Watase Teruhisa, Omiya Yasuhiro, Tokuno Shinichi
Gradutate School of Health Innovation Kanagawa University of Human Service Kawasaki, Kanagawa Japan.
Department of Bioengineering Graduate School of Engineering The University of Tokyo Tokyo Japan.
JMIR Biomed Eng. 2023 Nov 6;8:e50924. doi: 10.2196/50924. eCollection 2023.
In Japan, individuals with mild COVID-19 illness previously required to be monitored in designated areas and were hospitalized only if their condition worsened to moderate illness or worse. Daily monitoring using a pulse oximeter was a crucial indicator for hospitalization. However, a drastic increase in the number of patients resulted in a shortage of pulse oximeters for monitoring. Therefore, an alternative and cost-effective method for monitoring patients with mild illness was required. Previous studies have shown that voice biomarkers for Parkinson disease or Alzheimer disease are useful for classifying or monitoring symptoms; thus, we tried to adapt voice biomarkers for classifying the severity of COVID-19 using a dynamic time warping (DTW) algorithm where voice wavelets can be treated as 2D features; the differences between wavelet features are calculated as scores.
This feasibility study aimed to test whether DTW-based indices can generate voice biomarkers for a binary classification model using COVID-19 patients' voices to distinguish moderate illness from mild illness at a significant level.
We conducted a cross-sectional study using voice samples of COVID-19 patients. Three kinds of long vowels were processed into 10-cycle waveforms with standardized power and time axes. The DTW-based indices were generated by all pairs of waveforms and tested with the Mann-Whitney test (α<.01) and verified with a linear discrimination analysis and confusion matrix to determine which indices were better for binary classification of disease severity. A binary classification model was generated based on a generalized linear model (GLM) using the most promising indices as predictors. The receiver operating characteristic curve/area under the curve (ROC/AUC) validated the model performance, and the confusion matrix calculated the model accuracy.
Participants in this study (n=295) were infected with COVID-19 between June 2021 and March 2022, were aged 20 years or older, and recuperated in Kanagawa prefecture. Voice samples (n=110) were selected from the participants' attribution matrix based on age group, sex, time of infection, and whether they had mild illness (n=61) or moderate illness (n=49). The DTW-based variance indices were found to be significant (<.001, except for 1 of 6 indices), with a balanced accuracy in the range between 79% and 88.6% for the /a/, /e/, and /u/ vowel sounds. The GLM achieved a high balance accuracy of 86.3% (for /a/), 80.2% (for /e/), and 88% (for /u/) and ROC/AUC of 94.8% (95% CI 90.6%-94.8%) for /a/, 86.5% (95% CI 79.8%-86.5%) for /e/, and 95.6% (95% CI 92.1%-95.6%) for /u/.
The proposed model can be a voice biomarker for an alternative and cost-effective method of monitoring the progress of COVID-19 patients in care.
在日本，轻度新冠肺炎患者此前需要在指定区域接受监测，只有病情恶化为中度或更严重时才会住院。使用脉搏血氧仪进行每日监测是住院的关键指标。然而，患者数量的急剧增加导致用于监测的脉搏血氧仪短缺。因此，需要一种替代的、具有成本效益的方法来监测轻症患者。先前的研究表明，帕金森病或阿尔茨海默病的语音生物标志物有助于对症状进行分类或监测；因此，我们尝试采用动态时间规整（DTW）算法，将语音小波视为二维特征，利用语音生物标志物对新冠肺炎的严重程度进行分类；小波特征之间的差异计算为分数。
本可行性研究旨在测试基于DTW的指标能否为二元分类模型生成语音生物标志物，该模型使用新冠肺炎患者的声音在显著水平上区分中度疾病和轻度疾病。
我们使用新冠肺炎患者的语音样本进行了一项横断面研究。将三种长元音处理为具有标准化功率和时间轴的10周期波形。基于DTW的指标由所有波形对生成，并通过曼-惠特尼检验（α<0.01）进行测试，并用线性判别分析和混淆矩阵进行验证，以确定哪些指标在疾病严重程度的二元分类中表现更好。基于广义线性模型（GLM），使用最有前景的指标作为预测因子，生成二元分类模型。受试者工作特征曲线/曲线下面积（ROC/AUC）验证模型性能，混淆矩阵计算模型准确性。
本研究的参与者（n=295）于2021年6月至2022年3月感染新冠肺炎，年龄在20岁及以上，在神奈川县康复。根据年龄组、性别、感染时间以及是否患有轻症（n=61）或中度疾病（n=49），从参与者的属性矩阵中选择语音样本（n=110）。发现基于DTW的方差指标具有显著性（<0.001，6个指标中有1个除外），对于/a/、/e/和/u/元音，平衡准确率在79%至88.6%之间。GLM实现了较高的平衡准确率，/a/为86.3%，/e/为80.2%，/u/为88%；/a/的ROC/AUC为94.8%（95%CI 90.6%-94.8%），/e/为86.5%（95%CI 79.8%-86.5%），/u/为95.6%（95%CI 92.1%-95.6%）。
所提出的模型可以作为一种语音生物标志物，用于一种替代的、具有成本效益的方法来监测新冠肺炎患者在护理中的病情进展。
