[论文解读] IATos: AI-powered pre-screening tool for COVID-19 from cough audio samples

D. Trejo Pizzo, S. Esteban|arXiv (Cornell University)|Apr 27, 2021

COVID-19 diagnosis using AIMedicine被引用 18

一句话总结

IATos 是一款基于人工智能的预筛查工具，利用深度学习对咳嗽音频样本进行分类，以检测新冠肺炎，使用 RT-PCR 作为金标准，准确率达到 86%，能够有效区分阳性与阴性病例。该系统利用布宜诺斯艾利斯市的大规模音频数据，实现可扩展、低成本且非侵入性的群体筛查，作为临床检测的补充。

ABSTRACT

OBJECTIVE: Our objective is to evaluate the possibility of using cough audio recordings (spontaneous or simulated) to detect sound patterns in people who are diagnosed with COVID-19. The research question that led our work was: what is the sensitivity and specificity of a machine learning based COVID-19 cough classifier, using RT-PCR tests as gold standard? SETTING: The audio samples that were collected for this study belong to individuals who were swabbed in the City of Buenos Aires in 20 public and 1 private facilities where RT-PCR studies were carried out on patients suspected of COVID, and 14 out-of-hospital isolation units for patients with confirmed COVID mild cases. The audios were collected through the Buenos Aires city government WhatsApp chatbot that was specifically designed to address citizen inquiries related to the coronavirus pandemic (COVID-19). PARTICIPANTS: The data collected corresponds to 2821 individuals who were swabbed in the City of Buenos Aires, between August 11 and December 2, 2020. Individuals were divided into 1409 that tested positive for COVID-19 and 1412 that tested negative. From this sample group, 52.6% of the individuals were female and 47.4% were male. 2.5% were between the age of 0 and 20 , 61.1% between the age of 21 and 40 , 30.3% between the age of 41 and 60 and 6.1% were over 61 years of age. RESULTS: Using the dataset of 2821 individuals our results showed that the neural network classifier was able to discriminate between the COVID-19 positive and the healthy coughs with an accuracy of 86%. This accuracy obtained during the training process was later tested and confirmed with a second dataset corresponding to 492 individuals.

研究动机与目标

• 评估基于机器学习的咳嗽分类器在使用咳嗽音频记录检测新冠肺炎时的敏感性和特异性。
• 开发一种可扩展、低成本的数字预筛查工具，以减轻医疗系统压力并降低暴露风险。
• 通过布宜诺斯艾利斯市公共检测机构的现实世界 RT-PCR 确诊病例验证人工智能模型的性能。
• 探索在城市环境中大规模部署此类工具进行群体筛查的可行性。
• 为支持全球疫情应对技术的创新，贡献开源方法和数据。

提出的方法

• 本研究采用深度神经网络分类器，基于在疫情期间通过全市范围 WhatsApp 机器人收集的咳嗽音频样本进行训练。
• 根据 RT-PCR 检测结果将音频记录分类为阳性或阴性，作为模型训练和评估的基准真实值。
• 模型首先在包含 2,821 名个体（1,409 名阳性，1,412 名阴性）的数据集上进行训练，随后在更大规模的 143,351 名个体数据集（18,271 名阳性，125,080 名阴性）上重新训练，以提升鲁棒性。
• 系统使用音频信号处理技术，从自发性或强制性咳嗽中提取特征，重点关注与感染相关的呼吸音模式。
• 通过保留的测试集（492 名个体）验证模型性能，确保其在初始训练数据之外仍具泛化能力。
• 最终模型针对检测阴性病例的高特异性进行了优化，尽管阴性样本在数据中占主导地位（90%）。

实验结果

研究问题

• RQ1 基于机器学习的咳嗽分类器在使用 RT-PCR 作为金标准时，检测新冠肺炎的敏感性和特异性是多少？
• RQ2 来自多样化城市人群的咳嗽音频记录能否可靠地区分 SARS-CoV-2 感染者与非感染者？
• RQ3 当在疫情爆发期间收集的大规模真实世界数据集上进行训练时，模型性能如何扩展？
• RQ4 此类工具在多大程度上可通过预筛查减轻传统检测基础设施的负担？
• RQ5 该系统能否作为可扩展、低成本且安全的数字分诊工具，在公共卫生环境中有效部署？

主要发现

• 在初始 2,821 名个体的数据集上，神经网络分类器在区分新冠肺炎阳性与阴性咳嗽方面达到了 88% 的准确率。
• 在独立验证集（492 名个体）上测试时，模型保持了强劲性能，证实了其可靠性。
• 在重新训练包含 143,351 名个体的更大数据集后，模型准确率达到 86%，展现出可扩展性和鲁棒性。
• 模型在识别阴性病例方面表现更优，可能得益于第二组数据集中阴性样本占 90% 的高比例。
• 该系统展现出大规模部署作为预筛查工具的潜力，可在极少基础设施投入下，实现每天处理数千份咳嗽样本。
• 本研究证实，咳嗽音频分析可作为 RT-PCR 检测在疫情管理中的可行、非侵入性且成本效益高的补充手段。