recommended to evaluate the essential strategies of analysis and highlight the importance of consistent data collection, the homogeneity of nomenclature, comparable amounts of images, and other technical aspects related to image processing.³

It is important to consider that the performance of machine learning algorithms depends primarily on appropriate feature selection. One common theme that emerges is that models that include more objective-based features perform better than those that include subjective-based features based on patient self-reporting. Using different techniques to effect dimensionality reduction, such as principal component analysis and singular value decomposition, could help reveal weighted combinations of features that enhance the discriminating capacity of these models. The use of regularization and other techniques can help automate feature selection and bias inclusion of models with fewer parameters. Nonetheless, more research may be needed to help improve feature selection using weighted combinations and automated selection in larger population datasets.¹

Thanks for joining us on this episode of Health Connect. Don't miss out on our next episode. 

Discover more medical news on Viatris Connect.

References:

1. 1. Aiyer I, Shaik L, Sheta A, Surani S. Review of application of machine learning as a screening tool for diagnosis of obstructive sleep apnea. Medicina (Kaunas). 2022;58(11):1574. Available at: http://dx.doi.org/10.3390/medicina58111574
2. 2. Brennan HL, Kirby SD. The role of artificial intelligence in the treatment of obstructive sleep apnea. J Otolaryngol Head Neck Surg. 2023;52(1):7. Available at: The role of artificial intelligence in the treatment of obstructive sleep apnea
3. 3. Żurek M, Jasak K, Niemczyk K, Rzepakowska A. Artificial intelligence in laryngeal endoscopy: Systematic review and meta-analysis. J Clin Med. 2022;11(10):2752. Available at: http://dx.doi.org/10.3390/jcm11102752
4. 4. Wang Y, Chen H, Lin J, Dong S, Zhang W. Automatic detection and recognition of nasopharynx gross tumour volume (GTVnx) by deep learning for nasopharyngeal cancer radiotherapy through magnetic resonance imaging. Radiat Oncol. 2023;18(1):76. Available at: Automatic detection and recognition of nasopharynx gross tumour volume (GTVnx) by deep learning for nasopharyngeal cancer radiotherapy through magnetic resonance imaging
5. 5. Bulfamante AM, Ferella F, Miller AM, Rosso C, Pipolo C, Fuccillo E, et al. Artificial intelligence, machine learning, and deep learning in rhinology: a systematic review. Eur Arch Otorhinolaryngol. 2023;280(2):529–542. Available at: Artificial intelligence, machine learning, and deep learning in rhinology: a systematic review

Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.

NON-2023-9960
NON-2023-2512