JMIR Publications

ABSTRACT

Background:

In Japan, individuals with a mild COVID-19 illness were supposed to be monitored in designated areas and were hospitalized only if their condition worsened to moderate illness I 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 mild-illness patients was required. As voice biomarkers for Parkinson's disease or Alzheimer's disease are known to be significant for classifying or monitoring symptoms, they may be adaptable for classifying the severity of COVID-19 because it is a respiratory disease, and it has been reported that acoustic changes in the voice occur with this disease. In this study, voices are considered via their wave-like features using two-dimensional parameters, and we attempt to classify them in terms of their similarity between the waveforms of voices of mild-illness and moderate-illness I patients using a dynamic time warping (DTW) algorithm.

Objective:

This feasibility study aimed to test whether DTW-based voice biomarkers can be used to obtain a binary classification of mild illness and moderate illness I for COVID-19 at a significant level.

Methods:

Participants who agreed to participate in this study (n = 295) were infected by COVID-19 between July 2021 and June 2022 and were aged 20 years or older under recuperation in the Kanagawa Prefecture. In total, 110 participants were finally selected according to their age, sex, and time of occurrence of the variant and severity of their conditions as mild illness (61) or moderate illness I (49). Three kinds of long vowels were obtained from the patients, and as a preprocessing step, each voice was cut into a 10-cycle waveform and then standardized to the power (amplitude) and time axes. The DTW distances obtained for a given waveform (x) in combination with the remaining 109 waveforms (110-x) were divided into two groups according to the severity labels of the 109 waveforms. The average and variance of the two groups were then calculated as four indices associated with a single waveform of x. Among the 12 indices (four indices per vowel multiplied by three vowels), the significant indices for classification were determined using the Mann-Whitney U-test, linear discrimination analysis (LDA), and confusion matrix. Those indices were used as predictors for a generalized logistic model (GLM) as a classification model using a five-fold cross-validation method. Predictive validation of the model was confirmed using the confusion matrix and the receiver operating characteristic curve /the area under the curve (ROC/AUC).

Results:

The proposed DTW-distance-based GLM achieved a high balance accuracy of the confusion matrix for three vowels, ranging from 80.2% to 88.0%, and AUC ranged from 86.5% to 96.5%.

Conclusions:

The proposed model may be a useful tool for monitoring the progress of COVID-19 patients in care.