Early identification of exacerbation of COPD using voice and cough analysis through a machine learning model

• Grant number: 25/02011-7
• Support Opportunities: Scholarships abroad - Research
• Start date: August 01, 2025
• End date: July 31, 2026
• Field of knowledge: Health Sciences - Medicine - Medical Clinics
• Principal Investigator: Fábio Fernandes Neves
• Grantee: Fábio Fernandes Neves
• Host Investigator: John Robert Hurst
• Host Institution: Centro de Ciências Biológicas e da Saúde (CCBS). Universidade Federal de São Carlos (UFSCAR). São Carlos , SP, Brazil
• Institution abroad: University College London (UCL), England
• Associated research grant: 24/10232-0 - Speaking up for COPD through artificial intelligence in Brazil: Analysis of voice biomarkers for better diagnosis, early identification of exacerbation and assessing the effect of pulmonary rehabilitation, AP.TEM

Abstract

Background: COPD is a debilitating lung disease affecting millions globally, causing significant mortality and economic burden. Exacerbations worsen COPD, leading to increased mortality, readmissions, and healthcare costs. Early treatment of exacerbations improves recovery and reduces hospitalization, but often these episodes go unrecognized. Artificial intelligence, particularly machine learning applied to audio analysis of cough and voice, shows promise for improved ECOPD detection.Objectives: This research aims to develop and test the accuracy of machine learning models, trained on audio samples of cough and voice, to reliably identify ECOPD in hospitalized patients.Methods: This research project aims to develop an AI model for early detection of COPD exacerbations (ECOPD). The study will recruit 20 patients hospitalized for ECOPD at the Royal Free London Hospital. Data collection will involve audio recordings of coughs and phonation, along with EXACT questionnaires, at admission and four weeks post-discharge. Audio samples will be processed using Praat software, converted to wavelet signal images and spectrograms, and analyzed using machine learning algorithms (KNN, LDA, SVM). The best-performing model will be identified using five-fold cross-validation and Particle Swarm Optimization. Model performance will be evaluated using sensitivity, specificity, AUROC, and F1-score. The AI model's accuracy will be compared against the EXACT questionnaire and physician-based diagnosis (Rome criteria). Ethical approval will be obtained, and all participants will provide informed consent.Expected results: with this research it is expected to develop a superior AI model for ECOPD identification, outperforming current methods.