Development of machine learning-based predictive models for evaluating risk in the contralateral ear among patients with unilateral sudden sensorineural hearing loss

Abstract

Background:

Sequential bilateral sudden sensorineural hearing loss (Se-BSSHL) may occur after an initial unilateral episode, but the risk factors for contralateral ear involvement remain unclear. This research aims to identify risk factors associated with the occurrence of contralateral ear in patients suffering from sudden sensorineural hearing loss (SSHL). By leveraging machine learning algorithms, the study uses patients’ demographic and clinical data to develop a predictive model.

Methods:

We conducted an analysis on clinical data from 939 patients admitted to the Ear Department of the Chinese PLA General Hospital from 2008 to 2022, categorizing them into unilateral sudden sensorineural hearing loss (USSHL) and sequential bilateral sudden sensorineural hearing loss (Se-BSSHL) groups. Stratified sampling was executed to maintain a proportional representation of unilateral versus bilateral cases, leading to the creation of seven internal and three external datasets. Variable selection was performed using a decision tree-based recursive feature elimination method. We applied five-fold cross-validation on the internal datasets and conducted model testing on the external datasets. Four machine learning algorithms—random forest, extreme gradient boosting, logistic regression, and support vector machine—were used to build the predictive model and conduct external validation. Model efficacy was primarily appraised using the area under the receiver operating characteristic curve (AUC-ROC), calibration curves, decision curve analysis (DCA), and additional performance metrics including accuracy, precision, recall, and F1-score on the test set. The influence of the variables was illustrated through Shapley value plots. Statistical methods such as the t-test, analysis of variance, and chi-squared test were used.

Results:

We observed a 12.57% (118/939) incidence rate of contralateral ear complications in USSHL patients with otherwise normal hearing. The predictive model incorporated 19 variables, among them, age, hearing loss degree, and sex are the three most significant features influencing the occurrence in the healthy ear of the USSHL group. An increase in age and a more severe degree of hearing loss are risk factors for the occurrence of the disease, and females have a higher risk of disease occurrence. Other high-risk test indicators for disease occurrence can be summarized and categorized into two types of risk factors: lipoprotein dysfunction (total cholesterol, high-density lipoprotein cholesterol, apolipoprotein A1) and inflammation (eosinophils, lymphocyte count, neutrophil count). The AUC of the model on the external test reached 0.77, with an accuracy of 88%, a precision of 84%, a recall of 88%, and an F1 value of 0.85 using the default classification threshold of 0.5, demonstrating robust predictive capabilities.

Conclusion:

This investigation elucidates the risk factors for contralateral ear disease in USSHL patients and establishes a robust predictive model for such occurrence, which offers substantial reference value for managing and preventing complications in USSHL patients.

Introduction

Sudden sensorineural hearing loss (SSHL) is characterized by an abrupt, unexplained onset of sensorineural hearing loss, with a reduction of 20 dB or more at two consecutive frequencies.^[1] Bilateral sudden sensorineural hearing loss (BSSHL) involves sudden deafness that occurs either simultaneously or sequentially in both ears. This condition can be classified based on the interval between the onset in each ear into four categories: (1) simultaneous bilateral sudden sensorineural hearing loss (Si-BSSHL), where the interval is 3 days or less; (2) sequential bilateral sudden sensorineural hearing loss (Se-BSSHL), where the interval exceeds three days; (3) metachronous SSHL, where the interval exceeds 6 months; and (4) successive SSHL, where the interval exceeds 1 year.^[2–4] Recent data indicate an increasing incidence of SSHL, and this trend is similarly observed in BSSHL cases. According to a 1996 report by Fetterman et al,^[5] BSSHL accounted for 1.7% of all SSHL cases, whereas a meta-analysis by Bhat et al^[6] in 2024 showed this figure had risen to 11.9%. The growing prevalence of BSSHL has drawn more attention due to its unclear etiology, unknown mechanisms, and generally poor prognosis. Hearing loss profoundly impacts the physical, mental, and social well-being and the quality of life of patients.^[7,8] Moreover, the rehabilitation demand ranks first among requirements.^[9]

The pathophysiological mechanisms of SSHL may involve vascular complications, infections, immune responses, and tumors.^[1] Si-BSSHL is particularly concerning and is often associated with immune, cardiovascular, and hypertension-related factors.^[6] Scholars assert that Si-BSSHL and USSHL may represent distinct pathological entities, potentially indicative of severe systemic illnesses such as leukemia, metastatic meningeal tumors, cerebrovascular accidents, basilar artery insufficiency, postpartum hemorrhage, and ethylene glycol poisoning.^[10–13] This suggests that Si-BSSHL could serve as a precursor to more severe diseases. Se-BSSHL, a more common subtype of BSSHL, shares many characteristics with USSHL and is often idiopathic.^[6] Yet, the distinctions and relationships between these conditions have not been sufficiently explored in the literature. Consequently, this study will exclude the Si-BSSHL subtype and focus exclusively on investigating the risk factors for contralateral hearing loss in USSHL patients (Se-BSSHL).

In recent years, the application of machine learning and deep learning technologies in otology for disease prediction has seen considerable growth. In the backdrop of the continuous pursuit of precision and personalization in medical research and clinical practice, these advanced technologies have offered novel perspectives and methodologies for the diagnosis, treatment, and prognosis evaluation of otological disorders. For instance, in the study of Meniere’s disease, machine learning algorithms enable the prediction of whether patients will experience bilateral attacks. This predictive capability facilitates early intervention strategies, thereby helping to prevent further deterioration of hearing.^[14] When it comes to diagnosing hearing loss resulting from otitis media, the integration of otoscopic images with deep learning models significantly enhances the accuracy and efficiency of diagnosis. This provides clinicians with more reliable diagnostic evidence.^[15] In the case of cochlear implantation surgery, predicting the postoperative speech rehabilitation outcomes of patients is of paramount importance for devising rehabilitation programs. Machine learning technologies have played a pivotal role in this regard.^[16] Additionally, when assessing the function of the vestibular nerve, these advanced technologies allow for more precise analysis of relevant data, thus assisting physicians in making more accurate diagnoses and formulating appropriate treatment plans.^[17]

At present, logistic regression is the predominant technique used to develop prognostic prediction models for sudden deafness. Research has demonstrated that factors such as age, the timing of medical intervention, the severity of hearing loss, vertigo, and the duration and dosage of steroid administration significantly influence prognosis.^[18,19] SSHL, a frequent emergency in otology, not only raises concerns among patients regarding the treatment outcomes for their affected ear but also generates anxiety about the potential health risks to their unaffected ear. Despite this concern, the scarcity of Se-BSSHL cases has posed a significant obstacle. It has proven arduous to conduct comprehensive and in-depth research aimed at developing a reliable predictive model to assess the risk of disease occurrence in the unaffected ears of patients with SSHL. Accordingly, the objective of this study was to develop and validate a machine learning-based predictive model for estimating the risk of subsequent SSHL onset in the unaffected ear of patients with USSHL, thereby facilitating the early identification and subsequent prevention of disease in these patients.

Methods

Data collection

Data were obtained from the Ear Department’s inpatient records spanning from 2008 to 2022. The inclusion criteria encompassed patients discharged with a diagnosis of SSHL of unknown origin, specifically including 821 cases of USSHL (contralateral ear exhibiting normal hearing) and 118 cases of Se-BSSHL (onset in both ears separated by more than three days). A detailed flowchart illustrating patient selection and model development is shown in Figure 1. Exclusion criteria were as follows: (1) concurrent ear disorders, such as Meniere’s disease, benign paroxysmal positional vertigo, and superior canal dehiscence syndrome; (2) identifiable causes, such as noise-induced hearing loss, blast injuries, and genetic factors; (3) nonacute hearing losses, including progressive or fluctuating hearing loss; (4) cochlear and auditory nerve pathologies, such as auditory neuropathy; (5) absence of follow-up data; (6) simultaneous bilateral sudden hearing loss (onset in both ears within three days); and (7) patients with unilateral sudden hearing loss who also exhibited hearing loss in the contralateral ear (pure-tone audiometry [PTA] ≥20 dB). This study received approval from the Ethics Committee of the PLA General Hospital (approval No. S2017-024-01).

Figure 1.

Flowchart of model construction and evaluation. DCA: Decision curve analysis; ROC: Receiver operating characteristic; Se-BSSHL: Sequential bilateral sudden sensorineural hearing loss; SMOTE: Synthetic minority over-sampling technique; USSHL: Unilateral sudden sensorineural hearing loss.

For patients primarily presenting with an extended duration since the initial onset of hearing loss who did not seek treatment at our facility, data were collected at the time of the second ear’s symptom onset. Collected data included demographic information (age, sex, body mass index [BMI]), initial episode details (triggers, accompanying symptoms, and comorbidities), auditory assessments (hearing loss grade and audiogram type), and laboratory results (complete blood count, coagulation profile, and biochemical tests). These variables were used as features in the predictive model.

Data preprocessing

In this study, a series of critical steps were undertaken for data preprocessing prior to analysis. Initially, for the processing of auditory data, the initial examination results obtained upon hospital admission were used, and auditory profiles were categorized in accordance with the 2015 guidelines for the diagnosis and treatment of sudden deafness in China.^[1] The classification of hearing loss severity adhered to the 2021 World Health Organization grading standards for hearing impairment.^[20] Furthermore, consistent with guidelines of prediction models and missing data and the recommendations from established predictive models,^[21–23] variables exhibiting more than 30% missing data were excluded, while those with 30% or less missing data underwent multiple imputation. Additionally, continuous data were both standardized and discretized, and categorical variables were transformed into numerical formats through one-hot encoding.

Model training

The initial step involved dataset splitting and the application of the synthetic minority over-sampling technique augmentation. The dataset underwent stratified sampling, maintaining a positive-to-negative outcome ratio (Se-BSSHL:USSHL) of 1:7. The internal dataset was divided into seven subsets, while the external test set was comprised of three subsets, each maintaining an equivalent proportion of positive cases. Synthetic minority over-sampling technique was applied to enhance the dataset in the five internal subsets derived from five-fold cross-validation, aiming to adjust the positive-to-negative ratio to 1:2. The external test set was kept in its original form without modifications and was exclusively used for model evaluation.

Subsequently, recursive feature elimination (RFE) based on RFs was used for feature selection within the internal dataset. This technique constructs a random forest (RF), assesses feature importance, and iteratively discards the least significant features. Through this process, the model is trained repetitively, thereby progressively refining the selection of significant features.

Moreover, to address the imbalance in patient numbers between USSHL and Se-BSSHL (original ratio approximately 1:7), four machine learning algorithms were employed: RF, extreme gradient boosting (XGB), logistic regression (LR), and support vector machine (SVM). These algorithms are particularly effective in managing data imbalances. Initially, model training involved five-fold cross-validation, followed by hyperparameter optimization through grid search for each algorithm.

In this study, the sample size was estimated using PASS 2023 (NCSS, LLC, Kaysville, UT, USA) with the aim of evaluating whether the sample size of the predictive model satisfies statistical requirements. A one-sided test was adopted, with the significance level α specified as 0.05 and the statistical power set at 0.9. Given that the ratio of negative to positive cases in the actual data was 7:1, the group-allocation approach of “Enter R = N−/N+, solve for N+ and N−” was selected, with the R value set at 7. The area under the curve (AUC) under the null hypothesis (AUC0) was 0.50, and the AUC under the alternative hypothesis (AUC1) was 0.77. The data type was categorized as discrete (rating), the ratio of the standard deviations between the negative and positive groups was 1.00, and the lower and upper bounds of the false-positive rate were set as 0 and 1, respectively. The software calculations indicated that, to achieve a statistical power of 0.9, the sample size for the positive group (N+) was 11, and that for the negative group (N−) was 77, resulting in a total sample size of 88. The actual statistical power was 0.91462, and the actual value of R was 7, which was in close proximity to the target value. To account for potential data loss that might occur during the study, assuming a 20% data-loss rate, it was ultimately necessary to recruit 14 positive and 97 negative patients, totaling 111 patients.

Model evaluation

Model performance was assessed using the internal validation set and the external test set, employing receiver operating characteristic (ROC) curves, calibration curves, and clinical benefit curves. The optimal cutoff values derived from the ROC curves were used to calculate the AUC, along with additional metrics such as accuracy, precision, F1 score, and recall.

Statistical analysis

Data analysis, model training, and evaluation were conducted using SPSS 26 (IBM Corporation, Armonk, NY, USA) and Python 3.10 (Python Software Foundation, Wilmington, DE, USA). Continuous variables that conform to a normal distribution are expressed as mean ± standard deviation, and were analyzed using t-tests. For continuous variables that do not conform to a normal distribution, they are represented by the median (interquartile range), while the Mann–Whitney U test was adopted. Categorical variables are presented as the number of cases and the percentage (n, %), and they were examined using chi-square tests or Fisher’s exact test. A P value of less than 0.05 was deemed statistically significant.

Result

Demographic and clinical characteristics of patients

The study encompassed 939 participants, including 821 patients diagnosed with USSHL and 118 patients with Se-BSSHL. The estimated incidence rate for Se-BSSHL was approximately 12.57% (118/939). The median age of the entire group was 41 (interquartile range: 28.50–51.00) years, with males comprising 49.20% (462/939) of the cohort. The duration before seeking medical intervention was 16 (7, 30) days, and the hospital stay was 13 (11, 14) days.

Regarding the severity of hearing loss, around 66.86% (549/821) of USSHL patients experienced severe and above impairment, predominantly characterized by a flat steeping audiogram pattern. In contrast, Se-BSSHL patients frequently exhibited more severe hearing loss than USSHL (P <0.05, χ² = 22.26), with 79.66% (94/118) classified as severe and above. The most common type among them was total deafness, affecting 48.31% (57/118) of these patients. Additionally, the Se-BSSHL group was older than USSHL, with the age of 51.50 (39.00, 61.75) years compared to 39.00 (27.00, 50.00) years in the USSHL group (P <0.001, Z = −7.08), and had a higher average BMI (P <0.001, t = −3.37). The proportion of males was also greater in the Se-BSSHL group (P <0.05, χ² = 4.64).

In terms of accompanying symptoms and comorbidities, Se-BSSHL patients demonstrated lower incidence of ear fullness and vertigo compared to those with USSHL (P <0.05, χ² = 6.67; P <0.05, χ² = 5.03), and a higher prevalence of comorbid conditions such as hypertension and diabetes (P <0.001, χ² = 12.75; P <0.001, χ² = 12.76). Statistical differences in these indicators were significant between the two groups [Table 1].

Table 1.

Demographic and clinical characteristics of USSHL and Se-BSSHL patients.

Variables	Total (n = 939)	USSHL (n = 821)	Se-BSSHL (n = 118)	Statistics (t/Z/χ2)	P
Age	41.00 (28.50, 51.00)	39.00 (27.00, 50.00)	51.50 (39.00, 61.75)	−7.08†	<0.001||
Time of onset	16.00 (7.00, 30.00)	17.00 (7.00, 30.00)	14.50 (7.00, 30.00)	−0.94†	0.346
Hospitalization duration	13.00 (11.00, 14.00)	12.00 (11.00, 14.00)	13.00 (11.00, 14.00)	−0.12†	0.907
BMI	23.58 ± 4.00	23.41 ± 4.03	24.74 ± 3.59	−3.37*	<0.001||
Sex				4.64‡	<0.05§
Female	477 (50.80)	428 (52.13)	49 (41.53)
Male	462 (49.20)	393 (47.87)	69 (58.47)
Side of onset				0.06‡	0.813
Left	487 (51.86)	427 (52.01)	60 (50.85)
Right	452 (48.14)	394 (47.99)	58 (49.15)
Hearing loss grade				22.26‡	<0.05§
Normal	1 (0.11)	0 (0.00)	1 (0.85)
Mild	78 (8.31)	75 (9.14)	3 (2.54)
Moderate	83 (8.84)	79 (9.62)	4 (3.39)
Moderate to severe	134 (14.27)	118 (14.37)	16 (13.56)
Severe	198 (21.09)	163 (19.85)	35 (29.66)
Profound	198 (21.09)	174 (21.19)	24 (20.34)
Total deafness	247 (26.30)	212 (25.82)	35 (29.66)
Hearing loss type				2.01‡	0.570
Low-frequency	21 (2.24)	19 (2.31)	2 (1.69)
High-frequency	66 (7.03)	61 (7.43)	5 (4.24)
Flat	430 (45.79)	376 (45.80)	54 (45.76)
Profound	422 (44.94)	365 (44.46)	57 (48.31)
Ear fullness				6.76‡	<0.05§
No	508 (54.10)	431 (52.50)	77 (65.25)
Yes	431 (45.90)	390 (47.50)	41 (34.75)
Tinnitus				2.23‡	0.136
No	52 (5.54)	42 (5.12)	10 (8.47)
Yes	887 (94.46)	779 (94.88)	108 (91.53)
Vertigo				5.03‡	<0.05§
No	305 (32.48)	256 (31.18)	49 (41.53)
Yes	634 (67.52)	565 (68.82)	69 (58.47)
Hypertension				12.75‡	<0.001||
No	796 (84.77)	709 (86.36)	87 (73.73)
Yes	143 (15.23)	112 (13.64)	31 (26.27)
Diabetes				12.76‡	<0.001||
No	860 (91.59)	762 (92.81)	98 (83.05)
Yes	79 (8.41)	59 (7.19)	20 (16.95)

Values are presented as n (%), median (Q₁, Q₃), or mean ± SD. ^*t-value (t-test); ^†Z-value (Mann–Whitney U test); ^‡χ²-value (Chi-squared test); ^§P <0.05, ^||P <0.001. BMI: Body mass index; BSSHL: Bilateral sudden sensorineural hearing loss; M: Median, SD: Standard deviation; Se-BSSHL: Sequential bilateral sudden sensorineural hearing loss; USSHL: Unilateral sudden sensorineural hearing loss.

Model construction and evaluation

Model construction

Using RFE based on RF, 19 significant variables impacting patient outcomes were identified. Four machine learning algorithms—RF, XGB, LR, and SVM—were used to develop predictive models using the internal dataset.

Evaluation of model sensitivity and specificity—ROC curves

The ROC curves and AUC values for all models during internal cross-validation and for the external test set are depicted in Figure 2A–D. All models, except for LR, demonstrated AUC values above 0.94 in internal cross-validation. Notably, RF and XGB models exhibited AUC values exceeding 0.76 in the external test set, indicating their superior performance. The SVM model, however, showed a considerable discrepancy in AUC values between the internal and external evaluations, which may suggest issues of overfitting.

Figure 2.

Evaluation for four machine learning prediction models (RF, XGB, LR, and SVM) with five-fold cross-validation and external test set. (A) Displays the ROC curves of RF; (B) Displays the ROC curves of XGB; (C) Displays the ROC curves of LR; (D) Display the ROC curves of SVM; E–H: Calibration curves. (E) Presents the calibration curves of RF; (F) Presents the calibration curves of XGB; (G) Presents the calibration curves of LR; (H) Presents the calibration curves of SVM; I–L: DCA curves. (I) Shows the clinical benefit curves of RF; (J) Shows the clinical benefit curves of XGB; (K) Shows the clinical benefit curves of LR; (L) Shows the clinical benefit curves of SVM. DCA: Decision curve analysis; LR: Logistic regression; RF: Random forest; ROC: Receiver operating characteristic; SVM: Support vector machine; XGB: eXtreme gradient boosting.

Among these models, the RF emerged as the most effective, achieving an AUC of 0.77 in the external test set. The optimal cutoff value for this model was determined to be 0.31, at which the sensitivity was 0.57. This implies that 57% of the Se-BSSHL patients were accurately identified, while the false-positive rate (1–specificity) was 0.15, indicating that 15% of USSHL patients’ healthy ears were incorrectly predicted to be affected.

Evaluation of model calibration performance—calibration curves

Calibration curves were employed to assess the calibration performance of the models, focusing particularly on the consistency between predicted probabilities and actual observed values. The ideal calibration is depicted by a 45° diagonal line, where predicted probabilities match the actual probabilities exactly. The proximity of the calibration curve to this 45° reference line indicates the model’s calibration accuracy.

As depicted in Figure 2E–H, the calibration curves for the XGB, LR, and SVM models were primarily positioned below the 45° dashed line for both the internal validation set and the external test set. This positioning indicates that the actual probabilities were generally lower than the predicted probabilities. In contrast, the RF model demonstrated a better alignment with the 45° line, indicating superior calibration. Furthermore, the Brier score, another indicator of model calibration performance, showed that lower values correspond to higher prediction accuracy. The Brier score for the RF model in the internal validation set was significantly lower than that for the external test set. Conversely, the Brier scores for the XGB, LR, and SVM models exhibited minimal differences between the internal and external datasets, though all were higher than those of the RF model. In the low–predicted–probability region, the curve of the RF model is positioned below, which implies that the actual positive rate is lower than the predicted probability. This situation may result in the overdiagnosis of low-risk patients. In the high–predicted–probability region, the curve is positioned above, meaning that the actual positive rate is higher than the predicted probability. Consequently, this may lead to the underdiagnosis of high-risk patients.

Curve clinical benefit evaluation of models—decision curve analysis

Decision curve analysis (DCA) was used to evaluate the actual effectiveness of classification models in clinical applications. Among models, the dotted line labeled “Treat all” means treating all patients without considering the prediction results of the model. When the threshold probability is low, the “Treat all” strategy may have a positive net benefit. However, as the threshold probability increases, the net benefit of the “Treat all” strategy drops rapidly and turns negative. Therefore, when the threshold probability is relatively high, if all patients are treated, it will lead to overtreatment. The dashed line labeled “Treat none” indicates that no patients will be treated, and the natural course of the disease is completely relied on. The net benefit of “Treat none” is always zero, which means that if no intervention is carried out for patients, although the problem of overtreatment can be avoided, no benefits can be obtained from the treatment either. In the internal validation set, the DCA curves for all four models were notably higher than those in the external test set. For the external test set, the intersections of the DCA curves with the baseline were at threshold probabilities of 0.54 (RF model), 0.43 (XGB model), 0.09 (LR model), and 0.26 (SVM model). Of these, the RF model exhibited the best performance, demonstrating a net benefit greater than zero in the threshold range from 0.13 to 0.54, thus indicating its positive impact on clinical decision-making [Figure 2I–L].

Other performance metrics for machine learning models

This research evaluated four machine learning prediction models: RF, XGB, LR, and SVM, using five-fold cross-validation and an external test set. On the internal validation set, the RF model achieved an accuracy, precision, recall, and F1-score of 0.98 each, which was comparable to the XGB model (with all metrics being 0.98), and outperformed the LR model (with an accuracy of 0.68) and the SVM model (with an accuracy of 0.92).

The results of the test set demonstrate that the RF model has a significant advantage using the default classification threshold of 0.5. It achieves an accuracy of 0.88, a precision of 0.84, a recall of 0.88, and an F1-score of 0.85. In contrast, the XGB model has an accuracy of 0.87, a precision of 0.85, a recall of 0.87, and an F1-score of 0.86. Compared with the XGB model, the RF model showed slightly higher accuracy and recall, whereas the XGB model showed slightly higher precision and F1-score. The LR model shows poor performance with an accuracy of only 0.67, a precision of 0.85, a recall of 0.67, and an F1-score of 0.73. The SVM model, with an accuracy of 0.79, a precision of 0.82, a recall of 0.79, and an F1-score of 0.80, also fails to match the RF model. In conclusion, the RF model exhibits outstanding performance, with excellent predictive performance and stability, and has broad prospects in practical applications.

Explanation of model variables using a Shapley plot

As previously mentioned, the RF model demonstrates superior performance in predicting the incidence of contralateral ear involvement in cases of USSHL.

The Shapley plot is a tool used to elucidate the predictions made by machine learning models. In this plot, the horizontal axis quantifies the magnitude of feature Shapley values, which correspond to their impact on the prediction outcome. Features are ordered according to their influence. Positive and negative Shapley values signify beneficial or detrimental effects on the outcome, respectively, whereas the color on the vertical axis denotes the magnitude of the feature values themselves. The risk factors for contralateral ear involvement in USSHL include age (older), sex (female), severity of hearing loss (higher), type of hearing loss (more severe), delay before seeking medical consultation (longer), and accompanying symptoms such as ear fullness and vertigo. Additionally, several laboratory indicators are pertinent, including eosinophils (EOS), apolipoprotein A1 (APOA1), high-density lipoprotein cholesterol (HDL-C), fibrinogen (FIB), neutrophils (NEU), and high levels of lymphocytes (LB). Among these, the most influential factors affecting the outcome are age, severity of hearing loss, sex, EOS, and the hearing loss curve, with age being the most significant predictor of contralateral ear incidence [Figure 3].

Figure 3.

Shapley plot illustrating the 19 influential variables in the random forest prediction model. APOA1: Apolipoprotein A1; BMI: Body mass index; FIB: Fibrinogen; HDL-C: High-density lipoprotein cholesterol; LB: Lymphocyte count; EOS: Eosinophils; TCHO: Total cholesterol; URIC: Serum uric acid; NEU: Neutrophil count.

Discussion

Machine learning algorithms find extensive application in the medical domain, particularly in predictive modeling for SSHL, which tends to focus on prognostic predictions.^[18,24,25] The rarity and unpredictable onset of the condition pose significant challenges for tracking patient outcomes and medical histories. To date, no studies have been reported on predictive models specifically for USSHL in the contralateral ear of patients. This study retrospectively analyzed 939 cases of SSHL, including both USSHL and Se-BSSHL. It employed four machine learning algorithms to assess the likelihood of subsequent sudden hearing loss in the unaffected ear of USSHL patients. The developed RF model exhibited an AUC of 0.77 and an accuracy of 0.88, with the incidence of sudden hearing loss in the healthy ear of USSHL patients—who initially had completely normal hearing—reaching as high as 12.57% (118/939). This model represents the first of its kind, both nationally and internationally, to evaluate the risk of SSHL in the contralateral ear following an event. The findings are anticipated to enhance healthcare professionals’ understanding of the risks to the healthy ear in USSHL patients, thereby heightening awareness of potential onset factors and promoting early detection and preventive measures against this condition.

Machine learning algorithms are designed to identify patterns within inconsistent data sets by training on internal data and applying their generalized detection abilities to novel data sets. This process renders them highly effective tools for clinical evaluation and prediction. In this study, four machine learning algorithms—RF, XGB, LR, and SVM—were utilized. These algorithms are particularly adept at managing imbalanced datasets. Among them, the RF algorithm exhibited the superior model performance, achieving a high recall rate of 0.88 using the default classification hreshold of 0.5. This rate indicates that the model successfully identifies the majority of true positives—patients with SSHL in the ostensibly healthy ear—thereby substantially reducing the likelihood of missed diagnoses. Additionally, the model maintains a high precision, suggesting that the predictions it makes regarding the development of the condition are usually accurate, with a low false-positive rate, thus minimizing the potential harm of misdiagnosis. The F1-score, which is the harmonic mean of recall and precision, serves as a critical metric for thoroughly evaluating model performance, particularly in scenarios involving unbalanced outcome categories. In this research, the F1-score of the RF model on the external test set was 0.85, indicative of exemplary model performance. These metrics are vitally important for predictive modeling, and the exceptional performance of the RF model suggests that it can effectively predict the risk of SSHL onset in the contralateral ear of patients.

Clinical practitioners can benefit from understanding the key factors identified by the model, as well as the correlations between these factors and the risk of disease onset. This knowledge can significantly enhance clinical practice. For example, the feature importance illustrated by the Shapley plot enables clinicians to recognize that certain patient characteristics, such as age, the degree of hearing loss, sex, SSHL, and hearing loss curve, significantly influence the risk of disease onset in the healthy ear. When evaluating patients, clinicians should focus on these key factors and combine this information with their clinical experience to assess the overall risk of disease onset in the patient’s healthy ear. The features included in the prediction model, ranked by importance, are: age, hearing loss grade, sex, EOS, type of hearing loss, FIB levels, ear fullness, BMI, APOA1, HDL-C, hypertension, total cholesterol (TCHO), neutrophils count, duration of hospitalization, vertigo, LB, time of onset, side of onset, and serum uric acid (URIC) levels.

Age emerges as a crucial determinant in forecasting the onset of USSHL in the contralateral healthy ear. This correlation may stem from the typical interval between the onset of bilateral conditions, with a median interval of three years reported.^[26] As individuals age, a significant proportion experience progressive, symmetrical high-frequency hearing loss known as presbycusis—approximately one-third of those aged 65–74 years and nearly half of the individuals aged ≥75 years.^[27] The etiology of presbycusis encompasses a complex interplay among aging, oxidative damage, genetic predispositions, and environmental factors. Central to the aging process, the oxidative stress theory posits that reactive oxygen species (ROS) levels escalate with age. This excessive ROS production compromises the antioxidant system and inflicts mitochondrial damage, with mutations and deletions in mitochondrial DNA playing a pivotal role in the development of presbycusis.^[28] ROS are likewise instrumental in the pathogenesis of SSHL, capable of inducing damage and death in cochlear hair cells and triggering inflammatory mediator release in response to infections and immune reactions, thereby exacerbating oxidative damage. Additionally, comorbid conditions such as carotid atherosclerosis, ischemic stroke, diabetes, hypertension, the severity of hearing loss, and audiogram types also significantly influence the progression of age-related SSHL.^[29–31] Treatment outcomes for SSHL in elderly patients vary, ranging from improved to diminished effectiveness compared to outcomes in middle-aged and younger cohorts.^[32–34] Notably, effectiveness markedly declines when bilateral onset is observed.^[35] Prognostic indicators for a poor outcome include older age, greater severity of hearing loss, bilateral involvement, comorbid diabetes, and the presence of dizziness.^[36] In the elderly, hearing loss not only compromises communication capabilities but is also linked with increased risks of dementia, depression, cardiovascular disease, and falls. Moreover, elderly individuals often face socioeconomic challenges such as low income and unemployment, which profoundly affect their psychological, physical, and social well-being. The onset of Se-BSSHL thus warrants particular attention from healthcare providers and family members due to these associated risks.

The degree of hearing loss is closely associated with the risk of onset in the contralateral healthy ear, corroborating previous findings that more severe hearing loss is linked with bilateral SSHL.^[37] Patients with severe or profound SSHL accompanied by dizziness often suffer from abnormal blood supply to the cochlea, whereas cases of total deafness are predominantly caused by inner ear vascular embolism or thrombosis. Inner ear hemorrhage remains a relatively rare cause of profound or total SSHL. Consequently, the potential mechanism for the onset of SSHL in the contralateral ear (Se-BSSHL) may involve sequential damage to the hair cells in both ears, exacerbated by conditions affecting the inner ear’s blood supply. This study also identifies female as a risk factor for the onset of SSHL in the healthy ear. Although males accounted for a higher proportion of the Se-BSSHL8 group in the unadjusted comparison, and multiple clinical studies have supported a higher incidence of Se-BSSHL in males compared with females, the multivariable model suggested that female sex contributed positively to the prediction of Se-BSSHL.^{[2,4,38–40]} The mechanism underlying this association requires further investigation.

The indicators of high risk for the development of contralateral ear SSHL in patients diagnosed with USSHL can be stratified into two principal categories: lipoprotein dysfunction and inflammatory factors.

This study identified that patients with USSHL exhibit a significantly elevated risk of contralateral ear SSHL associated with reduced TCHO and elevated FIB, APOA1 and HDL-C levels. This phenomenon challenges the conventional protective perception of HDL-C. Its apparent paradox can be explained by the “dysfunctional HDL” hypothesis: Under chronic inflammation or vascular injury, elevated HDL-C and APOA1 may signify the accumulation of defective lipoprotein particles—characterized by loss of cholesterol reverse transport capacity and acquisition of pro-inflammatory/pro-oxidant properties—rather than an increase in functional protective particles. Concurrently, reduced TCHO likely reflects systemic inflammation-induced lipid consumption or metabolic imbalance.^[41] These factors act synergistically to disrupt cochlear microcirculatory homeostasis. As terminal vessels, the spiral modiolar arteries exhibit high sensitivity to hemorheological abnormalities and oxidative stress, ultimately mediating increased SSHL risk in the contralateral ear.^[42]

Recent studies have demonstrated that elevated levels of specific inflammatory markers—namely, EOS, LB, and NEU—significantly increase the risk of contralateral ear involvement in patients with USSHL. EOS is indicative of an allergic response within the body, while heightened levels of LB and NEU often predict the presence of viral or bacterial infections. Together, these markers suggest that immune or inflammatory responses may be occurring either systemically or within the inner ear. Notably, allergic reactions such as urticaria are associated with the onset of BSSHL, typically characterized by simultaneous bilateral hearing loss that progressively worsens.^[43,44] Viral infections are also recognized as significant etiological factors for SSHL, with three potential mechanisms implicated: direct viral infection of the cochlea or auditory nerve, possibly via hematogenous spread or through cerebrospinal fluid; reactivation of latent viruses within the inner ear; and peripheral infections that trigger an immune response, leading to a secondary antigen–antibody reaction in the inner ear.^[45] Cytomegalovirus is frequently identified in cases of virus-associated hearing loss, which can manifest heterogeneously as either unilateral or bilateral, and range from mild to severe impairment.^[46] Previous research has identified several viruses that can lead to BSSHL, including congenital cytomegalovirus, rubella virus, lymphocytic choriomeningitis virus, acquired measles virus, West Nile virus, and the novel coronavirus.^[47,48] However, these studies did not differentiate among BSSHL subtypes. Preliminary analyses by our research group have shown that clinical characteristics of BSSHL patients often include a higher incidence of viral infections compared to those with USSHL.^[49] Additionally, among BSSHL patients, a notable prevalence of co-infection with four specific pathogens—syphilis, cytomegalovirus, arboviruses, and herpes simplex virus—was observed at 3.9%.^[26] These findings, however, did not extend to subtype classification, indicating a need for more detailed investigation in this area.

In this study, sophisticated machine learning algorithms, adept at handling complex datasets, were utilized to develop the inaugural model for predicting the risk of onset in the healthy ear following USSHL. By incorporating a variety of factors—including patients’ basic demographic information, medical history, clinical symptoms, and laboratory results—a comprehensive disease prediction model was constructed, thereby addressing a previously unexplored area in this field. The model demonstrated robust performance, achieving an optimal AUC of 0.77. The risk factors identified as influencing the onset in the contralateral ear in SSHL patients include increasing age, females, severe hearing loss, lipoprotein dysfunction and systemic inflammation, which may manifest as allergic reactions or bacterial/viral infections. These findings carry significant implications for the management and prevention strategies for SSHL. However, this study is constrained by its retrospective design; it exclusively analyzed data from hospitalized patients, thus limiting its applicability to the outpatient population or individuals with less severe conditions. It is crucial to acknowledge that the predictive model only identifies correlations between variables and outcomes and does not establish causality. This limitation underscores the need for further research to explore causal relationships. Future developments in prediction models should aim to establish multicenter databases to facilitate external validation and refinement of the models. Additionally, there is potential for enhancing the accessibility and application of these models through web-based interfaces and other digital platforms. Further advancements could also involve the integration of multimodal data, including genomic and radiomic data, to augment the diagnostic and predictive accuracy of these models.

Acknowledgments

The authors would like to express sincere thanks to Gao Yun for their valuable assistance and support during the collection and for providing statistical analysis. The authors are also grateful to Wang Dayong for their helpful comments and suggestions.

Funding

This study received support from the National Natural Science Foundation of China (General Program) (No. 82171130), and the National Key Research and Development Program of China (Nos. 2023YFC2508400, 2023YFF1203504, and 2023YFC2509800).

Conflicts of interest

None.