Identifying patients with Long COVID: study of the effectiveness of voice signal analysis and machine learning (Preprint)

Abstract

The COVID-19 pandemic has had a profound global health impact, leading to the emergence of Post-Acute Sequelae of SARS-CoV-2 (PASC) infection, characterized by persistent symptoms in a significant proportion of recovered individuals. However, distinguishing between patients with PASC and healthy individuals remains challenging due to the wide range of symptoms associated with the condition. The existing scientific literature on PASC diagnosis has primarily focused on clinical manifestations and symptomatology, with limited attention given to voice signal analysis. To address this gap, our study utilizes a pioneering dataset collected from individuals diagnosed with PASC, making it the first of its kind to analyze voice signals in this specific population.

The primary objective of this study is to compare multiple algorithms and develop a classification model based-on Machine Learning (ML) that can effectively distinguish individuals with PASC. Additionally, this study aims to identify discernible differences in voice signals among patients with PASC as second objective. Furthermore, this study seeks to investigate the influence of physical effort and different audios (coughs or sustained phonations of vowel /a/) on the performance of the classification model.

This study presents a PASC recognition system that utilizes a classical ML classification strategy. The system includes an audio pre-processing phase, a feature extraction module, a classifier and a voting system to assign class labels. The pre-processing phase enhances voice signals quality, while feature extraction captures underlying characteristics such as spectral features or energy-related measures. The classifiers trained (Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Multi-Layer Perceptron (MLP)) on labeled data to predict the class of new voice signals, and a voting system determines the final classification result. The study provides insights into the performance of the methods employed and demonstrates the potential of the recognition system for distinguishing patients with PASC.

In terms of general performance, RF classifier achieved the highest performance (AUC: 93%, F1-score: 86%), followed by the MLP classifier (AUC: 89%, F1-score: 85%). Spectral features outperformed other feature sets. Coughs audios improved the classifiers performance compared to the sustained phonation of vowel /a/. Regarding the impact of physical exertion, there was an increase in performance metrics for all classifiers after physical exertion.

Our research aimed to automatically recognize PASC using voice signals. Our study revealed that spectral features were highly effective in accurately classifying patients with PASC. Among the classifiers tested, the RF classifiers demonstrated superior performance. Moreover, we observed that utilizing cough sounds and audio recordings taken after physical exertion improved the classifiers' performance. These significant findings contribute to the advancement of non-invasive diagnostic tools for PASC.

US Clinical Trials Registry NCT05629793