Recent Advances in the Application of Artificial Intelligence in Otorhinolaryngology-Head and Neck Surgery

Abstract

This study presents an up-to-date survey of the use of artificial intelligence (AI) in the field of otorhinolaryngology, considering opportunities, research challenges, and research directions. We searched PubMed, the Cochrane Central Register of Controlled Trials, Embase, and the Web of Science. We initially retrieved 458 articles. The exclusion of non-English publications and duplicates yielded a total of 90 remaining studies. These 90 studies were divided into those analyzing medical images, voice, medical devices, and clinical diagnoses and treatments. Most studies (42.2%, 38/90) used AI for image-based analysis, followed by clinical diagnoses and treatments (24 studies). Each of the remaining two subcategories included 14 studies. Machine learning and deep learning have been extensively applied in the field of otorhinolaryngology. However, the performance of AI models varies and research challenges remain.

INTRODUCTION

Artificial intelligence (AI) refers to the ability of machines to mimic human intelligence without explicit programming; AI can solve tasks that require complex decision-making [1,2]. Recent advances in computing power and big data handling have encouraged the use of AI to aid or substitute for conventional approaches. The results of AI applications are promising, and have attracted the attention of researchers and practitioners. In 2015, some AI applications began to outperform human intelligence: ResNet performed better than humans in the ImageNet Large Scale Visual Recognition Competition 2015 [3], and AlphaGo became the first computer Go program to beat a professional Go player in October 2015 [4]. Such technical advances have promising implications for medical applications, particularly because the amount of medical data is doubling every 73 days in 2020 [5]. As such, it is expected that AI will revolutionize healthcare because of its ability to handle data at a massive scale. Currently, AI-based medical platforms support diagnosis, treatment, and prognostic assessments at many healthcare facilities worldwide. The applications of AI include drug development, patient monitoring, and personalized treatment. For example, IBM Watson is a pioneering AI-based medical technology platform used by over 230 organizations worldwide. IBM Watson has consistently outperformed humans in several case studies. In 2016, IBM Watson diagnosed a rare form of leukemia by referring to a dataset of 20 million oncology records [6]. It is clear that the use of AI will fundamentally revolutionize medicine. Frost and Sullivan (a research company) forecast that AI will boost medical outcomes by 30%–40% and reduce treatment costs by up to 50%. The AI healthcare market is expected to attain a value of USD 31.3 billion by 2025 [7].

Otorhinolaryngologists use many instruments to examine patients. Since the early 1990s, AI has been increasingly used to analyze radiological and pathological images, audiometric data, and cochlear implant (CI). Performance [8-10]. As various methods of AI analysis have been developed and refined, the practical scope of AI in the otorhinolaryngological field has been broadened (e.g., virtual reality technology [11-13]). Therefore, it is essential for otorhinolaryngologists to understand the capabilities and limitations of AI. In addition, a data-driven approach to healthcare requires clinicians to ask the right questions and to fit well into interdisciplinary teams [8].

Herein, we review the basics of AI, its current status, and future opportunities for AI in the field of otorhinolaryngology. We seek to answer two questions: “Which areas of otorhinolaryngology have benefited most from AI?” and “ What does the future hold?”

MACHINE LEARNING AND DEEP LEARNING

AI has fascinated medical researchers and practitioners since the advent of machine learning (ML) and deep learning (DL) (two forms of AI) in 1990 and 2010, respectively. A flowchart of the literature search and study selection is presented in Fig. 1. Importantly, AI, ML, and DL overlap (Fig. 2). There is no single definition of AI; its purpose is to automate tasks that generally require the application of human intelligence [14]. Such tasks include object detection and recognition, visual understanding, and decision-making. Generally, AI incorporates both ML and DL, as well as many other techniques that are difficult to map onto recognized learning paradigms. ML is a data-driven technique that blends computer science with statistics, optimization, and probability [15]. An ML algorithm requires (1) input data, (2) examples of correct predictions, and (3) a means of validating algorithm performance. ML uses input data to build a model (i.e., a pattern) that allows humans to draw inferences [16,17]. DL is a subfield of ML, in which tens or hundreds of representative layers are learned with the aid of neural networks. A neural network is a learning structure that features several neurons; when combined with an activation function, a neural network delivers non-linear predictions. Unlike traditional ML algorithms, which typically only extract features, DL processes raw data to define the representations required for classification [18]. DL has been incorporated in many AI applications, including those for medical purposes [19]. The applications of DL thus far include image classification, speech recognition, autonomous driving, and text-to-speech conversion; in these domains, the performance of DL is at least as good as that of humans. Given the significant roles played by ML and DL in the medical field, clinicians must understand both the advantages and limitations of data-driven analytical tools.

Fig. 1.

Flowchart of the literature search and study selection.

Fig. 2.

Interconnections between artificial intelligence, machine learning, and deep learning.

AI IN THE FIELD OF OTORHINOLARYNGOLOGY

AI aids medical image-based analysis

Medical imaging yields a visual representation of an internal bodily region to facilitate analysis and treatment. Ear, nose, and throat-related diseases are imaged in various manners. Table 1 summarizes the 38 studies that used AI to assist medical image-based analysis in clinical otorhinolaryngology. Nine studies (23.7%) addressed hyperspectral imaging, nine studies (23.7%) analyzed computed tomography, six studies (15.8%) applied AI to magnetic resonance imaging, and one study (2.63%) analyzed panoramic radiography. Laryngoscopic and otoscopic imaging were addressed in three studies each (7.89% each). The remaining seven studies (18.39%) used AI to aid in the analysis of neuroimaging biomarker levels, biopsy specimens, simulated Raman scattering data, ultrasonography and mass spectrometry data, and digitized images. Nearly all AI algorithms comprised convolutional neural networks. Fig. 3 presents a schematic diagram of the application of convolutional neural networks in medical image-based analysis; the remaining algorithms consisted of support vector machines and random forests.

Table 1.

AI techniques used for medical image-based analysis

Study	Analysis modality	Objective	AI technique	Validation method	No. of samples in the training dataset	No. of samples in the testing dataset	Best result
							Accuracy (%)/AUC	Sensitivity (%)/specificity (%)
[20]	CT	Anterior ethmoidal artery anatomy	CNN: Inception-V3	Hold-out	675 Images from 388 patients	197 Images	82.7/0.86	-
[21]	CT	Osteomeatal complex occlusion	CNN: Inception-V3	-	1.28 Million images from 239 patients	-	85.0/0.87	-
[22]	CT	Chronic otitis media diagnosis	CNN: Inception-V3	Hold-out	975 Images	172 Images	-/0.92	83.3/91.4
[23]	DECT	HNSCC lymph nodes	RF, GBM	Hold-out	Training and testing set are randomly chosen with a ratio 70:30 from a total of 412 lymph nodes from 50 patients.		90.0/0.96	89.0/91.0
[24]	microCT	Intratemporal facial nerve anatomy	PCA+SSM	-	40 Cadaveric specimens from 21 donors	-	-	-
[25]	CT	Extranodal extension of HNSCC	CNN	Hold out	2,875 Lymph nodes	200 Lymph nodes	83.1/0.84	71.0/85.0
[26]	CT	Prediction of overall survival of head and neck cancer	NN, DT, boosting, Bayesian, bagging, RF, MARS, SVM, k-NN, GLM, PLSR	10-CV	101 Head and neck cancer patients, 440 radiomic features		-/0.67	-
[27]	DECT	Benign parotid tumors classification	RF	Hold-out	882 Images from 42 patients	Two-thirds of the samples	92.0/0.97	86.0/100
[28]	fMRI	Predicting the language outcomes following cochlear implantation	SVM	LOOCV	22 Training samples, including 15 labeled samples and 7 unlabeled samples		81.3/0.97	77.8/85.7
[29]	fMRI	Auditory perception	SVM	10-CV	42 Images from 6 participants		47.0/-	-
[30]	MRI	Relationship between tinnitus and thicknesses of internal auditory canal and nerves	ELM	Repeated hold-out	46 Images from 23 healthy subjects and 23 patients. Test was repeated 10 times for three training ratios, i.e., 50%, 60%, and 70%.		94.0/-	-
[31]	MRI	Prediction of treatment outcomes of sinonasal squamous cell carcinomas	SVM	9-CV	36 Lesions from 36 patients		92.0/-	100/82.0
[32]	Neuroimaging biomarkers	Tinnitus	SVM	5-CV	102 Images from 46 patients and 56 healthy subjects		80.0/0.86	-
[33]	MRI	Differentiate sinonasal squamous cell carcinoma from inverted papilloma	SVM	LOOCV	22 Patients with inverted papilloma and 24 patients with SCC		89.1/-	91.7/86.4
[34]	MRI	Speech improvement for CI candidates	SVM	LOOCV	37 Images from 37 children with hearing loss and 40 images from 40 children with normal hearing		84.0/0.84	80.0/88.0
[35]	Endoscopic images	Laryngeal soft tissue	Weighted voting (UNet+ErfNet)	Hold-out	200 Images	100 Images	84.7/-	-
[36]	Laryngoscope images	laryngeal neoplasms	CNN	Hold-out	14,340 Images from 5,250 patients	5,093 Images from 2,271 patients	96.24/-	92.8/98.9
[37]	Laryngoscope images	Laryngeal cancer	CNN	Hold-out	13,721 Images	1,176 Images	86.7/0.92	73.1/92.2
[38]	Laryngoscope images	Oropharyngeal cariconoma	Naive Bayes	Hold-out	4 Patients with oropharyngeal cariconoma and 1 healthy subject	16 Patients with oropharyngeal cariconoma and 9 healthy subjects	65.9/-	66.8/64.9
[39]	Otoscopic images	Otologic diseases	CNN	Hold-out	734 Images; 80% of the images were used for the training and 20% were used for validation.		84.4/-	-
[40]	Otoscopic images	Otitis media	MJSR	Hold-out	1,230 Images; 80% of and 20% were used the images were used for the training for validation.		91.41/-	89.48/93.33
[41]	Otoscopic images	Otoscopic diagnosis	AutoML	Hold-out	1,277 Images	89 Images	88.7/-	86.1/-
[42]	Digitized images	H&E-stained tissue of oral cavity squamous cell carcinoma	LDA, QDA, RF, SVM	Hold-out	50 Images	65 Images	88.0/0.87	78.0/93.0
[43]	PESI-MS	Intraoperative specimens of HNSCC	LR	LOOCV	114 Non-cancerous specimens and 141 cancerous specimens		95.35/-	-
[44]	Biopsy specimen	Frozen section of oral cavity cancer	SVM	LOOCV	176 Specimen pairs from 27 subjects		-/0.94	100/88.78
[45]	HSI	Head and neck cancer classification	CNN	LOOCV	88 Samples from 50 patients		80.0/-	81.0/78.0
[46]	HSI	Head and neck cancer classification	CNN	LOOCV	12 Tumor-bearing samples for 12 mice		91.36/-	86.05/93.36
[47]	HSI	Oral cancer	SVM, LDA, QDA, RF, RUSBoost	10-CV	10 Images from 10 mice		79.0/0.86	79.0/79.0
[48]	HSI	Head and neck cancer classification	LDA, QDA, ensemble LDA, SVM, RF	Repeated hold-out	20 Specimens from 20 patients	16 Specimens from 16 patients	94.0/0.97	95.0/90.0
[49]	HSI	Tissue surface shape reconstruction	SSRNet	5-CV	200 SL images		96.81/-	92.5/-
[50]	HSI	Tumor margin of HNSCC	CNN	5-CV	395 Surgical specimens		98.0/0.99	-
[51]	HSI	Tumor margin of HNSCC	LDA	10-CV	16 Surgical specimens		90.0/-	89.0/91.0
[52]	HSI	Optical biopsy of head and neck cancer	CNN	LOOCV	21 Surgical gross-tissue specimens		81.0/0.82	81.0/80.0
[53]	SRS	Frozen section of laryngeal squamous cell carcinoma	CNN	5-CV	18,750 Images from 45 patients		100/-	-
[54]	HSI	Cancer margins of ex-vivo human surgical specimens	CNN	Hold-out	11 Surgical specimens	9 Surgical specimens	81.0/0.86	84.0/77.0
[55]	USG	Genetic risk stratification of thyroid nodules	AutoML	Hold-out	556 Images from 21 patients	127 Images	77.4/-	45.0/97.0
[56]	CT	Concha bullosa on coronal sinus classification	CNN: Inception-V3	Hold-out	347 Images (163 concha bullosa images and 184 normal images)	100 Images (50 concha bullosa images and 50 normal images)	81.0/0.93	-
[57]	Panoramic radiography	Maxillary sinusitis diagnosis	AlexNet CNN	Hold-out	400 Healthy images and 400 inflamed maxillary sinuses images	60 Healthy and 60 inflamed maxillary sinuses images	87.5/0.875	86.7/88.3

AI, artificial intelligence; AUC, area under the receiver operating characteristic curve; CT, computed tomography; CNN, convolutional neural network; DECT, dual-energy computed tomography; HNSCC, head and neck squamous cell carcinoma; RF, random forest; GBM, gradient boosting machine; PCA, principle component analysis; SSM, statistical shape model; NN, neural network; DT, decision tree; MARS, multi adaptive regression splines; SVM, support vector machine; k-NN, k-nearest neighbor; GLM, generalized linear model; PLSR, partial least squares and principal component regression; CV, cross-validation; fMRI, functional magnetic resonance imaging; LOOCV, leave-one-out cross-validation; ELM, extreme learning machine; CI, cochlear implant; MJSR, multitask joint sparse representation; LDA, linear discriminant analysis; QDA, quadratic discriminant analysis; PESI-MS, probe electrospray ionization mass spectrometry; LR, logistic regression; HSI, hyperspectral imaging; SSRNet, super-spectral-resolution network; SRS, stimulated Raman scattering; USG, ultrasonography.

Fig. 3.

Artificial intelligence (AI) techniques used for medical image-based analysis.

AI aids voice-based analysis

The subfield of voice-based analysis within otorhinolaryngology seeks to improve speech, to detect voice disorders, and to reduce the noise experienced by patients with (CIs; Table 2 lists the 14 studies that used AI for speech-based analyses. Nine (64.29%) sought to improve speech intelligibility or reduce noise for patients with CIs. Two (14.29%) used acoustic signals to detect voice disorders [67] and “hot potato voice” [70]. In other studies, AI was used for symptoms, voice pathologies, or electromyographic signals as a way to detect voice disorders [68,69], or to restore the voice of a patient who had undergone total laryngectomy [71]. Neural networks were favored, followed by k-nearest neighbor methods, support vector machines, and other widely known classifiers (e.g., decision trees and XGBoost). Fig. 4 presents a schematic diagram of the application of convolutional neural networks in medical voice-based analysis.

Table 2.

AI techniques used for voice-based analysis

Study	Analysis modality	Objective	AI technique	Validation method	No. of samples in the training dataset	No. of samples in the testing dataset	Best result
[58]	CI	Noise reduction	NC+DDAE	Hold-out	120 Utterances	200 Utterances	Accuracy: 99.5%
[59]	CI	Segregated speech from background noise	DNN	Hold-out	560×50 Mixtures for each noise type and SNR	160 Noise segments from original unperturbed noise	Hit ratio: 84%; false alarm: 7%
[60]	CI	Improved pitch perception	ANN	Hold-out	1,500 Pitch pairs	10% of the training material	Accuracy: 95%
[61]	CI	Predicted speech recognition and QoL outcomes	k-NN, DT	10-CV	A total of 29 patients, including 48% unilateral CI users and 51% bimodal CI users		Accuracy: 81%
[62]	CI	Noise reduction	DDAE	Hold-out	12,600 Utterances	900 Noisy utterances	Accuracy: 36.2%
[63]	CI	Improved speech intelligibility in unknown noisy environments	DNN	Hold-out	640,000 Mixtures of sentences and noises	-	Accuracy: 90.4%
[64]	CI	Modeling electrode-to-nerve interface	ANN	Hold-out	360 Sets of fiber activation patterns per electrode	40 Sets of fiber activation patterns per electrode	-
[65]	CI	Provided digital signal processing plug-in for CI	WNN	Hold-out	120 Consonants and vowels, sampled at 16 kHz; half of data was used as training set and the rest was used as testing set.		SNR: 2.496; MSE: 0.086; LLR: 2.323
[66]	CI	Assessed disyllabic speech test performance in CI	k-NN	-	60 Patients	-	Accuracy: 90.83%
[67]	Acoustic signals	Voice disorders detection	CNN	10-CV	451 Images from 10 health adults and 70 adults with voice disorders		Accuracy: 90%
[68]	Dysphonic symptoms	Voice disorders detection	ANN	Repeated hold-out	100 Cases of neoplasm, 508 cases of benign phonotraumatic, 153 cases of vocal palsy		Accuracy: 83%
[69]	Pathological voice	Voice disorders detection	DNN, SVM, GMM	5-CV	60 Normal voice samples and 402 pathological voice samples		Accuracy: 94.26%
[70]	Acoustic signal	Hot potato voice detection	SVM	Hold-out	2,200 Synthetic voice samples	12 HPV samples from real patients	Accuracy: 88.3%
[71]	SEMG signals	Voice restoration for laryngectomy patients	XGBoost	Hold-out	75 Utterances using 7 SEMG sensors	-	Accuracy: 86.4%

AI, artificial intelligence; CI, cochlear implant; NC, noise classifier; DDAE, deep denoising autoencoder; DNN, deep neural network; SNR, signal-to-noise ratio; ANN, artificial neural network; QoL, quality of life; k-NN, k-nearest neighbors; DT, decision tree; CV, cross-validation; WNN, wavelet neural network; MSE, mean square error; LLR, log-likelihood ratio; CNN, convolutional neural network; GMM, Gaussian mixture model; SVM, support vector machine; HPV, human papillomavirus; SEMG, surface electromyographic.

Fig. 4.

Artificial intelligence (AI) techniques used for voice-based analysis.

AI analysis of biosignals detected from medical devices

Medical device-based analyses seek to predict the responses to clinical treatments in order to guide physicians who may wish to choose alternative or more aggressive therapies. AI has been used to assist polysomnography, to explore gene expression profiles, to interpret cellular cartographs, and to evaluate the outputs of non-contact devices. These studies are summarized in Table 3. Of these 14 studies, most (50%, seven studies) focused on analyses of gene expression data. Three studies (21.43%) used AI to examine polysomnography data in an effort to score sleep stages [72,73] or to identify long-term cardiovascular disease [74]. Most algorithms employed ensemble learning (random forests, Gentle Boost, XGBoost, and a general linear model+support vector machine ensemble); this approach was followed by neural networkbased algorithms (convolutional neural networks, autoencoders, and shallow artificial neural networks). Fig. 5 presents a schematic diagram of the application of the autoencoder and the support vector machine in the analysis of gene expression data.

Table 3.

AI analysis of biosignals detected from medical device

Study	Analysis modality	Objective	AI technique	Validation method	No. of samples in the training dataset	No. of samples in the testing dataset	Best result
[73]	EEG signal of PSG	Sleep stage scoring	CNN	5-CV	294 Sleep studies; 122 composed the training set, 20 composed the validation set, and 152 were used in the testing set.		Accuracy: 81.81%; F1 score: 81.50%; Cohen’s Kappa: 72.76%
[72]	EEG, EMG, EOG signals of PSG	Sleep stage scoring	CNN	Hold-out	42,560 Hours of PSG data from 5,213 patients	580 PSGs	Accuracy: 86%; F1 score: 81.0%; Cohen’s Kappa: 82.0%
[74]	Sleep heart rate variability in PSG	Long-term cardiovascular outcome prediction	XGBoost	5-CV	1,252 Patients with cardio vascular disease and 859 patients with non-cardio vascular disease		Accuracy: 75.3%
[87]	Sleep breathing sound using an air-conduction microphone	AHI prediction	Gaussian process, SVM, RF, LiR	10-CV	116 Patients with OSA		CC: 0.83; LMAE: 9.54 events/hr; RMSE: 13.72 events/hr
[88]	Gene signature	Thyroid cancer lymph node metastasis and recurrence rediction	LDA	6-CV	363 Samples	72 Samples	AUC: 0.86; sensitivity: 86%; specificity: 62%; PPV: 93%; NPV: 42%
[89]	Gene expression profile	Response prediction to chemotherapy in patient with HNSCC	SVM	LOOCV	16 TPF-sensitive patients and 13 non-TPF-sensitive patients		Sensitivity: 88.3%; specificity: 88.9%
[90]	Mucus cytokines	SNOT-22 scores prediction of CRS patients	RF, LiR	-	147 Patients with 65 patients with postoperative follow-up		R2: 0.398
[91]	Cellular cartography	Single-cell resolution mapping of the organ of Corti	Gentle boost, RF, CNN	Hold-out	20,416 Samples	19,594 Samples	Recall: 99.3%; precision: 99.3%; F1: 93.3%
[92]	RNA sequencing, miRNA sequencing, methylation data	HNSCC progress prediction	Autoencoder and SVM	2×5-CV	360 Samples from TCGA		C-index: 0.73; Brier score: 0.22
[93]	DNA repair defect	HNSCC progress prediction	CART	10×5-CV	180 HPV-negative HNSCC patients		AUC: 1.0
[94]	PESI-MS	Identified TGF-β signaling in HNSCC	LDA	LOOCV	A total of 240 and 90 mass spectra from TGF-β-unstimulated and stimulated HNSCC cells, respectively		Accuracy: 98.79%
[95]	Next generation sequencing of RNA	Classified the risk of malignancy in cytologically indeterminate thyroid nodules	Ensemble of elastic net GLM and SVM	40×5-CV	A total of 10,196 genes, among which are 1,115 core genes		Sensitivity: 91%; specificity: 68%
[96]	Gene expression profile	HPV-positive oropharyngeal squamous cell carcinoma detection	LR	500-CV	146 Genes from patients with node-negative disease and node-positive disease		AUC: 0.93
[97]	miRNA expression profile	Sensorineural hearing loss prediction	DF, DJ, LR, NN	LOOCV	16 Patients were included.		Accuracy: 100%

AI, artificial intelligence; EEG, electroencephalogram; PSG, polysomnography; CNN, convolutional neural network; CV, cross-validation; EMG, electromyography; EOG, electrooculogram; AHI, apnea-hypopnea index; SVM, support vector machine; RF, random forest; LiR, linear regression; OSA, obstructive sleep apnea; CC, correlation coefficient; LMAE, least mean absolute error; RMSE, root mean squared error; LDA, linear discriminant analysis; AUC, area under the receiver operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value; HNSCC, head and neck squamous cell carcinoma; LOOCV, leave-one-out cross validation; TPF, docetaxel, cisplatin, and 5-fluorouracil; SNOT-22, 22-item sinonasal outcome test; CRS, chronic rhinosinusitis; miRNA, microRNA; TCGA, the cancer genome atlas; CART, classification and regression trees; HPV, human papillomavirus; PESI-MS, probe electrospray ionization mass spectrometry; TGF-β, transforming growth factor beta; GLM, generalized linear model; LR, logistic regression; DF, decision forest; DJ, decision jungle; NN, neural network.

Fig. 5.

Artificial intelligence (AI) analyses of biosignals detected from medical devices. SVM, support vector machine.

AI for clinical diagnoses and treatments

Clinical diagnoses and treatments consider only symptoms, medical records, and other clinical documentation. We retrieved 24 relevant studies (Table 4). Of the ML algorithms, most used logistic regression for classification, followed by random forests and support vector machines. Notably, many studies used hold-outs to validate new methods. Fig. 6 presents a schematic diagram of the process cycle of utilizing AI for clinical diagnoses and treatments.

Table 4.

AI techniques used for clinical diagnoses and treatments

Study	Analysis modality	Objective	AI technique	Validation method	No. of samples in the training dataset	No. of samples in the testing dataset	Best result
[98]	Hearing aids	Hearing gain prediction	CRDN	Hold-out	2,182 Patients that were diagnosed with hearing loss; the percentages of randomly sampled training, validation, and test sets were 40%, 30%, and 30%, respectively.		MAPE: 9.2%
[99]	Hearing aids	Predicted CI outcomes	RF	LOOCV	121 Postlingually deaf adults with CI		MAE: 6.1; Pearson’s correlation coefficient: 0.96
[100]	Clinical data	SSHL prediction	DBN, LR, SVM, MLP	4-CV	1,220 Unilateral SSHL patients		Accuracy: 77.58%; AUC: 0.84
[101]	Clinical data including demographics and risk factors	Determined the risk of head and neck cancer	LR	Hold-out	1,005 Patients, containing 932 patients with no cancer outcome and 73 patients with cancer outcome	235 Patients, containing 212 patients with no cancer outcome and 23 patients with cancer outcome	AUC: 0.79
[102]	Clinical data including symptom	Peritonsillar abscess diagnosis prediction	NN	Hold-out	641 Patients	275 Patients	Accuracy: 72.3%; sensitivity: 6.0%; specificity: 50%
[103]	Vestibular test batteries	Vestibular function assessment	DT, RF, LR, AdaBoost, SVM	Hold-out	5,774 Individuals	100 Individuals	Accuracy: 93.4%
[104]	Speakers and microphones within existing smartphones	Middle ear fluid detection	LR	LOOCV	98 Patient ears		AUC: 0.9; sensitivity: 84.6%; specificity: 81.9%
[105]	Cancer data survival	5-Year survival patients with oral cavity squamous cell carcinoma	DF, DJ, LR, NN	Hold-out	26,452 Patients	6,613 Patients	AUC: 0.8; accuracy: 71%; precision: 71%; recall: 68%
[106]	Histological data	Occult lymph node metastases identification in clinically oral cavity squamous cell	RF, SVM, LR, C5.0	Hold-out	56 Patients	112 Patients	AUC: 0.89; accuracy: 88.0%; NPV: >95%
[107]	Clinicopathologic data	Head and neck free tissue transfer surgical complications prediction	GBDT	Hold-out	291 Patients	73 Patients	Specificity: 62.0%; sensitivity: 60.0%; F1: 60.0%
[108]	Clinicopathologic data	Delayed adjuvant radiation prediction	RF	Hold-out	61,258 Patients	15,315 Patients	Accuracy: 64.4%; precision: 58.5%
[109]	Clinicopathologic data	Occult nodal metastasis prediction in oral cavity squamous cell carcinoma	LR, RF, SVM, GBM	Hold-out	1,570 Patients	391 Patients	AUC: 0.71; sensitivity: 75.3%; specificity: 49.2%
[110]	Dataset of the center of pressure sway during foam posturography	Peripheral vestibular dysfunction prediction	GBDT, bagging, LR	CV	75 Patients with vestibular dysfunction and 163 healthy controls		AUC: 0.9; recall: 0.84
[111]	TEOAE signals	Meniere’s disease hearing outcome prediction	SVM	5-CV	30 Unilateral patients		Accuracy: 82.7%
[112]	Semantic and syntactic patterns in clinical documentation	Vestibular diagnoses	NLP+Naïve Bayes	10-CV	866 Physician-generated histories from vestibular patients		Sensitivity: 93.4%; specificity: 98.2%; AUC: 1.0
[113]	Endoscopic imaging	Nasal polyps diagnosis	ResNet50, Xception, and Inception V3	Hold-out	23,048 Patches (167 patients) as training set, 1,577 patches (12 patients) as internal validation set, and 1,964 patches (16 patients) as external test set		Inception V3: AUC: 0.974
[114]	Intradermal skin tests	Allergic rhinitis diagnosis	Associative classifier	10-CV	872 Patients with allergic symptoms		Accuracy: 88.31%
[115]	Clinical data	Identified phenotype and mucosal eosinophilia endotype subgroups of patients with medical refractory CRS	Cluster analysis	-	46 Patients with CRS without nasal polyps and 67 patients with nasal polyps		-
[116]	Clinical data	Prognostic information of patient with CRS	Discriminant analysis	-	690 Patients		-
[117]	Clinical data	Identified phenotypic subgroups of CRS patients	Discriminant analysis	-	382 Patients		-
[118]	Clinical data	Characterization of distinguishing clinical features between subgroups of patients with CRS	Cluster analysis	-	97 Surgical patients with CRS		-
[119]	Clinical data	Identified features of CRS without nasal polyposis	Cluster analysis	-	145 Patients of CRS without nasal polyposis		-
[120]	Clinical data	Identified inflammatory endotypes of CRS	Cluster analysis	-	682 Cases (65% with CRS without nasal polyps)		-
[121]	Clinical data	Identified features of CRS with nasal polyps	Cluster analysis	-	375 Patients		-

AI, artificial intelligence; CRDN, cascade recurring deep network; MAPE, mean absolute percentage error; RF, random forest; LOOCV, leave-one-out cross validation; CI, cochlear implant; MAE, mean absolute error; SSHL, sudden sensorineural hearing loss; DBN, deep belief network; LR, logistic regression; SVM, support vector machine; MLP, multilayer perceptron; CV, cross-validation; AUC, area under the receiver operating characteristic curve; NN, neural network; DT, decision tree; DF, decision forest; DJ, decision jungle; NPV, negative predictive value; GBDT, gradient boosted decision trees; GBM, gradient boosting machine; TEOAE, transient-evoked otoacoustic emission; NLP, natural language processing; CRS, chronic rhinosinusitis.

Fig. 6.

Artificial intelligence (AI) techniques used for clinical diagnoses and treatments. EMR, electronic medical record.

DISCUSSION

We systematically analyzed reports describing the integration of AI in the field of otorhinolaryngology, with an emphasis on how AI may best be implemented in various subfields. Various AI techniques and validation methods have found favor. As described above, advances in 2015 underscored that AI would play a major role in future medicine. Here, we reviewed post-2015 AI applications in the field of otorhinolaryngology. Before 2015, most AI-based technologies focused on CIs [10,75-86]. However, AI applications have expanded greatly in recent years. In terms of image-based analysis, images yielded by rigid endoscopes, laryngoscopes, stroboscopes, computed tomography, magnetic resonance imaging, and multispectral narrow-band imaging [38], as well as hyperspectral imaging [45-52,54], are now interpreted by AI. In voice-based analysis, AI is used to evaluate pathological voice conditions associated with vocal fold disorders, to analyze and decode phonation itself [67], to improve speech perception in noisy conditions, and to improve the hearing of patients with CIs. In medical device-based analyses, AI is used to evaluate tissue and blood test results, as well as the outcomes of otorhinolaryngology-specific tests (e.g., polysomnography) [72,73,122] and audiometry [123,124]. AI has also been used to support clinical diagnoses and treatments, decision-making, the prediction of prognoses [98-100,125,126], disease profiling, the construction of mass spectral databases [43,127-129], the identification or prediction of disease progress [101,105,107-110,130], and the confirmation of diagnoses and the utility of treatments [102-104,112,131].

Although many algorithms have been applied, some are not consistently reliable, and certain challenges remain. AI will presumably become embedded in all tools used for diagnosis, treatment selection, and outcome predictions; thus, AI will be used to analyze images, voices, and clinical records. These are the goals of most studies, but again, the results have been variable and are thus difficult to compare. The limitations include: (1) small training datasets and differences in the sizes of the training and test datasets; (2) differences in validation techniques (notably, some studies have not included data validation); and (3) the use of different performance measures during either classification (e.g., accuracy, sensitivity, specificity, F1, or area under the receiver operating characteristic curve) or regression (e.g., root mean square error, least mean absolute error, R-squared, or log-likelihood ratio).

ML algorithms always require large, labeled training datasets. The lack of such data was often a major limitation of the studies that we reviewed. AI-based predictions in the field of otorhinolaryngology must be rigorously validated. Often, as in the broader medical field, an element of uncertainty compromises an otherwise ideal predictive method, and other research disparities were also apparent in the studies that we reviewed. Recent promising advances in AI include the ensemble learning model, which is more intuitive and interpretable than other models; this model facilitates bias-free AI-based decision-making. The algorithm incorporates a concept of “fairness,” considers ethical and legal issues, and respects privacy during data mining tasks. In summary, although otorhinolaryngology-related AI applications were divided into four categories in the present study, the practical use of a particular AI method depends on the circumstances. AI will be helpful for use in real-world clinical treatment involving complex datasets with heterogeneous variables.

CONCLUSION

We have described several techniques and applications for AI; notably, AI can overcome existing technical limitations in otorhinolaryngology and aid in clinical decision-making. Otorhinolaryngologists have interpreted instrument-derived data for decades, and many algorithms have been developed and applied. However, the use of AI will refine these algorithms, and big health data and information from complex heterogeneous datasets will become available to clinicians, thereby opening new diagnostic, treatment, and research frontiers.

HIGHLIGHTS

▪ Ninety studies that implemented artificial intelligence (AI) in otorhinolaryngology were reviewed and classified.

▪ The studies were divided into four subcategories.

▪ Research challenges regarding future applications of AI in otorhinolaryngology are discussed.