Voice Analysis in Asthmatic and Healthy Individuals: A Comparative Assessment of Asthma Control Levels Using Machine Learning Models

Artificial intelligence (AI) and machine learning (ML) models offer promising tools for asthma diagnosis and monitoring. While previous studies have primarily focused on wheezing, coughing, breathing sounds, and phonetic analysis, limited research has explored word-based discrimination and asthma control prediction. This study aimed to leverage AI/ML models to distinguish asthmatic patients from healthy individuals using phonetic sounds and specific words and to classify asthmatic patients as controlled or uncontrolled based on the GINA symptom control criteria.

This study included 344 participants (284 asthmatics, 60 healthy individuals). Each repeated six Turkish words (“ana”, “araba”, “ordu”, “titiz”, “ünlem”) and the phonetic sound "a" for ten seconds. Voice recordings were analyzed using ten ML models: Naïve Bayes (NBA), K-Nearest Neighbors (KNN), Decision Trees (DT), Artificial Neural Networks (ANN), CatBoost, AdaBoost, GradientBoost, Random Forest (RF), Support Vector Machines (SVM), and Logistic Regression (LR). Performance was assessed via sensitivity, specificity, PPV, NPV, accuracy, balanced accuracy, and F1 score.

For distinguishing asthmatics, the highest sensitivity, specificity, accuracy, and F1 score were 99%, 67%, 83%, and 91%, respectively. For classifying uncontrolled asthma based on GINA criteria, these values reached 100%, 57%, 84%, and 91%. RF and CatBoost showed the best sensitivity, accuracy, and F1 scores, while NBA, SVM, and LR performed best in specificity.

Our AI/ML models showed strong potential in identifying asthma patients and distinguishing those with uncontrolled asthma.