Machine Learning Models Can Predict Tinnitus and Noise-Induced Hearing Loss

Abstract

Objectives:

Despite the extensive use of machine learning (ML) models in health sciences for outcome prediction and condition classification, their application in differentiating various types of auditory disorders remains limited. This study aimed to address this gap by evaluating the efficacy of five ML models in distinguishing (a) individuals with tinnitus from those without tinnitus and (b) noise-induced hearing loss (NIHL) from age-related hearing loss (ARHL).

Design:

We used data from a cross-sectional study of the Canadian population, which included audiologic and demographic information from 928 adults aged 30 to 100 years, diagnosed with either ARHL or NIHL due to long-term occupational noise exposure. The ML models applied in this study were artificial neural networks (ANNs), K-nearest neighbors, logistic regression, random forest (RF), and support vector machines.

Results:

The study revealed that tinnitus prevalence was over twice as high in the NIHL group compared with the ARHL group, with a frequency of 27.85% versus 8.85% in constant tinnitus and 18.55% versus 10.86% in intermittent tinnitus. In pattern recognition, significantly greater hearing loss was found at medium- and high-band frequencies in NIHL versus ARHL. In both NIHL and ARHL, individuals with tinnitus showed better pure-tone sensitivity than those without tinnitus. Among the ML models, ANN achieved the highest overall accuracy (70%), precision (60%), and F1-score (87%) for predicting tinnitus, with an area under the curve of 0.71. RF outperformed other models in differentiating NIHL from ARHL, with the highest precision (79% for NIHL, 85% for ARHL), recall (85% for NIHL), F1-score (81% for NIHL), and area under the curve (0.90).

Conclusions:

Our findings highlight the application of ML models, particularly ANN and RF, in advancing diagnostic precision for tinnitus and NIHL, potentially providing a framework for integrating ML techniques into clinical audiology for improved diagnostic precision. Future research is suggested to expand datasets to include diverse populations and integrate longitudinal data.

INTRODUCTION

Age-related hearing loss (ARHL), noise-induced hearing loss (NIHL), and tinnitus are prevalent auditory disorders that significantly impact the adult population (Jafari et al. 2019). ARHL, or presbyacusis, is a gradual decline in hearing ability resulting from aging, affecting nearly one-third of adults over 65 and over half of those aged 75 and older (Yamasoba et al. 2013; Jafari et al. 2020a, 2021). This condition is characterized by the deterioration of the sensory cells and neural pathways of the auditory system as part of the natural aging process (Gates & Mills 2005; Jafari et al. 2023). NIHL, resulting from prolonged exposure to high noise levels, contributes significantly to the burden of hearing loss, particularly in occupational settings (Jafari et al. 2020a, b; Natarajan et al. 2023). This type of hearing loss is often characterized by high-frequency sensorineural hearing loss owing to permanent damage to the hair cells in the cochlea (Nelson et al. 2005). Tinnitus, a condition characterized by the perception of sound without an external source is linked to various etiologies, particularly aging and noise exposure. Tinnitus affects approximately 10 to 15% of adults globally, with prevalence increasing with age and exposure to loud noises (Baguley et al. 2013; McCormack et al. 2016). It is notably higher among those with hearing impairments (Bhatt et al. 2016) making it a significant comorbidity in individuals with ARHL and NIHL (Jafari et al. 2020a, 2022).

Predicting tinnitus and differentiating NIHL from ARHL based on audiologic tests poses significant clinical challenges. Traditional diagnostic approaches rely on detailed audiological evaluations and patient histories (Jafari et al. 2022); however, these methods can be confounded by overlapping symptoms and the subjective nature of tinnitus (Sadegh-Zadeh et al. 2024). Furthermore, the latency period between noise exposure and the onset of NIHL complicates the differentiation from ARHL, which typically progresses more linearly with age (Dobie 2008). For instance, individuals with NIHL may not exhibit symptoms until years after the initial noise exposure, making it challenging to distinguish NIHL from ARHL, which develops gradually. Assessing noise exposure history is critical for an accurate diagnosis, but these assessments are often unreliable because of difficulties in accurately recalling and quantifying past noise exposure levels (Rabinowitz 2000). Tinnitus diagnosis is further complicated by its subjective reporting, varying degrees of severity, and predominately accompanying hearing loss (Henry et al. 2005; Han et al. 2009). Accurate differentiation is crucial for tailoring appropriate interventions, as the prevention and management strategies for ARHL and NIHL are different. For example, whereas ARHL management might focus on hearing aids and aural rehabilitation, NIHL prevention emphasizes noise control and hearing protection strategies (Pelegrin et al. 2015). These complexities underscore the importance of developing advanced diagnostic tools to accurately differentiate between these conditions.

In recent years, machine learning (ML) techniques have emerged as powerful tools in medicine and health, offering the potential to enhance diagnostic accuracy by analyzing large datasets and identifying patterns that may not be discernible through traditional methods (Ebnali et al. 2023; Hosseini et al. 2023). Various studies have applied ML models to audiometric data to predict tinnitus and differentiate between types of hearing loss. For example, Tomiazzi et al. (2019) evaluated the performance of ML algorithms for pattern recognition and classification of hearing impairment in Brazilian farmers. The study included 127 participants, and data were evaluated using artificial neural networks (ANN), K-nearest neighbors (K-NN), and support vector machines (SVM). The ML classifiers achieved good classification performance (control and exposed), with the K-NN test showing the best classification results in extended high-frequency threshold datasets (9 to 16 kHz, approximately 90% accuracy). Zhao et al. (2019a) explored ML models for predicting hearing impairment in 1113 workers from 17 factories located in Zhejiang Province, China, who were exposed to complex industrial noise. The prediction accuracy ranged from 78.6 to 80.1%, indicating that the 4 classification models—SVM, neural network multilayer perceptron, random forest (RF), and adaptive boosting—could be valuable tools for assessing NIHL among workers exposed to various complex occupational noises. Using both the receiver operating characteristic-area under the curve (ROC-AUC) and prediction accuracy, the SVM model demonstrated the highest performance, with an AUC value of 0.808, for predicting hearing impairment from occupational noise exposure. Elkhouly et al. (2023) developed a data-driven audiogram classifier utilizing data normalization and multi-stage feature selection, significantly improving the classification accuracy of various audiogram types. The study, which worked with a large dataset of 28,244 audiograms from a database at the Department of Audiology at Stockholm South Hospital (Bisgaard et al. 2010), used unsupervised spectral clustering to classify the audiograms according to their shapes. The results demonstrated better performance compared with existing models, with accuracy, precision, recall, specificity, and F-score values ranging from 0.93 to 0.99%. A systematic review by Chen et al. (2021) examined studies using ML to predict NIHL, including 8 studies with sample sizes ranging from 150 to 10,567. The most commonly used models were ANN, RF, and SVM. Accuracy ranged from 75.3 to 99%, with low prediction error/root-mean-square error in 3 studies. Despite these advancements, further research is needed to validate these models across diverse populations and to refine methodologies for feature selection and model evaluation.

Although ML models have been widely applied in health sciences to predict outcomes and classify conditions, their application in differentiating various types of auditory disorders is limited. Recent studies demonstrate the usefulness of ML models for predicting hearing loss (Lenatti et al. 2022; Gathman et al. 2023), NIHL (Zhao et al. 2019b; Madahana et al. 2024; Soylemez et al. 2024; Tian et al. 2024), tinnitus (Allgaier et al. 2022; Manta et al. 2023), and tinnitus-related distress (Manta et al. 2023). However, few studies have incorporated a large dataset that includes a control group to investigate the applicability of ML models for differentiating NIHL from ARHL and individuals with tinnitus from those without. To address this gap in the literature, we applied five ML models—ANNs, K-NN, logistic regression (LR), RF, and SVM—to a large demographic and audiometric dataset of adults aged 30 to 100 years with ARHL, including a subset with a history of long-term occupational noise exposure. The subset with chronic occupational noise exposure could have combined NIHL and ARHL, referred to as the NIHL group throughout this article. The study had two main objectives: (1) to assess whether ML models can predict the presence of tinnitus and differentiate individuals with tinnitus from those without, and (2) to determine if ML models can distinguish NIHL from ARHL based on demographic and audiometric data. We hypothesized that ML models provide a feasible approach to improve the accuracy of these differentiations compared with traditional diagnostic methods, facilitating more precise diagnoses and informing better-targeted interventions in the future. The significance of this study lies in its potential to assist in transforming clinical decision-making processes by leveraging advanced data analyses to address challenges in diagnosing tinnitus and differentiating NIHL from ARHL.

MATERIALS AND METHODS

Study Design and Participants

For this study, we used audiologic and demographic information from a dataset including records of 928 participants aged 30 to 100 years (605 men and 323 women) with hearing loss (i.e., hearing thresholds of 25 dB HL or worse from 0.25 to 8 kHz) (Omidvar et al. 2018) and/or tinnitus collected in Lethbridge, AB, Canada, between 2015 and 2019 (Jafari et al. 2020a). Participants with tinnitus reported either constant tinnitus (a permanent sensation of sound) or intermittent tinnitus (episodes of sound sensation lasting more than 5 minutes per week) (Henry et al. 2010) in one or both ears for at least 6 months (Jafari et al. 2019, 2020a). The dataset included demographic and auditory information, consisting of the year of birth, gender (men, women), tinnitus status (constant tinnitus [C-T], intermittent tinnitus [I-T], or no tinnitus [N-T]), hearing thresholds at standard audiometric frequencies (0.25, 0.5, 1, 1.5, 2, 3, 4, 6, 8 kHz), Word Recognition Test (WRT) in quiet, and the static compliance of the tympanic membrane (mmho). The WRT entailed the percentage of correctly repeated words from separate lists of phonetically balanced (PB-50) monosyllabic words (CID W-22). The eardrum static compliance was measured by the tympanometry test using the Titan/IMP440 (Interacoustica, Middelfart, Denmark) with a 226 Hz low-frequency probe tone to confirm the absence of middle ear problems.

In this study, the ground truth for tinnitus and hearing groups (tinnitus versus non-tinnitus and NIHL versus ARHL) was established through a combination of clinical assessments and audiologic data from a cross-sectional study of 928 adults, which included detailed audiometric evaluations and demographic information. The motivation for applying ML models stems from their capacity to analyze complex patterns within these datasets, enabling objective differentiation of auditory disorders like NIHL and ARHL. Unlike manual classification methods that rely solely on self-reported data, ML models integrate multiple variables—such as audiometric thresholds and demographic features—to improve diagnostic precision and potentially uncover subtle patterns or risk factors. This approach supports scalable, data-driven predictions, complementing clinical workflows and addressing the variability and subjectivity inherent in manual classification methods.

To focus the study on the impact of aging and/or chronic occupational noise exposure, individuals with a family history of hearing loss, chronic ear infections, ototoxic medication use, sudden hearing loss, known ear diseases, dizziness, diabetes, stroke, hypertension, atherosclerosis, cardiovascular diseases, ear surgery, and neurodegenerative diseases were excluded. The study included 2 groups: the ARHL group (n = 497, 47.28% men), with no self-reports of exposure to environmental or occupational noise, serving as the control group; and the noise group (n = 431, 91.30% men) with NIHL, based on self-reports of long-term occupational noise exposure. This exposure included noise from industries such as oil and gas, steel, food, heavy machinery, railroads, power tools, mining, construction, farming, military, air force, fire service, hunting, shooting, carpentry, laundry, and bus driving. This study was approved by the University of Lethbridge Human Participant Research Committee, REB Protocol #2020-009, in accordance with university policy and the Tri-Council Policy Statement: Ethical Conduct for Research Involving Humans.

Data Preprocessing

The dataset underwent preprocessing to ensure its suitability for ML analyses. Initially, any infinite values were considered missing data and removed to maintain data integrity. These values appeared in two cases due to calculation errors during data collection. They were removed from the analysis to ensure data integrity and prevent any potential bias in the model’s performance. Given the rarity of these occurrences, their removal did not affect the representativeness of the dataset. Hearing thresholds were averaged into three distinct frequency bands: the low band (average of thresholds at 0.25 and 0.5 kHz for both ears), the medium band (average of thresholds at 1, 1.5, and 2 kHz for both ears), and the high band (average of thresholds at 3, 4, and 6 kHz for both ears).

Data Included in ML Models

Five variables were included in the ML models to differentiate individuals with tinnitus from those without, including year of birth, averaged hearing thresholds in three frequency bands, WRT scores, eardrum static compliance, and chronic noise exposure (yes/no). To investigate the ability of ML models to differentiate ARHL from NIHL, we considered the following five variables: year of birth, averaged hearing thresholds across three frequency bands, WRT scores, eardrum static compliance, and tinnitus status (yes/no). These variables were selected as they represent the fundamental hearing-related data available in the dataset.

ML Models

The five most common ML algorithms were employed to predict the noise exposure group using the selected features (Mahesh 2020).

Logistic Regression

LR is a linear model used for binary classification problems (Tan et al. 2016). It models the probability of the default class using a logistic function, which is particularly useful for datasets where the relationship between the features and the target variable is linear. For this study, the LR model was optimized for a maximum of 1000 iterations to ensure convergence (Agakov et al. 2006). Convergence typically occurred between 100 and 500 iterations, as confirmed by monitoring the loss function. An early stopping criterion was applied, terminating training if the relative improvement in the loss function fell below 1e-4 for 10 consecutive iterations. This threshold was chosen based on established practices in ML (Pedregosa et al. 2011) to balance convergence with computational efficiency. Loss and accuracy metrics were monitored throughout training, and no significant performance degradation, indicative of overfitting, was observed in the validation set.

K-Nearest Neighbors

K-NN is a nonparametric algorithm used for classification and regression tasks. For classification, it assigns the class of a sample based on the majority class among its K-NNs (Guo et al. 2003). The distance metric, typically Euclidean distance, is used to determine the nearest neighbors. K-NN is known for its simplicity and effectiveness in scenarios where the decision boundary is not linear. In this study, the number of neighbors was optimized using GridSearchCV to find the best-performing k-value (Ahmad et al. 2022).

Support Vector Machine

SVM is a powerful classifier that works by finding the hyperplane that best separates the classes in the feature space (Suthaharan 2016). SVM can use different kernel functions (linear, polynomial, radial basis function) to handle nonlinear classification problems. The algorithm aims to maximize the margin between the support vectors of the two classes. Hyperparameters such as the penalty parameter C and the kernel coefficient gamma (Kalita et al. 2020), which defines the influence of a single training example, were optimized using GridSearchCV (Ahmad et al. 2022). SVM was chosen for its robustness and effectiveness in high-dimensional spaces.

Artificial Neural Network

ANNs are inspired by the structure of the human brain and consist of interconnected nodes (neurons) organized in layers (Wu et al. 2014). ANNs were selected for their ability to model complex, nonlinear relationships in the data (Saddi Kadhim et al. 2022). Their layered architecture includes an input layer, one or more hidden layers, and an output layer. Each neuron in a layer receives input from neurons in the previous layer, applies a weighted sum followed by an activation function, and passes the result to the next layer. In this study, the ANN model was optimized to achieve robust performance. Hyperparameter tuning was conducted using GridSearchCV to systematically evaluate combinations of key parameters, including the number of hidden layers (1 to 3), neurons per layer (10 to 100), activation functions (ReLU, tanh), solvers (SGD, Adam), and learning rates (0.001 to 0.01). Regularization was applied using an L2 penalty (α values ranging from 0.001 to 0.1) to mitigate overfitting. The training was performed using the Adam optimizer, which adapts learning rates during optimization for improved convergence. In addition, early stopping was implemented to terminate training if the validation loss did not improve after 10 epochs, preventing overfitting and enhancing generalizability.

Random Forest

RF is an ensemble learning method that constructs multiple decision trees during training and outputs the mode of the classes (classification) or mean prediction (regression) of the individual trees (Qi 2012). Each tree in the forest is built from a bootstrap sample of the training data, and at each split in the tree, a random subset of features is considered for splitting. This randomness helps to ensure that the model is robust and reduces the risk of overfitting. The number of trees in the forest (n_estimators) and other hyperparameters were optimized using GridSearchCV (Ahmad et al. 2022).

Model Training and Evaluation

The dataset was split into training and test sets using an 80-20 split, with stratification based on the target variable to ensure balanced classes (Dobbin & Simon 2011). In addition, fivefold cross-validation was used during training to validate the model’s performance across multiple subsets of the data, ensuring robustness and mitigating the risk of bias from a single train-test split. The preprocessing pipeline included standardizing numerical features using StandardScaler and one-hot encoding categorical features (Pedregosa et al. 2011). This pipeline was applied consistently to both the training and test sets across all models to ensure comparability. To assess the performance of the ML models, several evaluation metrics were used, including accuracy, precision, recall, F1-score, and ROC-AUC, which were averaged across the folds during cross-validation to provide a more reliable assessment of model performance. Accuracy measures the proportion of correctly predicted instances among the total instances, offering a general sense of the model’s performance (Sokolova & Lapalme 2009). Precision, the ratio of true positive predictions to total predicted positives, indicates how many predicted positives are actually correct (Hicks et al. 2022). Recall, also known as sensitivity, is the ratio of true positive predictions to total actual positives, reflecting the model’s ability to identify all relevant instances (Sokolova & Lapalme 2009). The F1-score is the harmonic mean of precision and recall, providing a single metric that balances both concerns (Powers 2011). In addition, the ROC-AUC score evaluates the model’s ability to distinguish between classes, with the ROC curve plotting the true positive rate against the false-positive rate at various threshold settings. The AUC score indicates the likelihood that the model will rank a randomly chosen positive instance higher than a randomly chosen negative one, with a score closer to 1.0 representing excellent discrimination capability (Bradley 1997). To address the class imbalance observed in the dataset, the widely used oversampling technique, Synthetic Minority Over-sampling Technique, and class weighting were incorporated during model training. These methods aimed to balance the representation of the No-T class and improve model performance metrics, particularly precision and recall.

Statistical Analysis

All statistical analyses were conducted using Python libraries including NumPy, Pandas, Scikit-learn, SciPy, and Statsmodels at a significant level of 0.05 or better. To compare the audiometric characteristics between individuals with and without tinnitus, χ² tests (categorical variables) and t tests (numerical values) were conducted. The Friedman test was used for within-group comparison (across three frequency bands: low, medium, and high), and the analysis of variance test was conducted for between-group comparison (across three tinnitus groups C-T, I-T, and N-T). Violin plots were generated to visualize the severity of hearing loss across three frequency bands NIHL and ARHL groups divided into three tinnitus subgroups. The ML models were evaluated using cross-validation to measure accuracy, precision, recall, F1-score, and ROC-AUC.

RESULTS

Demographic Results

Table 1 presents the demographic characteristics of the population included in the ML analyses. Tinnitus prevalence was significantly greater in the NIHL versus ARHL group (46.40% versus 19.71%) in both constant (27.85% versus 8.85%) and intermittent (18.55% versus 10.86%). Figure 1A summarizes the types of occupational noise, divided into seven categories (transportation, music, mechanical/technical, military/police/fire service, farming/agriculture, and others) in the NIHL group. Figure 1B ranks the percentage of individuals with tinnitus in each noise category, with 78% being the highest in the recreational/music group and 38% being the lowest in the farming/agriculture group. Figure 1C exhibits the percentage of individuals in each tinnitus category, highlighting the higher frequency of tinnitus in the NIHL relative to the ARHL group.

TABLE 1.

Descriptive characteristics of variables included in the machine learning models

	NIHL (n = 431)		ARHL (n = 497)
Variable	n	%	n	%	p
Men (n, %)	392	91.30	235	47.28
Women (n, %)	39	8.30	262	51.72
Tinnitus (n, %)	200	46.40	98	19.71	<0.001
Constant (n, %)	120	27.85	44	8.85	<0.001
Intermittent (n, %)	80	18.55	54	10.86	<0.001
	NIHL (n = 431)		ARHL (n = 497)
Variable	n	Mean (SD)	n	Mean (SD)	p
Age (y)	431	73.37, 15.64	497	69.16, 13.67
PTA2,4,8-r, (dBHL)	58.65	18.41	50.25	21.37	<0.001
PTA2,4,8-l, (dBHL)	60.31	18.64	50.95	21.06	<0.001
WRT-r (%)	83.74	17.36	80.7	17.40	<0.001
WRT-l (%)	83.24	17.57	80.06	17.44	<0.001
Tymp-r (mmho)	0.84	0.54	0.66	0.37	<0.001
Tymp-l (mmho)	0.82	0.47	0.64	0.34	<0.001

ARHL, age-related hearing loss; l, left ear; NIHL, noise-induced hearing loss; PTA, pure-tone audiometry; r, right ear; WRT, Word Recognition Test.

Fig. 1.

Demographic characteristics of the study population. A, Types of occupational noise in the NIHL group. B, The percentage of tinnitus reported per type of exposed noise. C, The percentage of individuals in each tinnitus category. ARHL indicates age-related hearing loss; NIHL, noise-induced hearing loss. ***p < 0.001.

Between-Group Comparisons (NIHL Versus ARHL)

Figure 2 shows violin plots comparing hearing loss severity between NIHL and ARHL across three frequency bands, with further division into three tinnitus subgroups (C-T, I-T, and N-T). In the low-frequency band, no significant difference was observed between the two groups. In the medium-frequency band, hearing loss severity was greater in the NIHL compared with the ARHL across all three tinnitus subgroups: C-T (p = 0.002), I-T (p = 0.019), and N-T (p < 0.001). The same results were found in the high-frequency band, with greater hearing loss severity in the NIHL versus ARHL across all three tinnitus subgroups: C-T (p < 0.001), I-T (p < 0.001), and N-T (p < 0.001). Table S1, Supplemental Digital Content (http://links.lww.com/EANDH/B652) provides detailed statistical results for between-group comparisons.

Fig. 2.

Violin plots comparing hearing loss severity across (A) low-, (B) medium-, and (C) high-frequency bands, between NIHL and ARHL groups divided into three tinnitus subgroups (C-T, I-T, and N-T). Medians are shown. **p < 0.01, or ***p < 0.001. Tables 2 and 3 present detailed statistical results for within and between-group comparisons, respectively. ARHL indicates age-related hearing loss; C-T, constant tinnitus; I-T, intermittent tinnitus; NIHL, noise-induced hearing loss; N-T, no tinnitus.

Within-Group Comparisons Across Three Tinnitus Subgroups

In the noise group, there was a significant difference between I-T and N-T (p < 0.001) in the low-frequency band, between C-T and N-T (p = 0.015) and I-T and N-T (p = 0.002) in the medium-frequency band, and between C-T and N-T (p = 0.004) and I-T and N-T (p < 0.001) in the high-frequency band.

In the control group, there was a significant difference between C-T and N-T (p < 0.001) and I-T and N-T (p = 0.041) in the low-frequency band, between C-T and N-T (p < 0.001) and I-T and N-T (p < 0.001) in the medium-frequency band, and between C-T and N-T (p = 0.012) and I-T and N-T (p < 0.001) in the high-frequency band. Table S2, Supplemental Digital Content (http://links.lww.com/EANDH/B652) provides detailed statistical results for within-group comparisons across the three tinnitus subgroups in each noise and control group.

In all comparisons, hearing loss severity was significantly greater in the N-T group compared with the C-T group.

Distinguishing Individuals With Tinnitus From Those Without

Table 2 presents performance metrics for five ML models in differentiating individuals with tinnitus (either constant or intermittent) from those without. Key metrics include precision, recall, F1-score for both tinnitus (C-T and I-T) and N-T categories, overall accuracy, and ROC-AUC.

TABLE 2.

Performance metrics of five ML models for differentiating individuals with tinnitus from those without

Model	Precision (Tin)	Precision (No-Tin)	Recall (Tin)	Recall (No-Tin)	F1-Score (Tin)	F1-Score (No-Tin)	Accuracy	ROC-AUC
LR	0.58	0.32	0.41	0.72	0.88	0.79	0.69	0.71
K-NN	0.49	0.36	0.41	0.71	0.81	0.76	0.66	0.70
SVM	0.52	0.39	0.45	0.73	0.82	0.77	0.67	0.64
ANN	0.60	0.38	0.46	0.73	0.87	0.80	0.70	0.71
RF	0.57	0.36	0.44	0.73	0.86	0.79	0.69	0.71

The bold numbers just present the highest values in each column.

ANN, artificial neural network; K-NN: K-nearest neighbors; LR, logistic regression; ML, machine learning; RF, random forest; ROC-AUC, receiver operating characteristic-area under the curve; SVM, support vector machine.

Precision

The ANN model achieved the highest precision for tinnitus (0.60), indicating a better positive predictive value, whereas K-NN had the lowest (0.49). Precision for No-T was uniformly lower, with LR performing the lowest (0.32) and SVM the best (0.39), suggesting difficulties in correctly identifying individuals without tinnitus.

Recall

Recall for tinnitus was highest in ANN (0.46), showing superior sensitivity. All models performed similarly in recalling individuals without tinnitus, with a slight edge for SVM, ANN, and RF (0.73).

F1-Score

The F1-score was highest for tinnitus in ANN (0.87) and RF (0.86). For N-T, the highest F1-score was achieved by ANN (0.80), indicating it also manages a good balance for No-T predictions.

Accuracy

Overall accuracy was highest in ANN (0.70), closely followed by LR and RF (both at 0.69). This suggests that ANN has the highest overall correctness.

Receiver Operating Characteristic-Area Under the Curve

ROC-AUC scores were highest for LR, ANN, and RF (0.71), indicating these models have a better ability to discriminate between individuals with and without tinnitus. SVM had the lowest ROC-AUC (0.64). Figure 3 exhibits the ROC plots for each model. ANN shows the largest area under the curve (AUC) which demonstrates its superior performance in distinguishing between tinnitus and No-T.

Fig. 3.

ROC-AUC for distinguishing individuals with tinnitus from those without. ROC-AUC scores were highest for LR, ANN, and RF (0.71), indicating these models have the best discrimination ability between individuals with and without tinnitus. ANN indicates artificial neural network; K-NN, K-nearest neighbors; LR, logistic regression; ML, machine learning; RF, random forest; ROC-AUC, receiver operating characteristic-area under the curve; SVM, support vector machine.

Although Synthetic Minority Over-sampling Technique and class weighting were implemented to address the class imbalance, these methods did not result in meaningful improvements in model performance. As the results remained consistent with the original analysis, we decided to report the original findings for clarity and to avoid introducing redundant information.

Differentiating NIHL From ARHL

Table 3 summarizes performance metrics for five ML models in differentiating individuals with NIHL from those with ARHL. Key metrics include precision, recall, F1-score for both NIHL and ARHL categories, overall accuracy, and ROC-AUC.

TABLE 3.

Performance metrics of five machine learning models for differentiating NIHL from ARHL

Model	Precision (NIHL)	Precision (ARHL)	Recall (NIHL)	Recall (ARHL)	F1-Score (NIHL)	F1-Score (ARHL)	Accuracy	ROC-AUC
LR	0.77	0.84	0.83	0.78	0.80	0.81	0.81	0.89
K-NN	0.75	0.78	0.74	0.78	0.75	0.78	0.77	0.83
SVM	0.79	0.82	0.79	0.82	0.79	0.82	0.81	0.85
ANN	0.78	0.82	0.81	0.80	0.79	0.81	0.80	0.89
RF	0.79	0.85	0.85	0.80	0.81	0.82	0.82	0.90

The bold numbers just present the highest values in each column.

ANN, artificial neural network; ARHL, age-related hearing loss; K-NN: K-nearest neighbors; LR, logistic regression; ML, machine learning; NIHL, noise-induced hearing loss; RF, random forest; ROC-AUC, receiver operating characteristic-area under the curve; SVM, support vector machine.

Precision

The RF model achieved the highest precision for both NIHL (0.79) and ARHL (0.85). These results suggest the RF’s highest ability to correctly identify both NIHL and ARHL.

Recall

Recall for NIHL was highest in RF (0.85), showing superior sensitivity. The recall for ARHL was highest in SVM (0.82).

F1-Score

The F1-score was highest in RF for both NIHL (0.81) and ARHL (0.82).

Accuracy

Overall accuracy was highest in RF (0.70), suggesting that RF has the highest overall correctness.

Receiver Operating Characteristic-Area Under the Curve

ROC-AUC score was highest for RF (0.90), suggesting its high performance in differentiating NIHL from ARHL. Figure 4 demonstrates the ROC plots for each model, with RF showing the largest AUC.

Fig. 4.

ROC-AUC for distinguishing NIHL from ARHL. The ROC-AUC score was highest for RF (0.90), indicating its superior performance in differentiating NIHL from ARHL. ANN indicates artificial neural network; ARHL, age-related hearing loss; K-NN, K-nearest neighbors; LR, logistic regression; ML, machine learning; NIHL, noise-induced hearing loss; RF, random forest; ROC-AUC, receiver operating characteristic-area under the curve; SVM, support vector machine.

Conclusion

As Table 3 presents, the results of the five ML models were close together, with RF demonstrating the best performance across all but one metric (ARHL recall).

DISCUSSION

In this study, we used a large dataset of demographic and audiometric information to investigate the ability of five ML models to distinguish NIHL from ARHL and differentiate individuals with tinnitus from those without. The study yielded five major findings: (1) the prevalence of tinnitus was more than double in the NIHL group compared with the ARHL group (46.40% versus 19.71%), indicating the prominent role of loud noise exposure to experiencing tinnitus; (2) a distinct pattern of significantly greater hearing loss at medium- and high-frequency ranges was observed in the NIHL group compared with the ARHL group, irrespective of tinnitus; (3) a distinct pattern of better pure-tone sensitivity in individuals with tinnitus compared with those without tinnitus, (4) among the ML models, ANN demonstrated superior performance in predicting tinnitus; and (5) the RF model showed the best performance in distinguishing NIHL from ARHL. These findings highlight the potential of ML models in enhancing the accuracy of audiological diagnoses, offering more precise and individualized treatment options. The following sections discuss these findings in detail.

High Prevalence of Tinnitus in NIHL Versus ARHL

In the present study, the prevalence of tinnitus was significantly higher in the NIHL group compared with the ARHL group (46.40% versus 19.71%). Loud noise exposure had a remarkable contribution to the occurrence of both constant (28.31% versus 8.91%) and intermittent (19.15% versus 11.13%) tinnitus, leading to a 3.5 times higher rate of constant tinnitus and around two times higher rates of intermittent tinnitus. These findings underscore the prominent role of loud noise exposure in the development of tinnitus and align with numerous studies highlighting the adverse impact of occupational noise exposure on tinnitus occurrence. For instance, a systematic review and meta-analysis reported the global prevalence and incidence of tinnitus, with point estimates of 31% among Veterans and 26% among musicians, much higher than among the general adult population (14%) (McCormack et al. 2016). Similarly, Hong et al. (2016) found that around 40% of firefighters and operating engineers experienced tinnitus, with higher rates among those with high-frequency hearing loss. Recently, Molaug et al. (2023) also reported that self-reported high exposure to noise was associated with a higher prevalence of tinnitus in those with hearing loss (prevalence ratio 1.3, 1.0 to 1.7), indicating that noise-induced tinnitus often accompanies elevated hearing thresholds.

Exposure to intense noise can cause both temporary threshold shift and permanent threshold shift. Temporary threshold shift involves reversible damage to hair cell stereocilia or synapses and can reflect a protective adaptation in the auditory system. In contrast, permanent threshold shift represents permanent damage, including the loss of hair cells and synapses. This damage often results from the accumulation of reactive oxygen species and activation of intracellular stress pathways, leading to cell death (Kurabi et al. 2017). Extensive animal studies demonstrate a cascade of neurophysiological changes owing to prolonged exposure to loud noise. These changes include the release of excessive glutamate, leading to excitotoxicity and damage to auditory nerve fibers (Liberman & Kujawa 2017; Hu et al. 2020). In addition, noise exposure can result in abnormal neural activity in the auditory pathways and central auditory cortex, contributing to the perception of tinnitus (Haider et al. 2018; Cai et al. 2019). Moreover, noise-induced damage to the cochlea can lead to decreased input to the auditory cortex, resulting in neural plasticity changes that manifest as tinnitus (Henry et al. 2014). Studies have shown that the brain attempts to compensate for the reduced auditory input by increasing neural gain, which can exacerbate tinnitus perception (Schaette & Kempter 2006; Sedley 2019). These neurophysiological alterations highlight the complex interplay between peripheral auditory damage and central auditory processing in the development of tinnitus.

Greater Hearing Loss in NIHL Versus ARHL and Tinnitus Versus No-Tinnitus

Our findings demonstrate significantly greater hearing loss at medium-band (1, 1.5, and 2 kHz) and high-frequency bands (3, 4, and 6 kHz) in the NIHL group compared with the ARHL group, and in those without tinnitus compared with those with tinnitus. Extensive evidence supports this pattern, emphasizing that chronic noise exposure exacerbates auditory aging, particularly in medium-band to high-band frequencies in standard audiometry (Natarajan et al. 2023). Extensive research has shown that early or moderately advanced NIHL is often marked by a distinct notch at 4 kHz, with adjacent frequencies at 3 and 6 kHz also affected (Rabinowitz et al. 2006). Hearing recovery is commonly observed at 8 kHz (Le et al. 2017). Typically, hearing loss from noise exposure does not exceed 75 dB in high frequencies and 40 dB in lower frequencies (Masterson et al. 2013). However, in some cases, chronic noise exposure can lead to severe to profound sensorineural hearing loss (Hong 2005; Jansen et al. 2014). Since high frequencies are also impacted by presbyacusis, the 4 kHz notch may diminish with age, complicating the differentiation between NIHL and ARHL (Le et al. 2017). Prolonged noise exposure can deepen and widen the notch, potentially affecting lower frequencies such as 2, 1, and 0.5 kHz (Hong et al. 2013). The potential for chronic noise exposure to cause hearing loss at 8 kHz remains debated (Le et al. 2017).

In addition to the influence of the resonance frequency of the outer ear and the mechanical properties of the middle ear on the 4 kHz notch (Pierson et al. 1994), the increased susceptibility of medium and high frequencies to noise damage can be attributed to physiological factors. High-frequency hair cells, located at the cochlear base, are particularly vulnerable to mechanical stress and oxidative damage caused by noise exposure (Hickman et al. 2021; Kishimoto-Urata et al. 2022). This damage is often accompanied by synaptic loss at the cochlear level, disrupting communication between hair cells and auditory nerve fibers, further contributing to hearing loss (Kujawa & Liberman 2015; Hickman et al. 2021). In addition, chronic noise exposure triggers inflammatory responses that exacerbate age-related auditory changes, leading to increased hearing loss in both medium and high frequencies (Kujawa & Liberman 2006; Le et al. 2017).

In our study, we observed that the severity of hearing loss was greater in individuals without tinnitus compared with those with tinnitus, across both ARHL and NIHL groups. One possible explanation is that individuals without tinnitus might delay seeking treatment for their hearing loss, allowing it to progress to a more severe state before intervention. Moreover, tinnitus can cause significant distress, prompting individuals to seek help earlier and more frequently, which might help mitigate further hearing damage. Further research is necessary to replicate our findings and provide deeper insight into the complex interplay between tinnitus and hearing loss severity.

ML Models in Predicting Tinnitus

Among the ML models evaluated in this study, the ANN model showed the best performance in distinguishing individuals with tinnitus from those without. This was evidenced by the model’s precision (0.60), recall (0.46), F1-score (0.87), accuracy (0.70), and ROC-AUC (0.71). In comparison, RF and LR also showed strong performance, whereas K-NN and SVM were less effective. Overall accuracy was highest in ANN at 0.70. The ROC-AUC scores, which measure the ability of the model to distinguish between classes, were highest for LR, ANN, and RF, all scoring 0.71. These findings indicate the ANN model’s superior performance across most metrics. The high recall for the No-T group across all models (between 0.71 and 0.73) also demonstrates their capability to identify true negatives, which may reduce the risk of misdiagnosis.

ANNs are computational models inspired by the human brain’s network of neurons. ANNs are particularly adept at handling complex patterns and nonlinear relationships within data (Agatonovic-Kustrin & Beresford 2000). The adaptability of ANNs allows for continuous learning and improvement as more data becomes available, enhancing their predictive accuracy over time. These characteristics make them well-suited for tasks like predicting tinnitus. The superior performance of the ANN model in our study aligns with a broader body of research demonstrating the efficacy of ML models in health and disease prediction (Ghaffar Nia et al. 2023), including auditory disorders. For instance, Tomiazzi et al. (2019) evaluated ML algorithms for classifying hearing impairment among Brazilian farmers exposed to pesticides and cigarette smoke. They found that ANN models were effective in recognizing patterns of hearing impairment, similar to our study’s success in identifying tinnitus. In a study by Manta et al. (2023), seven ML models were evaluated, including RF, linear, radial, and polynomial SVM, naive Bayes, ANN, and linear discriminant analysis. The study used both electrophysiologic data—specifically wavelet-transformed auditory evoked potential signals, such as auditory brainstem responses and auditory middle latency responses—and clinical data to classify tinnitus patients based on their level of distress. Among the models, the SVM classifier showed the best performance, with an AUC value of 92.53%, sensitivity of 84.84%, and specificity of 83.04%. The authors concluded that auditory middle latency response signals offer valuable insights into tinnitus distress.

The robustness of ANNs in our study is also supported by their successful application in other areas of healthcare. For example, ANNs have been used to predict cardiovascular diseases, diabetes (Abbas et al. 2024), and even cancer outcomes with high accuracy (Esteva et al. 2017), as reported in various studies (Ghaffar Nia et al. 2023). These applications highlight the versatility and potential of ANNs in improving diagnostic processes across different medical fields.

Models in Distinguishing NIHL From ARHL

Among the models assessed, the RF model showed better performance across multiple metrics in differentiating NIHL from ARHL. The model demonstrated the highest precision for both NIHL (0.79) and ARHL (0.85), the highest recall for NIHL (0.85), the highest F1-score for both NIHL (0.81) and ARHL (0.82), the highest overall accuracy (0.70), and highest ROC-AUC score (0.90). The RF model’s highest precision for NIHL and ARHL indicates its superior ability to differentiate these conditions correctly. Precision measures the proportion of true positives among the predicted positives, reflecting the model’s accuracy in identifying individuals with NIHL and ARHL. The high recall for NIHL in RF and ARHL in SVM demonstrates the models’ sensitivity, indicating their effectiveness in capturing true positives. The F1-score was highest for RF in both NIHL and ARHL, showing the model’s robust performance in differentiating between these two types of hearing loss. Overall accuracy was highest in RF, suggesting its highest overall correctness. The ROC-AUC score, which indicates the model’s ability to distinguish between classes, was also highest for RF (0.90).

The adaptability and versatility of RF models make them well-suited for differentiating between NIHL and ARHL. RF models are ensemble learning methods that construct multiple decision trees and merge their outputs to improve predictive accuracy and control overfitting. This capability to handle large datasets with higher dimensionality and their inherent feature importance estimation makes RF particularly valuable in medical diagnostics (Khan et al. 2024). Furthermore, RF models have been successfully applied in various health domains (Cruz & Wishart 2007; Boukhatem et al. 2022), highlighting their potential to improve diagnostic accuracy and patient outcomes.

The superior performance of the RF model in our study aligns with existing literature that emphasizes the efficacy of ML models in predicting hearing loss. For instance, Zhao et al. (2019a) demonstrated the effectiveness of RF and three additional models (SVM, adaptive boosting, and multilayer perceptron) in predicting hearing impairment among workers exposed to complex industrial noise. In a recent study from the same research group, a new prediction model using asymmetric convolution algorithms was developed to identify workers exposed to complex industrial noise (Tian et al. 2024). By incorporating time-frequency features, this model outperformed the ISO 1999 and SVM models in predicting NIHL, demonstrating an AUC of 0.7768 compared with 0.75 for SVM and 0.51 for ISO 1999. Similarly, Ellis and Souza (2021) explored the use of RF algorithms in classifying audiometric data on a large dataset, using variables such as hearing thresholds, gender, age, military experience, and self-reported hearing ability (Ellis & Souza 2021). The RF model achieved an accurate classification for 54.79% of individuals, with the majority of misclassifications predicting milder losses than measured. Their findings corroborate our results, highlighting ML models’ capability to handle complex data and capture subtle differences between conditions.

Study Strengths, Limitations, and Directions for Future Research

The motivation for applying ML models in this study stems from their potential to address limitations inherent in traditional diagnostic approaches. While current manual classification methods rely heavily on self-reports and basic audiometric data, these methods may lack the granularity needed to uncover subtle patterns across diverse populations or predict outcomes under varying conditions. The integration of ML enables the synthesis of complex, multidimensional datasets—such as demographic variables, audiometric thresholds, and frequency-specific hearing loss patterns—into predictive frameworks that can enhance diagnostic precision and offer deeper insights into the mechanisms underlying tinnitus and NIHL. The need for ML models becomes particularly evident in scenarios where large-scale data analysis is required, such as identifying risk factors across populations or tailoring interventions for individuals. For tinnitus, which is multifaceted and influenced by a range of physiological, psychological, and environmental factors, ML models allow for the development of predictive tools that can integrate and weigh these diverse inputs. This approach has the potential to move beyond binary classification and toward nuanced, personalized assessments that could support targeted therapeutic interventions. Furthermore, the use of ML models creates opportunities for scalability and reproducibility. For instance, these models can be trained and validated across datasets to generalize findings beyond the immediate study population, a process that manual methods cannot easily replicate. While concerns about generalizability remain valid, ongoing refinement of these models through larger, more diverse datasets and longitudinal data will improve their robustness and applicability. By leveraging ML, the study aimed to contribute to a future where clinical audiology benefits from data-driven, precision-oriented decision-making frameworks.

Our study presents several notable strengths in investigating ML models for differentiating NIHL and ARHL and predicting tinnitus. First, the application of diverse ML models (LR, K-NN, SVM, ANN, and RF) offers opportunities for the automated evaluation of various predictive techniques. In addition, our study benefits from a well-defined dataset and methodical evaluation metrics. We provided a detailed analysis of several metrics, including precision, recall, F1-score, accuracy, and ROC-AUC, to compare the performance of the ML methods. The high ROC-AUC scores for the RF and ANN models potentially reflect their discriminatory power, which is critical for effective diagnostic differentiation.

One potential limitation of our study is the risk of model overfitting, particularly with complex algorithms like ANN, which may perform exceptionally well on training data but less so on unseen data. Although cross-validation techniques were employed to mitigate this risk, further validation on independent datasets is necessary to confirm the generalizability of our findings. Another potential limitation involves the specificity of the dataset used. Although the dataset was comprehensive, the characteristics of the sample—such as the auditory data available and the demographic composition—might not fully represent all possible variations in the population. Expanding the dataset to include more diverse populations and a wider range of audiometric and demographic variables could improve model robustness and generalizability. In addition, incorporating longitudinal data could provide insights into how well the models predict changes in hearing data over time, enhancing their utility for tracking the progression of NIHL, ARHL, and tinnitus.

Using fivefold cross-validation with the 80-20 train-test split ensured more robust and representative performance metrics by reducing bias from a single data split. Consistent preprocessing across all models facilitated comparability. However, slight overfitting and imbalanced performance for the minority class (N-T) persisted, highlighting the need for further refinement and additional data for this group in future studies. Although oversampling techniques and class weighting were applied, they did not result in significant improvements. Collecting more data from the minority class (N-T) could enhance its representation and improve the generalizability of the models.

Further exploration of hybrid models that combine multiple ML techniques on large datasets could also yield improved predictive performance (Chen et al. 2021). Investigating the use of ensemble methods, such as combining RF with ANN or other models, may enhance the ability to capture complex patterns and interactions within the data. In addition, real-world application and validation of these models in clinical settings are crucial. Pilot studies evaluating the effectiveness of ML-based diagnostic tools in clinical assessments and treatment planning will provide valuable insights into their feasibility and impact on patient outcomes.

CONCLUSIONS

We used a comprehensive dataset encompassing audiometric and demographic information to evaluate the performance of five ML models in predicting tinnitus and differentiating NIHL from ARHL. In this study, NIHL was associated with substantially greater hearing loss in the medium and high frequencies compared with ARHL. It is interesting that individuals without tinnitus exhibited a greater degree of hearing loss compared with those with tinnitus. Among the ML models, ANN demonstrated the highest performance in predicting tinnitus, and RF outperformed the other models in distinguishing NIHL from ARHL. These results underscore the potential of ML models to enhance diagnostic accuracy in hearing disorders. Given the limitations of current research on ML prediction models—such as variability in target populations, sample sizes, research objectives, types of data, and the ML models used—further studies using larger and more diverse datasets are essential. Such research will be critical for validating these findings and drawing comprehensive conclusions about the applicability of ML approaches to clinical practice. Future investigations should focus on expanding datasets to include varied populations, incorporating longitudinal data to track changes in hearing over time, and exploring hybrid ML approaches to refine predictive accuracy.

ACKNOWLEDGMENTS

Z.J. conceptualized, designed, and managed the study,. Z.J. and R.E.H. performed the data analyses and wrote the first draft of the article. All authors (Z.J., R.E.H., G.H., B.E.K., and M.H.M.) discussed the results and implications and reviewed and commented on the final revision.