比较机器学习算法在分类职业性噪声性听力损失中的性能
Comparing the performance of machine learning algorithms in classifying occupational noise-induced hearing loss.
文献信息
| PMID | 42858046 |
|---|---|
| 原文 | 在 PubMed 查看原文 ↗ |
| 发表日期 | 2026 |
| 作者 | Vida Rezaei-Hachesu |
| 作者单位 | Department of Occupational Health, School of Public Health, Tehran University of Medical Sciences, Tehran, Iran. |
| 期刊 | Work (Reading, Mass.) |
| SCI 分区 | Q3 |
| IF | 1.8 |
| 研究类型 | AI/ML · 临床 |
| 所属专科 | 耳科 |
中文摘要
背景： 纯音测听是诊断职业性噪声性听力损失（ONIHL）的金标准；然而，其依赖昂贵的设备、训练有素的人员和主观听觉反应，限制了其在大规模职业筛查中的应用。
目的： 本研究旨在开发和比较仅使用听力测量数据以及结合人口统计学和临床特征的听力测量数据的机器学习（ML）算法，用于ONIHL分类，并评估这些变量是否在听力阈值之外提供额外的预测价值。
方法： 在这项针对1524名伊朗石化行业工人的横断面研究中，开发了四种ML算法：决策树（DT）、随机森林（RF）、支持向量机（SVM）和极端梯度提升（XGBoost），用于分类ONIHL。数据按80:20分为训练集和测试集。使用准确率、精确率、灵敏度、F1分数和kappa统计量评估模型性能。
结果： 基于树的模型在不同ONIHL水平上均优于SVM。添加人口统计学-临床变量使准确率、灵敏度和F1分数提高了2-5%。所有模型对正常听力均达到100%的精确率。对于中度至重度损失，XGBoost达到85%的精确率和89%的灵敏度，而RF分别达到81%和87%。SVM在轻度损失方面表现较差，可能是由于对类别不平衡的敏感性。6000 Hz阈值是模型中最重要的特征。
结论： ML算法，特别是基于树的模型，能有效识别ONIHL。结合听力测量和人口统计学-临床数据可提高性能，并可能支持在高风险职业人群中进行更准确的筛查和及时干预。
英文摘要
BackgroundPure-tone audiometry is the gold standard for diagnosing occupational noise-induced hearing loss (ONIHL); however, its reliance on expensive equipment, trained personnel, and subjective auditory responses limits its use in large-scale occupational screening.ObjectiveThis study aimed to develop and compare machine learning (ML) algorithms for ONIHL classification using audiometric data alone and audiometric data combined with demographic and clinical features, and to assess whether these variables provide additional predictive value beyond audiometric thresholds alone.MethodsIn this cross-sectional study of 1524 workers employed in the Iranian petrochemical industry, four ML algorithms, decision tree (DT), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost), were developed to classify ONIHL. Data were split 80:20 into training and testing sets. Model performance was evaluated using accuracy, precision, sensitivity, F1-score, and the kappa statistic.ResultsTree-based models outperformed SVM across ONIHL levels. Adding demographic-clinical variables improved accuracy, sensitivity, and F1-score by 2-5%. All models achieved 100% precision for normal hearing. For moderate-to-severe loss, XGBoost achieved 85% precision and 89% sensitivity, whereas RF achieved 81% and 87%, respectively. SVM performed less well for mild loss, likely because of sensitivity to class imbalance. The 6000 Hz threshold was the most important feature across models.ConclusionsML algorithms, particularly tree-based models, effectively identified ONIHL. Combining audiometric and demographic-clinical data enhanced performance and may support more accurate screening and timely intervention in high-risk occupational populations.