Engineering Research Center of EMR and Intelligent Expert System, Ministry of Education, Key Laboratory for Biomedical Engineering of Ministry of Education, College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, Zhejiang Province, China.
These authors contributed equally to this study.
Ear Hear. 2024;45(3):648-657. doi: 10.1097/AUD.0000000000001454. Epub 2024 Jan 10.
Current approaches for evaluating noise-induced hearing loss (NIHL), such as the International Standards Organization 1999 (ISO) 1999 prediction model, rely mainly on noise energy and exposure time, thus ignoring the intricate time-frequency characteristics of noise, which also play an important role in NIHL evaluation. In this study, an innovative NIHL prediction model based on temporal and spectral feature extraction using an asymmetric convolution algorithm is proposed.
Personal data and individual occupational noise records from 2214 workers across 23 factories in Zhejiang Province, China, were used in this study. In addition to traditional metrics like noise energy and exposure duration, the importance of time-frequency features in NIHL assessment was also emphasized. To capture these features, operations such as random sampling, windowing, short-time Fourier transform, and splicing were performed to create time-frequency spectrograms from noise recordings. Two asymmetric convolution kernels then were used to extract these critical features. These features, combined with personal information (e.g., age, length of service) in various configurations, were used as model inputs. The optimal network structure was selected based on the area under the curve (AUC) from 10-fold cross-validation, alongside the Wilcoxon signed ranks test. The proposed model was compared with the support vector machine (SVM) and ISO 1999 models, and the superiority of the new approach was verified by ablation experiments.
The proposed model had an AUC of 0.7768 ± 0.0223 (mean ± SD), outperforming both the SVM model (AUC: 0.7504 ± 0.0273) and the ISO 1999 model (AUC: 0.5094 ± 0.0071). Wilcoxon signed ranks tests confirmed the significant improvement of the proposed model ( p = 0.0025 compared with ISO 1999, and p = 0.00142 compared with SVM).
This study introduced a new NIHL prediction method that provides deeper insights into industrial noise exposure data. The results demonstrated the superior performance of the new model over ISO 1999 and SVM models. By combining time-frequency features and personal information, the proposed approach bridged the gap between conventional noise assessment and machine learning-based methods, effectively improving the ability to protect workers' hearing.
当前评估噪声性听力损失（NIHL）的方法，如国际标准化组织 1999 年（ISO）1999 预测模型，主要依赖于噪声能量和暴露时间，因此忽略了噪声的复杂时频特征，而这些特征在 NIHL 评估中也起着重要作用。本研究提出了一种基于时间和频谱特征提取的创新 NIHL 预测模型，使用非对称卷积算法。
本研究使用了来自中国浙江省 23 家工厂的 2214 名工人的个人数据和个人职业噪声记录。除了传统的噪声能量和暴露时间等指标外，还强调了时频特征在 NIHL 评估中的重要性。为了捕捉这些特征，对噪声记录进行了随机采样、加窗、短时傅里叶变换和拼接等操作，以创建时频谱图。然后使用两个非对称卷积核提取这些关键特征。将这些特征与个人信息（如年龄、工龄）以各种配置组合在一起作为模型输入。根据 10 折交叉验证的曲线下面积（AUC）和 Wilcoxon 符号秩检验选择最佳网络结构。将所提出的模型与支持向量机（SVM）和 ISO 1999 模型进行了比较，并通过消融实验验证了新方法的优越性。
所提出的模型 AUC 为 0.7768 ± 0.0223（均值 ± 标准差），优于 SVM 模型（AUC：0.7504 ± 0.0273）和 ISO 1999 模型（AUC：0.5094 ± 0.0071）。Wilcoxon 符号秩检验证实了所提出模型的显著改进（与 ISO 1999 相比，p = 0.0025，与 SVM 相比，p = 0.00142）。
本研究介绍了一种新的 NIHL 预测方法，为工业噪声暴露数据提供了更深入的见解。结果表明，新模型在 ISO 1999 和 SVM 模型上的性能优于传统模型。通过结合时频特征和个人信息，该方法弥补了传统噪声评估和基于机器学习方法之间的差距，有效提高了保护工人听力的能力。
