Idiopathic Sudden Sensorineural Hearing Loss (ISSHL) is an otological emergency that occurs within a 72-h window and is characterized by unilateral or bilateral hearing loss of ≥30 decibels at three consecutive audiometric frequencies. Although the cause of ISSHL has not been precisely identified, it is thought to be related to viral infection, vascular impairment, or autoimmune diseases.1 ISSHL is generally treated with high-dose systemic steroids. However, in patients unable to use systemic steroids, Intratympanic Steroid Injection (ITSI) and hyperbaric oxygen treatment are effective as salvage treatments.2 When the patient does not receive proper treatments after the onset of ISSHL, the resulting hearing loss leads to decreased ability of sound localization and speech perception in noise and tinnitus, thereby deteriorating the patient’s quality of life in the long term. Hence, proper and timely treatment is very important.3 Notably, the prediction of prognosis may lead to better treatment results because more additional treatments, such as salvage ITSI or hyperbaric oxygen treatment, can be added when a poor prognosis is predicted.4 Due to developments in artificial intelligence technology, studies have been conducted to predict the prognosis of various diseases using machine learning models, including ISSHL.5–8

The purpose of this study was to overcome the limitations of previous studies, analyze larger number of patients with ISSHL, and predict hearing recovery using the Deep Neural Network (DNN) model. In addition, we identified factors that had prognostic value.

The average age of all included patients was 51.7±14.3 years, and the proportions of males and females were almost the same. Continuous variables, such as age, weight, height, BMI, duration of hospital admission, treatment delay from symptom onset, and hearing thresholds before and after treatment, are summarized in Table 1. Categorical variables, such as sex, history of smoking and alcohol consumption, history of otological diseases, current underlying diseases, initial otological symptoms accompanying hearing loss, and laterality, are summarized in Table 2.

In all patients, the initial average hearing thresholds of the affected and unaffected ears were 76.4±28.3dB and 21.8±23.7dB, respectively. The average hearing thresholds of the affected ear at 2-weeks and 3-months post-treatment were 55.8±36.6dB and 50.4±30.2dB, respectively. The initial and post-treatment hearing thresholds of the affected ear and the initial hearing threshold of the non-affected ear were significantly better in the recovery group than in the non-recovery group (Table 1).

Significant differences in specific variables were evaluated between the two groups. Patients in the recovery group were significantly younger than those in the non-recovery group. The duration of hospital admission was significantly longer in the non-recovery group than in the recovery group. There were significantly more patients with a history of hearing loss, hypertension, and dizziness as an accompanying symptom in the non-recovery group. In contrast, the number of patients with ear fullness as an accompanying symptom was significantly higher in the recovery group.

Treatment outcomes based on Siegel’s criteria were as follows: CR, PR, SI, and NI were achieved in 73 (24.5%), 44 (14.8%), 82 (27.5%), and 99 (33.2%) patients, respectively (Fig. 2). The recovery rate by defining CR and PR as recovery was 39.3%. According to the two groups, the average hearing threshold at each frequency was calculated (Fig. 3).

Figure 2.

Hearing recovery in idiopathic sudden sensorineural hearing loss patients according to Siegel’s criteria.

(0.11MB).

Figure 3.

Initial and post-treatment average hearing levels in the affected and non-affected ears according to the recovery status. (A) Combined treatment recovery group. (B) Combined treatment non-recovery group. AE, Affected Ear; NAE, Non-Affected Ear.

(0.34MB).

The performance of various machine learning methods in predicting the prognosis of ISSHL is summarized in Table 3. The DNN model showed the highest predictive power [accuracy, 88.81%; area under the receiver operating characteristic curve (AUC), 0.9448], followed by the random forest model (accuracy, 86.76%; AUC, 0.9442).

To identify the significant prognostic factors for recovery from ISSHL, the variable importance, which is the difference in loss function between models based on the presence or absence of a specific variable, was calculated in the DNN method. If the value of the variable importance becomes negative, it can be judged as a significant factor for predicting the prognosis. According to the analysis, initial hearing level of the affected ear and non-affected ear, post-treatment (2-weeks) hearing level of the affected ear, smoking, tinnitus, laterality, and BMI were significant factors for predicting the prognosis of ISSHL (Table 4, S2 Table). The whole variable importance values are indicated in S2 Table.

The prognosis of ISSHL is generally predicted using conventional statistical models, such as logistic regression. Recently, studies have utilized machine learning methods to better predict the prognosis of ISSHL.5–8 In this study, we attempted to predict the prognosis of patients with ISSHL using various machine learning methods based on clinical characteristics. Significant differences in various factors were identified between the recovery and non-recovery groups. As demonstrated previously,10–13 patients in the recovery group were younger than those in the non-recovery group.

In our study, hypertension was more common in the non-recovery group. On the other hand, hypertension was not a significant prognostic factor in the DNN method. Some previous studies showed poor prognosis in patients with hypertension,10,12 although other studies found that the presence of hypertension was not a prognostic factor.13,14 In hypertensive patients, the blood vessel elasticity of the inner ear may decrease, thereby causing atherosclerotic change, which may narrow the blood vessels and aggravate damage to the inner ear.15 In addition, we found no differences in the presence of diabetes between the recovery and non-recovery groups. Although some studies showed similar results,12,14 other studies demonstrated poor prognosis in the presence of hyperglycemia.13,16 Regarding the relationship between other systemic diseases and the recovery from ISSHL, poor recovery of hearing loss has been reported in the presence of the metabolic syndrome.15 Some studies have suggested that cardiac disease or cardiovascular accident is not a prognostic factor for ISSHL;12,13 however, others have reported an increased risk of stroke and myocardial infarction after the onset of ISSHL.17 It has also been hypothesized that cochlear microangiopathy due to endothelial vascular abnormalities may affect the prognosis of ISSHL in the presence of comorbidities, such as diabetes or metabolic syndrome.15 Thus, there is controversy on whether underlying diseases such as hypertension and diabetes affect the prognosis in patients with ISSHL. Since the characteristics of the target group vary between studies, the association between the underlying disease and ISSHL remains to be further investigated.

Regarding the early accompanying otologic symptoms associated with ISSHL, we found that there were more patients with ear fullness in the recovery group, and no difference in the number of patients with tinnitus was observed between the recovery and non-recovery groups. Tinnitus was identified as a possible prognostic factor for predicting the prognosis of ISSHL in the DNN method. Previous studies showed that the prognosis was usually better when tinnitus and ear fullness were present,12,18 probably because these symptoms can be recognized more quickly and treatment can be initiated early. Tinnitus indicates the presence of residual hearing,19 and ISSHL patients with dizziness as an initial accompanying symptom often show a poor prognosis.11,13,14,20,21 Notably, it has been suggested that the cochlea and vestibule receive blood supply from the internal auditory artery; therefore, when ischemia occurs, both parts are affected, resulting in a poor prognosis in ISSHL.21 Likewise, we found that recovery was poor when dizziness was an initial accompanying symptom. However, dizziness was not a significant prognostic factor of ISSHL in the DNN method. Therefore, additional research is needed to confirm its role as a prognostic factor.

A poor initial hearing threshold of the affected ear has been demonstrated to be associated with a poor prognosis.11,13,20 In this study, both initial and post-treatment (2-weeks) hearing thresholds of the affected ear were better in the recovery group. Initial hearing threshold and post-treatment (2-weeks) hearing thresholds of the affected ear were also identified as significant prognostic factors of ISSHL in the DNN method. A similar trend was observed in previous studies: higher hearing recovery rates were observed in groups showing quick hearing recovery within 7-days,11,22 or 2 weeks.23 Moreover, we found that there were more patients with a previous history of hearing loss, and the initial hearing threshold of the non-affected ear was worse in the non-recovery group than that in the recovery group. Furthermore, the initial hearing threshold of the non-affected ear was a significant prognostic factor based on the DNN model. This could be due to an existing underlying bilateral hearing dysfunction or systemic disorder.20

Some variables require additional considerations in this study. First, studies have demonstrated that a shorter delay between symptom onset and treatment leads to a better prognosis.13,16,20 In this study, we could not find any differences between the recovery group and the non-recovery group in the delay between symptom onset and treatment. Second, the duration of hospital admission was significantly longer in the non-recovery group than in the recovery group, probably because additional ITSI is often performed when hearing recovery is not observed during hospitalization. Third, in general, additional treatments documented in this study are not well-accepted currently, although some studies have implied that additional treatments, such as lipo-prostaglandins, may affect the prognosis.14

In this study, we found that the DNN method had a better predictive performance than the approaches used in a similar previous study.7 There are several reasons for its better predictive performance. First, the sample size in this study was relatively larger than that in the former study; a larger sample size enhances the prediction performance.24,25 Second, more independent variables were analyzed. The DNN method has better predictive performance than conventional statistical methods and has been widely used recently for prediction and classification in the medical research field.16,26 According to the previous studies,5,6,8,27,28 among machine learning methods, DNN is known to have the best predictive power. However, this method is a so-called “black box” artificial intelligence model, and although its predictive performance is high, it is not clearly known how it works or why the performance is high.29 Therefore, in this study, we used the loss function based on a previous study27 to identify significant prognostic factors and pursue an explainable artificial intelligence model. Variables with negative variable importance (initial hearing level of affected and non-affected ear, post-treatment [2-weeks] hearing level of affected ear, smoking, tinnitus, laterality, and BMI) were considered significant prognostic factors.

This study has several limitations. First, although the sample size was larger than those in the previous study, it was still small for predicting the prognosis with accuracy using a machine learning algorithm. In this study, to increase the number of patients included in the analysis, the ITSI group or oral steroid group was included initially in the study. However, since the clinical characteristics of the patient group according to the treatment method were different, it was judged that it would bias the prognosis predictions; therefore, only the combined treatment group was analyzed. The higher the number of patients, the more precise the prediction; therefore, analysis using well-organized datasets containing more patients selected through well-planned multicenter studies are needed. Second, we only evaluated simple categorical variables, general demographics, and pre- and post-treatment hearing threshold levels. To overcome these limitations, it is necessary to utilize more diverse variables. Recent studies have evaluated auditory brainstem response, otoacoustic emissions,30 neutrophil-to-lymphocyte ratio, and C-reactive protein-albumin ratio31,32 in patients with ISSHL and vestibular function in patients with ISSHL with dizziness21 to predict the prognosis. Therefore, analyzing these variables in future studies may allow the prediction of disease prognosis with better accuracy. Third, in predicting the recovery of ISSHL, the definition of recovery and non-recovery was analyzed as a categorical variable. A previous study has reported a method to calculate the numerical recovery rate by comparing the pre- and post-treatment hearing levels of the affected ear and the contralateral hearing level.20 However, in this study, the difference in loss function was used to identify the importance of factors for predicting prognosis in the DNN method. The application of the numerical recovery rate required an overly complicated process to identify significant factors for predicting prognosis in the DNN method; therefore, the method described in this study was used instead. Lastly, since Gingko Biloba, lipo-prostaglandin, and carbogen are not generally accepted treatments, the treatment outcomes observed in this study should be interpreted with caution.

In conclusion, we evaluated multiple variables to predict the treatment prognosis using machine learning methods in patients with ISSHL. The DNN method showed the highest predictive power (accuracy: 88.81%, AUC=0.9448) and was considered the most useful for prognostic prediction. Initial hearing level of affected and non-affected ear, post-treatment (2-weeks) hearing level of affected ear were significant prognostic factors. Our prognosis prediction model could help design an appropriate treatment plan for patients with ISSHL. Further studies with more patients, more analyzed variables, and randomly assigned treatment methods should be undertaken.

Conceptualization: UTW and HML. Data curation: TWU and SBY. Formal analysis: TWU, SWC and HML. Funding acquisition: HML.dd Methodology: TWU, SJO and HML. Project administration: HML and IWL. Advising: SKK and IWL. Visualization: TWU and HML. Writing-original draft: TWU and HML Writing-review & editing: TWU and HML.

This study was supported by Research institute for Convergence of biomedical science and technology (30-2017-015), Pusan National University Yangsan Hospital, Yangsan, Republic of Korea.

The authors declare no conflicts of interest.