机器学习用于声门癌与喉活动障碍的无创鉴别

文献信息

中文摘要

目的： 随着人工智能和机器学习的发展，用于医学诊断的嗓音分析正在迅速进步。无创嗓音分析可优化耳鼻喉科（ENT）的诊断流程，补充内镜评估并指导治疗策略。本研究旨在使用机器学习在发声困难患者中区分早期声门癌（EGC）与单侧喉活动障碍（ULI）。
研究设计： 在Python中开发了一个机器学习模型，整合了数据分割、归一化、超参数调优，以及通过参数变换进行的数据增强，并使用合成少数类过采样技术（SMOTE）来平衡类别分布。一个结合多种算法（随机森林分类器、支持向量机和极端梯度提升）的元模型提高了预测准确性。
地点： 三级医疗中心（耳鼻喉科喉科单元）。
方法： 人群为诊断为ULI或EGC的发声困难患者，以及无症状对照。排除了患有其他影响嗓音疾病的患者。63例病理病例（ULI：38例，EGC：25例）和40例对照。数据增强和再平衡将数据集扩展至650个样本。使用机器学习分析持续元音的嗓音录音，以获取嗓音参数（基频、抖动和闪烁）。使用了高质量音频设备。通过精确率、召回率和受试者工作特征曲线下面积（AUC）评估模型在区分EGC与ULI方面的性能。
结果： 该模型在区分ULI与EGC方面达到了72%的准确率，并通过测试得到验证。
结论： 机器学习在区分EGC与ULI方面显示出强大潜力。这种无创方法可以提高诊断准确性并实现更早干预，尤其是在专科资源有限的地区。进一步改进可增强其在公共卫生中的作用。

英文摘要

OBJECTIVE: Voice analysis for medical diagnostics is advancing rapidly with artificial intelligence and machine learning. Non-invasive voice analysis could optimize diagnostic procedures in ear, nose, and throat (ENT), supplementing endoscopic evaluations and guiding treatment strategies. This study aims to distinguish early glottic carcinomas (EGCs) from unilateral laryngeal immobility (ULI) in dysphonic patients using machine learning.
STUDY DESIGN: A machine learning model was developed in Python, incorporating data segmentation, normalization, hyperparameter tuning, and data augmentation via parameter transformation and using Synthetic Minority Over-sampling Technique (SMOTE) to balance class distributions. A metamodel combining multiple algorithms (Random Forest Classifier, Support Vector Machines, and Extreme Gradient Boosting) enhanced predictive accuracy.
SETTING: Tertiary Medical Center (laryngology unit, ENT Department).
METHODS: Population is dysphonic patients diagnosed with ULI or EGC, and asymptomatic controls. Patients with other voice-affecting conditions were excluded. Sixty-three pathologic cases (ULI: 38, EGC: 25) and 40 controls. Data augmentation and rebalance expanded the data set to 650 samples. Voice recordings of sustained vowels were analyzed for vocal parameters (fundamental frequency, jitter, and shimmer) using machine learning. High-quality audio equipment was used. Model performance was evaluated by precision, recall, and area under receiver operating characteristic curve (AUC) in differentiating EGC from ULI.
RESULTS: The model achieved 72% accuracy in distinguishing ULI from EGC, validated through tests.
CONCLUSION: Machine learning shows strong potential for distinguishing EGC from ULI. This non-invasive method could improve diagnostic accuracy and enable earlier intervention, especially in regions with limited specialist access. Further refinement could enhance its role in public health.