You are on page 1of 9

Research

JAMA Otolaryngology–Head & Neck Surgery | Original Investigation

A Deep Learning Approach to Predict Conductive Hearing Loss


in Patients With Otitis Media With Effusion Using Otoscopic Images
Junbo Zeng, MD; Weibiao Kang, MD; Suijun Chen, MD; Yi Lin, MEng; Wenting Deng, MD; Yajing Wang, MD;
Guisheng Chen, MD; Kai Ma, PhD; Fei Zhao, MD, PhD; Yefeng Zheng, PhD; Maojin Liang, MD; Linqi Zeng, MD;
Weijie Ye, MD; Peng Li, PhD; Yubin Chen, MD; Guoping Chen, MD; Jinliang Gao, MD; Minjian Wu, MD;
Yuejia Su, MD; Yiqing Zheng, MD; Yuexin Cai, MD, PhD

Supplemental content
IMPORTANCE Otitis media with effusion (OME) is one of the most common causes of
acquired conductive hearing loss (CHL). Persistent hearing loss is associated with poor
childhood speech and language development and other adverse consequence. However, to
obtain accurate and reliable hearing thresholds largely requires a high degree of cooperation
from the patients.
OBJECTIVE To predict CHL from otoscopic images using deep learning (DL) techniques and a
logistic regression model based on tympanic membrane features.

DESIGN, SETTING, AND PARTICIPANTS A retrospective diagnostic/prognostic study was


conducted using 2790 otoscopic images obtained from multiple centers between January
2015 and November 2020. Participants were aged between 4 and 89 years. Of 1239
participants, there were 209 ears from children and adolescents (aged 4-18 years [16.87%]),
804 ears from adults (aged 18-60 years [64.89%]), and 226 ears from older people (aged
>60 years, [18.24%]). Overall, 679 ears (54.8%) were from men. The 2790 otoscopic images
were randomly assigned into a training set (2232 [80%]), and validation set (558 [20%]). The
DL model was developed to predict an average air-bone gap greater than 10 dB. A logistic
regression model was also developed based on otoscopic features.

MAIN OUTCOMES AND MEASURES The performance of the DL model in predicting CHL was
measured using the area under the receiver operating curve (AUC), accuracy, and F1 score (a
measure of the quality of a classifier, which is the harmonic mean of precision and recall; a
higher F1 score means better performance). In addition, these evaluation parameters were
compared to results obtained from the logistic regression model and predictions made by
three otologists.

RESULTS The performance of the DL model in predicting CHL showed the AUC of 0.74,
accuracy of 81%, and F1 score of 0.89. This was better than the results from the logistic
regression model (ie, AUC of 0.60, accuracy of 76%, and F1 score of 0.82), and much
improved on the performance of the 3 otologists; accuracy of 16%, 30%, 39%, and F1 scores
of 0.09, 0.18, and 0.25, respectively. Furthermore, the DL model took 2.5 seconds to predict
from 205 otoscopic images, whereas the 3 otologists spent 633 seconds, 645 seconds, and
692 seconds, respectively.

CONCLUSIONS AND RELEVANCE The model in this diagnostic/prognostic study provided


greater accuracy in prediction of CHL in ears with OME than those obtained from the logistic
regression model and otologists. This indicates great potential for the use of artificial
intelligence tools to facilitate CHL evaluation when CHL is unable to be measured.

Author Affiliations: Author


affiliations are listed at the end of this
article.
Corresponding Author: Yuexin Cai,
MD, PhD (caiyx25@mail.sysu.edu.
cn), and Yiqing Zheng, MD (zhengyiq
@mail.sysu.edu.cn), Department of
Otolaryngology, Sun Yat-sen
Memorial Hospital, Sun Yat-sen
JAMA Otolaryngol Head Neck Surg. 2022;148(7):612-620. doi:10.1001/jamaoto.2022.0900 University, 107# Yanjiang West Rd,
Published online May 19, 2022. Guangzhou 510120, China.

612 (Reprinted) jamaotolaryngology.com

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024
A Deep Learning Approach to Predict Conductive Hearing Loss in Patients With Otitis Media With Effusion Original Investigation Research

O
titis media with effusion (OME) is one of the most com-
mon middle ear conditions presenting at otolaryngol- Key Points
ogy and audiology clinics. Evidence suggests that ap-
Question How effective are deep learning (DL) models in
proximately 90% of preschool children will experience OME predicting conductive hearing loss (CHL) from otoscopic images of
at least once, and 30% to 40% are likely to have recurrent ears with otitis media with effusion?
episodes.1,2 In the natural course of OME, more than 20% of
Findings In this diagnostic/prognostic study including 2790
children will have a conductive hearing loss (CHL) of more
otoscopic images from 1239 patients, a DL model was developed
than 35 dB and 5% to 10% more than 50 dB.3-5 The CHL in pa- using otoscopic images and validated in multiple centers. The DL
tients with OME is largely caused by fluid accumulation in the model predicted CHL with an accuracy of 81%.
middle ear cavity, which increases stiffness and mass of the
Meaning This research to predict CHL using a DL model revealed
tympanum.6 Without timely and appropriate intervention, per-
great potential for the quick and accurate prediction of degree of
sistent CHL in preschool children with OME is associated with conductive hearing loss, and offers opportunities to reduce
poor speech and language development and other adverse con- patient waiting time and avoid unnecessary referral.
sequences, for example behavior issues and educational
difficulties.6-8
According to the clinical practice guidelines recom- cal tests (PTA and tympanometry). Inclusion criteria re-
mended by the American Academy of Otolaryngology in 2016,6 quired that otoscopic images and audiological assessment re-
otoscopy, pure tone audiometry (PTA), and tympanometry are sults were measured at the same time and on individual OME
suggested as essential tools for OME diagnosis. Although PTA ears. Ears with OME with a history of middle ear surgery (eg,
is considered the gold standard for assessing hearing sensitiv- grommet insertion) were excluded. The high incidence of na-
ity, it has several limitations owing to the subjective nature of its sopharyngeal tumors in the south of China is associated with
measurement, requiring a high degree of cooperation from the 8% to 29% of patients there having OME after radiotherapy.23,24
patients to achieve accurate and reliable hearing thresholds. However, owing to the low incidence in the other regions of
Indeed, Browning et al9 and Kemaloglu et al10 have indicated China, to avoid confusion, radiation-induced OME ears were ex-
that younger children are unlikely to be able to concentrate long cluded in the present study. In addition, otoscopic images of OME
enough to complete the test. ears labeled with an incorrect ear side were excluded. This study
The otoscope is a fundamental tool used in otologic prac- was approved by the institutional review board of Sun Yat-sen
tice to examine tympanic membrane condition by observing Memorial Hospital (reference number SYSEC-KY-KS-2021-191),
its shape, color, and integrity.11,12 However, otoscopic exami- and written informed consent was waived owing to the retro-
nation and otoscopic images are currently unable to provide spective nature of the study.
information on hearing levels. A number of studies have shown
that deep learning (DL) is an effective tool to classify patho- Quality Control for Otoscopic Images
logic conditions from medical images, information that can All otoscopic images were taken by otologists in routine
then support diagnosis and treatment.13-15 In the field of otol- otologic practice. They were captured in standard high-
ogy, previous DL studies have mainly used otoscopic images resolution using 2 different otoscopes, a 4-mm STORZ 0° en-
as input data and focused on how to achieve a high diagnos- doscope (KARL STORZ, Germany) or a 2.7-mm TIAN SONG 0°
tic accuracy in classifying middle ear disorders.16-22 endoscope (TIAN SONG, China). Images were stored in JPEG
The aim of this study was to predict the degree of CHL from or PNG format in the range 500 × 500 to 700 × 700 pixels. Only
otoscopic images using DL techniques and a logistic regres- otoscopic images where the pars flaccida and pars tensa were
sion model based on tympanic membrane features. The im- captured intact were included in this study. Furthermore, only
portance of this novel research is that it could enable a quick otoscopic images taken using standard white light were deemed
and accurate prediction of the degree of CHL, and thus pro- eligible for developing the DL model. Overexposed and un-
vide solutions to reduce patient waiting time and avoid un- derexposed images, and images where the tympanic mem-
necessary referral. brane was largely blocked by earwax were excluded. The 1 to
3 best quality otoscopic images captured from different angles
of each ear were selected for analysis. Overall, PTA hearing
thresholds at frequencies of 500, 1000, 2000, and 4000 Hz were
Methods measured in left and right ears. The air-bone gap (ABG) was cal-
Study Design and Data Set culated at each frequency and an average of the 4 frequencies
In this multicenter study, otoscopic images of OME cases were calculated. We chose averaged ABG across the 4 frequencies
obtained from 3 tertiary care medical centers in south China, greater than 10 dB as the criteria of conductive hearing loss.25
including Sun Yat-sen Memorial Hospital of Sun Yat-sen The severity of OME and changes to the tympanic membrane is
University, The Third Affiliated Hospital of Sun Yat-sen Uni- associated with different degrees of conductive hearing loss at
versity, and Zhongshan People’s Hospital. All data were col- different individual frequencies.26-28 Consequently, the DL
lected between January 2015 and November 2020 and re- model was used to predict CHL using different criteria: ABG
viewed retrospectively. Cases of OME were diagnosed greater than 10 dB at each frequency, averaged air conduction
according to clinical otologic practice that included medical hearing threshold more than 500, 1000, 2000, 4000 Hz greater
history, physical examination with otoscopes, and audiologi- than 25 dB, and averaged ABG of the 4 frequencies greater than

jamaotolaryngology.com (Reprinted) JAMA Otolaryngology–Head & Neck Surgery July 2022 Volume 148, Number 7 613

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024
Research Original Investigation A Deep Learning Approach to Predict Conductive Hearing Loss in Patients With Otitis Media With Effusion

20 dB. Changes in the tympanic membrane used to develop ternal validation. Within the data set, 738 ears with 1678 oto-
the logistic regression model included attic retraction pocket, scopic images were from Sun Yat-sen Memorial Hospital of Sun
atelectasis, type of fluid (full, air-fluid level, bubble), color of Yat-sen University; and 330 ears with 721 otoscopic images were
tympanic membrane, color of fluid, myringosclerosis, and from The Third Affiliated Hospital of Sun Yat-sen University;
opacification.16,25,27 An otologist (J.B.Z.) was assigned to col- and 171 ears with 391 otoscopic images were from Zhongshan
lect otoscopic images, other clinical data, and record changes people’s hospital. The OME ears with otoscopic images were
in the tympanic membrane. Another otologist (Y.X.C.) with randomly split into the training set (1034 ears with 2232
more than 10 years of clinical experience reviewed the data in- otoscopic images) and validation set (205 ears with 558
dependently. Any discrepancies were discussed with an- otoscopic images).
other experienced otologist (S.J.C.) until consensus was
reached. The detailed criteria for the otologist to predict CHL Summary of Clinical Characteristics
on the basis of otoscopic images are described in the eMethods Clinical diagnosis of CHL was taken on the basis of the aver-
in the Supplement. aged ABG across 4 frequencies being greater than 10 dB. Over-
all, 996 ears (80.4%) demonstrated CHL. The data set com-
Convolutional Neural Network Model prised 1239 participants aged between 4 and 89 years. This
Development and Validation included 209 ears from children and adolescents (aged 4-18
Convolutional neural network (CNN) model development and years 16.9%), 804 ears from adults (aged 18-60 years, 64.9%),
validation included data preprocessing, strategies to mini- and 226 ears from older people (aged >60 years, 18.2%); 679
mize overfittings, model selection, construction of a predic- ears (54.8%) were from men. Ears with tympanogram A (5.0%)
tion system, and main hyperparameters and are detailed in were further confirmed with tympanotomy, whereas 835 ears
the eResults in the Supplement. To evaluate the performance (67.40%) with tympanogram B, and 342 ears (27.6%) with tym-
of the DL model, the discriminative performance was panogram C did not need further testing to confirm.
assessed from the area under the receiver operating character- Atelectasis (753 [60.7%]) and attic retraction pocket (922
istic (AUC) curve, accuracy, F1 score (a measure of the quality [74.3%]) of any stage were most commonly seen. Pearl gray (736
of a classifier, which is the harmonic mean of precision and [59.5%]) was the most common tympanic membrane color, and
recall, higher F1 score means better performance) and com- amber (1088 [87.8%]) the most common fluid color. Full (898
parison with otologists’ prediction and the outcomes [72.5%]) was the most common type of fluid status. Statisti-
obtained using the logistic regression model, all detailed in cal analysis revealed that demographic factors, type of tym-
eFigure 1 in the Supplement. panogram, sex, changes of tympanic membrane, and age were
not significantly different between the training and valida-
Deep Learning Model Interpretability tion sets. Detailed characteristics of the ears in the training and
To interpret how the DL model produced the diagnostic out- validation sets are summarized and compared in Table 1.
puts it was crucial to visualize what the discriminative
regions in the otoscopic images used were by the DL model. Evaluation of DL Model Performance
A gradient-weighted heatmap was obtained from the final To commence, receiver operating characteristic (ROC) curves
layer of the neural network, which highlighted the discrimi- were plotted and AUC of the DL model calculated to show di-
native regions in the original otoscopic image using blue to agnostic performance. The DL diagnostic model achieved an
red colors. AUC of 0.74 (95% CI, 0.67-0.80) with ABG over 4 frequencies
greater than 10 dB. The F1 score was 0.97 (95% CI, 0.95-
Experimental Equipment and Statistical Analysis 0.99), and accuracy 81% (95% CI, 77%-85%). With ABG at 0.5
The experiments were performed using Python (version 3.6 in kHz, the AUC was 0.74 (95% CI, 0.65-0.83), 1 kHz 0.79 (95%
Keras, version 2.2.4) to develop the diagnostic model based CI, 0.68-0.85), 2 kHz 0.68 (95% CI, 0.63-0.75), and at 4 kHz
on TensorFlow with 4 Titan XP 256 GB GPU (version 1.12.0). 0.65 (95% CI, 0.55-0.78), respectively. When the DL model was
All statistical analyses were carried out using SPSS statistical applied to diagnose those with a CHL using averaged ABG
software (version 26, IBM). The categorical variables were com- across the 4 frequencies greater than 20 dB, the AUC was 0.69
pared using the χ2 test and Fisher exact test. The continuous (95% CI, 0.60-0.78), with F1 score of 0.83 (95% CI, 0.77-
variables were compared using the Kruskal-Wallis rank-sum 0.88), and accuracy of 73% (95% CI, 65%-80%) (Table 2). When
test. The AUC, accuracy, and F1 score were calculated, and 95% human and artificial intelligence were compared, the perfor-
CIs of AUC were estimated using bootstrapping with 1000 rep- mance of the 3 otologists in predicting CHL by reviewing the
licates. Two-sided P < .05 indicated statistical significance. otoscopic images was lower compared with the DL model in
terms of accuracy (81% vs 30%, 16%, and 39%). The DL model
took just 2.5 seconds to complete its prediction of CHL, whereas
the 3 otologists took 633 seconds, 645 seconds, and 692 sec-
Results onds, respectively. The logistic regression model based on tym-
Summary of Otoscopic Images panic membrane changes provided a lower AUC of 0.62 (95%
After application of inclusion and exclusion criteria, a total of CI, 0.45-0.75), F1 score of 0.82 (95% CI, 0.79-0.85), and accu-
1239 ears with clinical data and 2790 otoscopic images were racy of 76% (95% CI, 70%-81%) (eFigure 2 in the Supplement;
taken from the patients for use in model development and in- Table 3).

614 JAMA Otolaryngology–Head & Neck Surgery July 2022 Volume 148, Number 7 (Reprinted) jamaotolaryngology.com

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024
A Deep Learning Approach to Predict Conductive Hearing Loss in Patients With Otitis Media With Effusion Original Investigation Research

Table 1. Basic Characteristics of Participants

No. (%)
Characteristic Training set (n = 1024) Validation set (n = 215) P value
Age, median (IQR), y 44 (27-58) 44 (25-60) .93
Male sex 562 (54.9) 117 (54.4) .90
Tympanogram type
Tympanogram A 52 (5.0) 10 (4.7)
Tympanogram B 698 (68.2) 137 (13.4) .35
Tympanogram C 274 (26.8) 68 (31.6)
Pathologic features of tympanic membranes
Atelectasis 633 (61.8) 120 (55.9)
Stage I 562 (54.9) 106 (49.3)
Stage II 51 (5.0) 11 (5.1)
.44
Stage III 10 (1.0) 1 (0.5)
Stage IV 10 (1.0) 2 (0.9)
Attic retraction pocket 768 (75) 154 (71.6)
Stage I 561 (54.8) 111 (51.6)
Stage II 189 (18.5) 38 (17.7)
.30
Stage III 9 (0.9) 4 (1.9)
Stage IV 9 (0.9) 1 (0.5)
Color of tympanic membrane
Pearl gray 593 (57.9) 143 (66.5)
White 41 (4) 19 (8.8) .07
Transparent 179 (17.5) 53 (24.7)
Color of fluid
Amber 904 (88.3) 184 (85.6)
Transparent 5 (0.5) 2 (0.9)
.54
Yellow-white 39 (3.8) 12 (5.7)
Dark 76 (7.6) 17 (7.9)
Volume of fluid
Bubble 89 (8.7) 17 (7.9)
Air-fluid level 192 (18.8) 43 (20.0) .87
Full 743 (72.6) 155 (72.1)
Atrophy 97 (9.5) 20 (9.3) .94
Opacification
Marginal opacification 117 (11.4) 17 (7.9)
.01
Full opacification 53 (5.2) 23 (10.7)
Myringosclerosis 56 (5.5) 9 (4.2) .44

Table 2. Performance in Different Characteristics Groups

Group AUC (95%CI) F1 score (95% CI) Accuracy, % (95% CI)


Male 0.75 (0.62-0.86) 0.91 (0.88-0.95) 84 (79-90)
Female 0.71 (0.60-0.80) 0.86 (0.81-0.91) 76 (68-83)
Age, y
<18 0.65 (0.47-0.81) 0.83 (0.76-0.90) 72 (62-82)
18-60 0.76 (0.67-0.90) 0.89 (0.85-0.93) 80 (75-87)
>60 0.78 (0.63-0.90) 0.92 (0.88-0.96) 86 (78-93)
Tympanogram A 0.78 (0.44-1.00) 0.58 (0.21-0.80) 49 (20-73)
Tympanogram B 0.74 (0.64-0.84) 0.93 (0.90-0.95) 87 (82-91)
Abbreviation: AUC, area under the
Tympanogram C 0.72 (0.60-0.82) 0.85 (0.78-0.90) 74 (66-82)
receiver operating curve.

Performance of the model was lower in the children and C had a higher accuracy than tympanogram A (87%, 74%, 49%,
adolescent (aged 4-18 years) group than in the adult (aged 18-60 respectively) (Table 2). The clinical characteristics of OME ears
years) and older (aged >60 years) groups. Tympanogram B and where CHL was predicted rightly or wrongly by the DL are

jamaotolaryngology.com (Reprinted) JAMA Otolaryngology–Head & Neck Surgery July 2022 Volume 148, Number 7 615

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024
Research Original Investigation A Deep Learning Approach to Predict Conductive Hearing Loss in Patients With Otitis Media With Effusion

Table 3. Performance of the DL Model, Logistic Regression Model, and Otologists

Group AUC (95% CI) F1 score (95% CI) Accuracy, % (95% CI)
Otologist 1 (5 y) NA 0.18 (0.20-0.38) 30 (18-33)
Otologist 2 (10 y) NA 0.09 (0.06-0.24) 16 (11-23)
Otologist 3 (10 y) NA 0.25 (0.22-0.47) 39 (22-48)
ABG of 500 Hz >10 dBa 0.74 (0.65-0.83) 0.87 (0.81-0.90) 78 (70-84)
ABG of 1000 Hz >10 dBa 0.79 (0.68-0.85) 0.92 (0.86-0.94) 85 (78-88)
ABG of 2000 Hz >10 dBa 0.68 (0.63-0.75) 0.53 (0.48-0.59) 66 (60-73)
ABG of 4000 Hz >10 dBa 0.65 (0.55-0.78) 0.81 (0.75-0.87) 72 (63-80) Abbreviations: ABG, air-bone gap;
Averaged ABG>10 dBa 0.74 (0.67-0.81) 0.89 (0.86-0.92) 81 (77-85) AUC, area under the receiver
operating curve; DL, deep learning;
Averaged ABG>20 dBa 0.69 (0.60-0.78) 0.83 (0.77-0.88) 73 (65-80)
NA, not applicable.
Logistic regression model 0.62 (0.45-0.75) 0.82 (0.79-0.85) 76 (70-81) a
Deep learning model.

Table 4. Characteristics of the DL Model Missed and Identified CHL

No. (%)
Characteristic Missed (n = 40) Identified (n = 175) P value
Age, median (IQR), y 39 (18-65) 46 (29-60) .11
Male sex 16 (40) 101 (57.8) .04
Tympanogram
Tympanogram A 6 (15) 4 (2.3)
Tympanogram B 19 (47.5) 118 (67.4) .001
Tympanogram C 15 (37.5) 53 (30.3)
Features of tympanic membranes
Atelectasis 22 (55) 98 (56)
Stage I 21 (52.5) 85 (48.6)
Stage II 1 (2.5) 10 (5.7)
.01
Stage III 0 1 (0.6)
Stage IV 0 2 (1.1)
Attic retraction pocket 24 (60) 130 (74.3)
Stage I 22 (55) 89 (50.9)
Stage II 2 (5) 36 (20.6)
.07
Stage III 0 4 (2.3)
Stage IV 0 1 (0.6)
Color of tympanic membranes
Pearl gray 26 (65) 117 (66.9)
White 3 (7.5) 16 (9.1) .87
Transparent 11 (27.5) 42 (24.0)
Color of fluid
Amber 34 (85) 150 (85.7)
Transparent 0 2 (1.1)
.85
Yellow-white 3 (7.5) 9 (5.1)
Dark 3 (7.5) 14 (8.0)
Volume of fluid
Bubble 8 (20) 9 (5.1)
Air-fluid level 16 (40) 27 (15.4) <.001
Full 16 (40) 139 (79.4)
Atrophy 1 (2.5) 19 (10.9) .10
Opacification
Marginal opacification 3 (7.5) 14 (8.0)
.70
Full opacification 3 (7.5) 21 (12.0)
Abbreviations: CHL, conductive
Myringosclerosis 0 9 (5.1) .14
hearing loss; DL, deep learning.

illustrated in Table 4. A higher percentage of wrong cases was traction pocket (55.0% vs 50.9%), and bubble (20.0% vs 5.1%),
seen in stage I atelectasis (52.5% vs 48.6%), stage I attic re- with few errors in myringosclerosis (0% vs 5.1%).

616 JAMA Otolaryngology–Head & Neck Surgery July 2022 Volume 148, Number 7 (Reprinted) jamaotolaryngology.com

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024
A Deep Learning Approach to Predict Conductive Hearing Loss in Patients With Otitis Media With Effusion Original Investigation Research

Figure. Heatmap Illustration Overlaid by Deep Learning on Original Otoscopic Images

A No tympanic membrane changes

B Air-fluid level

C Atelectasis

D Attic retraction pocket

The red color represents a high


discriminative region of the otoscopic
images for identification of
conductive hearing loss. A, No
tympanic membrane changes;
B, air-fluid level; C and D, atelectasis
and attic retraction pocket.

The DL Model Interpretability


The heatmap (Figure) highlights the changes in the tympanic Discussion
membrane in the otoscopic images. These regions are what
otologists based their estimates of CHL on, such as atelecta- The importance of this novel research to predict hearing level
sis, attic retraction pocket, and air-fluid level. using a DL model is that it provides solutions to reduce pa-

jamaotolaryngology.com (Reprinted) JAMA Otolaryngology–Head & Neck Surgery July 2022 Volume 148, Number 7 617

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024
Research Original Investigation A Deep Learning Approach to Predict Conductive Hearing Loss in Patients With Otitis Media With Effusion

tient waiting time and avoid unnecessary referral. In the pre- changes did not contain sufficient detail to reflect the condi-
sent study, we have demonstrated that DL was used to pre- tion of tympanic membrane, such as air-fluid level, not sup-
dict CHL in averaged ABG greater than 10 dB (accuracy of 81%). plying information on how much fluid was in the tympanic cav-
Notably, the DL model acquired the highest accuracy at 1 kHz ity. However, the DL model used the whole otoscopic image
(85%). The DL model provided values of AUC, F1 score, and ac- as input data to avoid missing important detailed informa-
curacy greater than the logistic regression model and otolo- tion. As a result, the DL model offered a faster and more
gists reviewing otoscopic images. Moreover, the DL model only accurate prediction of CHL than the otologists, which may be
took 2.5 seconds to complete prediction of CHL in the valida- useful when applied to the clinical environment.
tion set, whereas the 3 otologists took 633 seconds, 645 sec- Although the group younger than 18 years showed the low-
onds, and 692 seconds, respectively. The study also demon- est accuracy level (72%), it was still significantly higher than
strated that the DL model used features changes in the the prediction perceived by parents using a symptom ques-
tympanic membrane to predict CHL, such as atelectasis, attic tionnaire in the study by Swierniak et al,37 with an accuracy
retraction pocket, and air-fluid level. of 17.8%. Previous studies have suggested that the diagnostic
Evidence has shown that the applications of artificially in- accuracy for detecting OME using otoscopy is between 60%
telligent technology in health care may be an alternative so- and 70%,38 whereas the sensitivity and specificity of tympa-
lution, or provide additional support, for diagnostic decision nometry can reach to 70% to 90% for the detection of OME,
making. In the field of ENT and clinical audiology, previous but it is dependent on patient cooperation.39,40 Further-
studies have shown DL models to be useful diagnostic tools more, it should be noted that there are several influencing
to assist otologists in terms of distinguishing various otologic factors that affect the observation of the tympanic mem-
diseases on the basis of otoscopic images.17-22 For example, brane and thus the accuracy in diagnosis during otoscopic
Cai et al22 proved the capability of artificial intelligence to examination, such as in the conditions when the external
diagnose OME automatically with an accuracy of 93%. In ad- ear canal is partly or completely blocked by cerumen, or when
dition, a recent literature review29 indicated that DL models children are unable or unwilling to cooperate with the
have achieved accuracy from 75.3% to 99% in predicting noise- examination.41-44 Therefore, it is crucial to obtain high-quality
induced hearing loss. However, it is noteworthy that the DL otoscopic images as input data to ensure a reliable and accurate
algorithm for predicting the degree of CHL using otoscopic im- DL model for predicting CHL in ears with OME.
ages did not attain the excellent performance seen in auto- These results that explore the relationship between oto-
mated diagnosis of ear diseases in previous studies (ie, accu- scopic images and CHL provide great potential for internal
racy over 90%).18-22 The possible explanation is that OME may medicien physicians, general practitioners, pediatricians, and
cause structural alterations in the tympanic cavity,12 not just other clinicians in rural areas to use the current DL model to
tympanic membrane changes. The otoscopic images only pro- predict CHL by sending their otoscopic images to a central-
vide features of the tympanic membrane and fluid, without ized server for diagnosis of OME. Patients can then be triaged
providing other features associated with CHL, such as status to otologists through new medical patterns.
of the structures in the tympanic cavity and mobility of the
tympanic membrane.25,30 Similar experiences occurred in Limitations and Future Studies
studies in other fields. For example, Noorbakhsh et al31 used Although our algorithm achieved an overall accuracy greater
hematoxylin and eosin–scanned images to develop CNN ar- than 80%, there are some limitations to this study. First, ow-
chitectures for tumor or normal and gene variation classifica- ing to the nature of retrospective data, lack of a standardized
tion. Their CNN architectures acquired an AUC of 0.99 in 19 otoscopic image acquisition protocol may have limited their
cancer type diagnosis, with only AUCs of 0.65 to 0.80 in TP53 quality and consistency. Therefore, to improve the DL model
variations. In addition, Coudary et al32 developed a DL model performance, further prospective randomized studies with
that classified lung histopathologic slides into adenocarci- standardized otoscopic image acquisition and annotation
noma (LUAD), squamous cell carcinoma, and normal lung tis- protocols are recommended. Second, the mobility of the tym-
sue. This model showed an averaged AUC of 0.97. When DL panic membrane, status of structures in the tympanic cavity
was used to predict commonly mutated genes in LUAD, the and other clinical factors associated with CHL could have been
AUCs were decreased, ranging from 0.64 to 0.86. included in the prediction model. Classifying on the basis of
The heatmap in the Figure shows that tympanic mem- morphological features of the tympanic membrane and their
brane changes such as atelectasis and attic retraction pocket changes could improve performance accuracy in evaluating
are the distinctive regions used by the DL model to predict CHL. hearing loss.
Previous studies have correlated these features with CHL.25,27
Tympanic membrane changes and fluid could limit the mo-
bility of the tympanic membrane and cause erosion of middle
ear structures.6,25,27,30,33 However, otologists and the logistic
Conclusions
regression model based on tympanic membrane changes show In this diagnostic/prognostic study, a clinically useful DL model
a worse performance in predicting CHL. The possible reason has been developed using DL training on 2790 otoscopic im-
is that these tympanic membrane changes were classified for ages from patients located in multiple centers. These results first
OME diagnostic assessment, rather than CHL assessment.34-36 show that DL algorithms can predict CHL using otoscopic im-
Moreover, the manual classification of tympanic membrane ages to an accuracy of 81%. The DL model acquired a better per-

618 JAMA Otolaryngology–Head & Neck Surgery July 2022 Volume 148, Number 7 (Reprinted) jamaotolaryngology.com

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024
A Deep Learning Approach to Predict Conductive Hearing Loss in Patients With Otitis Media With Effusion Original Investigation Research

formance in less time than a logistic regression model and otolo- A standardized otoscopic images acquisition and annotation pro-
gists at predicting CHL. We intend to improve the performance tocol should be discussed and published, which will further en-
and application of this DL model with a larger prospective study. hance DL used on otoscopic images.

ARTICLE INFORMATION Foundation (Grant No. 2021A1515012038). Dr Zhao 11. Lieberthal AS, Carroll AE, Chonmaitree T, et al.
Accepted for Publication: April 1, 2022. was funded by NIHR (AI Award, 02305), Sêr Cymru The diagnosis and management of acute otitis
III Enhancing Competitiveness Infrastructure Award media. Pediatrics. 2013;131(3):e964-e999. doi:10.
Published Online: May 19, 2022. (MA/KW/5554/19), Cardiff Metropolitan University 1542/peds.2012-3488
doi:10.1001/jamaoto.2022.0900 The Global Academies Research and Innovation 12. Areias B, Parente MPL, Santos C, Gentil F, Natal
Author Affiliations: Department of Development Fund. Jorge RM. The human otitis media with effusion:
Otolaryngology, Sun Yat-sen Memorial Hospital, Role of the Funder/Sponsor: The funding sources a numerical-based study. Comput Methods Biomech
Sun Yat-sen University, Guangzhou, China (J. Zeng, had no role in the design and conduct of the study; Biomed Engin. 2017;20(9):958-966. doi:10.1080/
S. Chen, Deng, Wang, Guisheng Chen, Liang, Wu, collection, management, analysis, and 10255842.2017.1316382
Su, Yiqing Zheng, Cai); The second Hospital, interpretation of the data; preparation, review, or
Medical College, Shantou University, Shantou, 13. Gargeya R, Leng T. Automated identification of
approval of the manuscript; and decision to submit diabetic retinopathy using deep learning.
Guangdong Province, China (Kang); Jarvis Lab, the manuscript for publication.
Tencent, Shen Zhen city, Guangdong Province, Ophthalmology. 2017;124(7):962-969. doi:10.1016/
China (Lin, Ma, Yefeng Zheng); Hong Kong Additional Contributions: We are thankful for the j.ophtha.2017.02.008
University of Science and Technology, Hong Kong, proofreading of Professor Christopher Wigham, 14. Ai T, Yang Z, Hou H, et al. Correlation of chest
China (Lin); Centre for Speech and Language PhD, Swansea University. He was not compensated. CT and RT-PCR testing for coronavirus disease 2019
Therapy and Hearing Science, Cardiff School of (COVID-19) in China: a report of 1014 cases. Radiology.
Sport and Health Sciences, Cardiff Metropolitan REFERENCES 2020;296(2):E32-E40. doi:10.1148/radiol.
University, Wales, United Kingdom (Zhao); 1. Tos M. Epidemiology and natural history of 2020200642
Zhongshan School of Medicine, Sun Yat-sen secretory otitis. Am J Otol. 1984;5(6):459-462. 15. Bai HX, Wang R, Xiong Z, et al. Artificial
University, Guangzhou, China (L. Zeng, Ye); 2. Williamson IG, Dunleavey J, Bain J, Robinson D. intelligence augmentation of radiologist
Department of Otolaryngology–Head and Neck The natural history of otitis media with effusion--a performance in distinguishing COVID-19 from
Surgery, The Third Affiliated Hospital of Sun Yat-sen three-year study of the incidence and prevalence of pneumonia of other origin at chest CT. Radiology.
University, Guangzhou, China (Li, Y. Chen); abnormal tympanograms in four South West 2020;296(3):E156-E165. doi:10.1148/radiol.
Department of Otolaryngology, Zhongshan City Hampshire infant and first schools. J Laryngol Otol. 2020201491
People’s Hospital, Zhongshan Affiliated Hospital of 1994;108(11):930-934.
Sun Yat-sen University, Zhongshan, Guangdong 16. Myburgh HC, van Zijl WH, Swanepoel D,
Province, China (Guoping Chen); Department of 3. Fria TJ, Cantekin EI, Eichler JA. Hearing acuity of Hellström S, Laurent C. Otitis media diagnosis for
Otolaryngology, Shenzhen Baoan Women’s and children with otitis media with effusion. Arch developing countries using tympanic membrane
Children’s Hospital, Shenzhen, Guangdong Otolaryngol. 1985;111(1):10-16. doi:10.1001/archotol. image-analysis. EBioMedicine. 2016;5:156-160.
Province, China (Gao); Shenzhen-Shanwei Central 1985.00800030044003 doi:10.1016/j.ebiom.2016.02.017
Hospital, Sun Yat-sen Memorial Hospital, Sun 4. Roberts J, Hunter L, Gravel J, et al. Otitis media, 17. Myburgh HC, Jose S, Swanepoel DW, Laurent C.
Yat-sen University, Shanwei City, Guangdong hearing loss, and language learning: controversies Towards low cost automated smartphone- and
Province, China (Yiqing Zheng, Cai). and current research. J Dev Behav Pediatr. 2004;25 cloud-based otitis media diagnosis. Biomed Signal
Author Contributions: Drs J. Zeng and Cai had full (2):110-122. doi:10.1097/00004703-200404000- Process Control. 2018;39:34-52. doi:10.1016/
access to all of the data in the study and take 00007 j.bspc.2017.07.015
responsibility for the integrity of the data and the 5. Vanneste P, Page C. Otitis media with effusion in 18. Cha D, Pae C, Seong SB, Choi JY, Park HJ.
accuracy of the data analysis. Drs J. Zeng and Kang children: pathophysiology, diagnosis, and Automated diagnosis of ear disease using ensemble
contributed equally to the work. treatment. A review. J Otol. 2019;14(2):33-39. doi: deep learning with a big otoendoscopy image
Concept and design: J. Zeng, S. Chen, Ma, Zhao, 10.1016/j.joto.2019.01.005 database. EBioMedicine. 2019;45:606-614. doi:10.
Yefeng Zheng, Yiqing Zheng, Cai. 6. Rosenfeld RM, Shin JJ, Schwartz SR, et al. 1016/j.ebiom.2019.06.050
Acquisition, analysis, or interpretation of data: J. Clinical practice guideline: otitis media with effusion 19. Lee JY, Choi S-H, Chung JW. Automated
Zeng, Kang, Deng, Wang, Guisheng Chen, Zhao, executive summary (update). Otolaryngol Head classification of the tympanic membrane using a
Liang, I. Zeng, Ye, Li, Y. Chen, Guoping Chen, Gao, Neck Surg. 2016;154(2):201-214. doi:10.1177/ convolutional neural network. Appl Sci (Basel).
Su, YiQing Zheng, Cai. 0194599815624407 2019;9(9). doi:10.3390/app9091827
Drafting of the manuscript: J. Zeng, Kang, Zhao, I.
Zeng, Ye, Yiqing Zheng, Cai. 7. Ito M, Takahashi H, Iino Y, et al. Clinical practice 20. Khan MA, Kwon S, Choo J, et al. Automatic
Critical revision of the manuscript for important guidelines for the diagnosis and management of detection of tympanic membrane and middle ear
intellectual content: J. Zeng, S. Chen, Deng, Wang, otitis media with effusion (OME) in children in infection from oto-endoscopic images via
Guisheng Chen, Ma, Zhao, Yefeng Zheng, Liang, Li, Japan, 2015. Auris Nasus Larynx. 2017;44(5):501-508. convolutional neural networks. Neural Netw. 2020;
Y. Chen, Guoping Chen, Gao, Su, Yiqing Zheng, Cai. doi:10.1016/j.anl.2017.03.018 126:384-394. doi:10.1016/j.neunet.2020.03.023
Statistical analysis: J. Zeng, Wang, Ye. 8. Bess FH, Dodd-Murphy J, Parker RA. Children 21. Wu Z, Lin Z, Li L, et al. Deep learning for
Obtained funding: Cai. with minimal sensorineural hearing loss: classification of pediatric otitis media. Laryngoscope.
Administrative, technical, or material support: Kang, prevalence, educational performance, and 2021;131(7):E2344-E2351. doi:10.1002/lary.29302
S. Chen, Deng, Guisheng Chen, Yefeng Zheng, functional status. Ear Hear. 1998;19(5):339-354. 22. Cai Y, Yu JG, Chen Y, et al. Investigating the use
Liang, Li, Y. Chen, Guoping Chen, Gao, Su, Yiqing doi:10.1097/00003446-199810000-00001 of a two-stage attention-aware convolutional
Zheng, Cai. 9. Browning GG; MRC Multi-centre Otitis Media neural network for the automated diagnosis of
Supervision: J. Zeng, Kang, Ma, Zhao, Yefeng Zheng, Study Group. Influence of age, type of audiometry otitis media from tympanic membrane images:
I. Zeng, Yiqing Zheng, Cai. and child’s concentration on hearing thresholds. Br a prediction model development and validation
Conflict of Interest Disclosures: None reported. J Audiol. 2000;34(4):231-240. doi:10.3109/ study. BMJ Open. 2021;11(1):e041139. doi:10.1136/
Funding/Support: This work was funded by the 03005364000000133 bmjopen-2020-041139
Medical artificial intelligence project of Sun Yat-sen 10. Kemaloğlu YK, Gündüz B, Gökmen S, Yilmaz M. 23. Chao WY, Wang CF, Chang SJ. Ventilation tube
Memorial Hospital (YXYGZN201904), Key R&D Pure tone audiometry in children. Int J Pediatr in adults with middle-ear effusion. J Otolaryngol.
Program of Guangdong Province, China (Grant No. Otorhinolaryngol. 2005;69(2):209-214. doi:10.1016/ 1999;28(5):278-281.
2018B030339001), National Natural Science j.ijporl.2004.08.018 24. Christensen JG, Wessel I, Gothelf AB, Homøe P.
Foundation of China (Grant No. 82071062), and Otitis media with effusion after radiotherapy of the
Guangdong Basic and Applied Basic Research

jamaotolaryngology.com (Reprinted) JAMA Otolaryngology–Head & Neck Surgery July 2022 Volume 148, Number 7 619

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024
Research Original Investigation A Deep Learning Approach to Predict Conductive Hearing Loss in Patients With Otitis Media With Effusion

head and neck: a systematic review. Acta Oncol. reveal conserved spatial behaviors within tumor media with effusion. Pediatrics. 2004;113(5):1412-
2018;57(8):1011-1016. doi:10.1080/0284186X.2018. histological images. Nat Commun. 2020;11(1):6367. 1429. doi:10.1542/peds.113.5.1412
1468085 doi:10.1038/s41467-020-20030-5 39. Watters GW, Jones JE, Freeland AP. The
25. Ungkanont K, Charuluxananan S, Komoltri C. 32. Coudray N, Ocampo PS, Sakellaropoulos T, et al. predictive value of tympanometry in the diagnosis
Association of otoscopic findings and hearing level Classification and mutation prediction from of middle ear effusion. Clin Otolaryngol Allied Sci.
in pediatric patients with otitis media with effusion. non-small cell lung cancer histopathology images 1997;22(4):343-345. doi:10.1046/j.1365-2273.1997.
Int J Pediatr Otorhinolaryngol. 2010;74(9):1063-1066. using deep learning. Nat Med. 2018;24(10):1559-1567. 00023.x
doi:10.1016/j.ijporl.2010.06.006 doi:10.1038/s41591-018-0177-5 40. Anwar K, Khan S, Rehman HU, Javaid M,
26. Cai T, McPherson B, Li C, Yang F. Pure tone 33. Li Y, Hunter LL, Margolis RH, et al. Prospective Shahabi I. Otitis media with effusion: accuracy of
hearing profiles in children with otitis media with study of tympanic membrane retraction, hearing tympanometry in detecting fluid in the middle ears
effusion. Disabil Rehabil. 2018;40(10):1166-1175. loss, and multifrequency tympanometry. of children at myringotomies. Pak J Med Sci. 2016;
doi:10.1080/09638288.2017.1290698 Otolaryngol Head Neck Surg. 1999;121(5):514-522. 32(2):466-470. doi:10.12669/pjms.322.9009
27. Canali I, Rosito LPS, Longo VD, da Costa SS. doi:10.1016/S0194-5998(99)70049-5 41. Brodie KD, David AP, Kriss H, Chan DK.
Audiometric pattern in moderate and severe 34. Kaleida PH, Fireman P. Diagnostic assessment Outcomes of an early childhood hearing screening
tympanic membrane retraction. Otol Neurotol. of otitis media. Clin Allergy Immunol. 2000;15: program in a low-income setting. JAMA Otolaryngol
2021;42(6):e716-e723. doi:10.1097/MAO. 247-262. Head Neck Surg. 2022. doi:10.1001/jamaoto.
0000000000003099 35. Sadé J, Berco E. Atelectasis and secretory otitis 2021.4430
28. Cai T, McPherson B. Hearing loss in children media. Ann Otol Rhinol Laryngol. 1976;85(2)(suppl 25 42. McManus B, Harbarger C, Grillis A, et al.
with otitis media with effusion: a systematic review. Pt 2):66-72. doi:10.1177/00034894760850S214 Otoscopy and tympanometry outcomes from the
Int J Audiol. 2017;56(2):65-76. doi:10.1080/ 36. Tos M, Stangerup SE, Larsen P. Dynamics of National Health and Nutrition Examination Survey
14992027.2016.1250960 eardrum changes following secretory otitis. a (NHANES). Am J Otolaryngol. 2022;43(2):103332.
29. Chen F, Cao Z, Grais EM, Zhao F. Contributions prospective study. Arch Otolaryngol Head Neck Surg. doi:10.1016/j.amjoto.2021.103332
and limitations of using machine learning to predict 1987;113(4):380-385. doi:10.1001/archotol.1987. 43. Esposito S, Bianchini S, Argentiero A, Gobbi R,
noise-induced hearing loss. Int Arch Occup Environ 01860040042014 Vicini C, Principi N. New approaches and
Health. 2021;94(5):1097-1111. doi:10.1007/ 37. Swierniak W, Gos E, Skarzynski PH, Czajka N, technologies to improve accuracy of acute otitis
s00420-020-01648-w Skarzynski H. The accuracy of parental suspicion of media diagnosis. Diagnostics (Basel). 2021;11(12):
30. Takahashi H, Honjo I, Hasebe S, Sudo M, hearing loss in children. Int J Pediatr Otorhinolaryngol. 2392. doi:10.3390/diagnostics11122392
Tanabe M. The diagnostic and prognostic value of 2021;141:110552. doi:10.1016/j.ijporl.2020.110552 44. Marchisio P, Pipolo C, Landi M, et al; Italian
eardrum mobility in otitis media with effusion. Eur 38. American Academy of Family Physicians; Earwax Study Group. Cerumen: A fundamental but
Arch Otorhinolaryngol. 1999;256(4):189-191. doi:10. American Academy of Otolaryngology-Head and neglected problem by pediatricians. Int J Pediatr
1007/s004050050137 Neck Surgery; American Academy of Pediatrics Otorhinolaryngol. 2016;87:55-60. doi:10.1016/
31. Noorbakhsh J, Farahmand S, Foroughi Pour A, Subcommittee on Otitis Media With Effusion. Otitis j.ijporl.2016.05.014
et al. Deep learning-based cross-classifications

620 JAMA Otolaryngology–Head & Neck Surgery July 2022 Volume 148, Number 7 (Reprinted) jamaotolaryngology.com

© 2022 American Medical Association. All rights reserved.


Downloaded from jamanetwork.com by Abdullah Fatih on 02/29/2024

You might also like