Detecting COPD Through Speech Analysis: A Dataset and Machine Learning Approach
Authors
Sankey-Olsen, Cuno ; Olesen, Rasmus Hvass ; Eberhard, Tobias Oliver
Term
4. Term
Publication year
2025
Submitted on
2025-06-05
Pages
10
Abstract
Kronisk Obstruktiv Lungesygdom (KOL) er en langvarig lungesygdom, der gør vejrtrækningen besværlig. Dette studie undersøger, om enkle stemmeoptagelser, analyseret med maskinlæring, kan hjælpe med at opdage tidlige tegn på KOL. Vi indsamlede tre korte stemmeopgaver fra 96 deltagere og udtrak akustiske egenskaber—målbare mønstre i lyden—ved hjælp af openSMILE og SpeechBrain, to værktøjer til lydanalyse. Fire maskinlæringsmodeller blev trænet og testet på tværs af flere datakonfigurationer. Blandt modellerne præsterede Support Vector Machine (SVM) og Random Forest mest stabilt, især når vi brugte features fra openSMILE. Begrænsningerne omfatter, at diagnoser var selvrapporterede, og at deltagerne ikke altid udførte opgaverne ens. På trods af dette peger resultaterne på, at stemmebaseret analyse kan blive en ikke-invasiv og skalerbar metode til screening for KOL.
Chronic Obstructive Pulmonary Disease (COPD) is a long-term lung condition that makes breathing difficult. This study investigates whether simple voice recordings, analyzed with machine learning, can help detect early signs of COPD. We recorded three short voice tasks from 96 participants and extracted acoustic features—measurable patterns in the sound—using openSMILE and SpeechBrain, two audio-analysis toolkits. Four machine learning models were trained and tested across multiple data configurations. Among the models, Support Vector Machine (SVM) and Random Forest performed most consistently, particularly with features from openSMILE. Limitations include reliance on self-reported diagnoses and inconsistent task execution. Even so, the findings suggest that voice-based analysis could be a non-invasive, scalable screening tool for COPD.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents