A novel framework for Covid-19 detection and clinical triage using multimodal physiological signals on a portable medical assistant
Nenhuma Miniatura disponível
Data
2025-08-15
Autores
Silva, Leticia Araújo
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal do Espírito Santo
Resumo
Emergency and urgent care systems face growing challenges in providing timely and accurate triage, especially in resource-constrained environments where subjectivity, lack of infrastructure, and high patient volumes compromise clinical decisions. These limitations became even more evident during the Coronavirus Disease 2019 (COVID-19) pandemic, which exposed critical gaps in diagnostic capacity and highlighted the absence of scalable, non-invasive tools for autonomous assessment. This research investigates whether multimodal physiological signals — cough, speech, breath, and vital signs — collected through a portable equipment called Integrated Portable Medical Assistant (IPMA) may support intelligent triage and COVID-19 inference via Machine Learning (ML) models. To address this, a two-part experimental design was conducted. The first part focused on COVID-19 detection using public datasets and real-world data collected with the IPMA. Mel-spectrograms were extracted from audio signals, followed by texture-based feature extraction using Local Binary Pattern (LBP) and Local Ternary Pattern (LTP). LBP consistently outperformed LTP across classification tasks, with speech showing the highest discriminative power, and SpO2 and temperature emerging as the most informative physiological indicators. Although trained on public datasets, models achieved moderate generalization to IPMA data, particularly for speech and breath signals. The second part evaluated clinical risk classification based on the Manchester Triage System through a structured approach that included data preprocessing, comparison of ML and Deep Learning (DL) models, and usability assessment. Using a public pediatric dataset, ensemble classifiers such as XGBoost and Stacking achieved F1-scores above 0.99 when trained on comprehensive clinical features. Additionally, promising results were obtained using primarily vital signs and low-subjectivity variables, with models reaching F1-scores around 0.74, demonstrating the potential of objective data for low-bias risk stratification in autonomous systems. However, when tested on adult data collected with IPMA, the models showed limited performance, indicating challenges in generalizing across different populations and clinical contexts. Usability was also a central component of this study. Standardized evaluations using the System Usability Scale (SUS) (and Post-Study System Usability Questionnaire — PSSUQ — for the COVID-19 task) indicated high user acceptance of the IPMA by both patients and healthcare professionals. Reported scores reflect the system’s ease of use and perceived integration into clinical workflows, reinforcing its potential for deployment in real-world triage and screening scenarios.
Descrição
Palavras-chave
Detecção de COVID-19 , Protocolo de triagem de Manchester , Aprendizado de máquina , Equipamento médico portátil , Avaliação de usabilidade