MeLoDicA AI- Machine Learning Based Detection of Asthma via Vocal Audio Analysis

Mild asthma symptoms can be easily mistaken for those of other health conditions, e.g., allergy and obesity-related symptoms. This might lead to delayed diagnosis and wrong treatment, risking complications and even permanent lung damage. In this study, we introduce a pioneering method for asthma detection. By integrating machine learning (ML) algorithms and advanced audio analysis techniques, the proposed method rapidly and accurately detects asthmatic patients from audio clips containing vocal recordings without the need for dedicated medical devices. By utilising various ML algorithms, the examination of audio clips analyses diverse asthma-related airflow control and respiratory aspects. To strengthen the reliability and accuracy of the model, feature selection algorithms are integrated into the proposed method to address overfitting concerns and retain the most informative features. The model evaluation employs various performance metrics, e.g., accuracy and confusion matrix, to ensure a comprehensive and quantitative assessment of asthma detection. Another novelty of the proposed method is the inclusion of established data analytics tools, e.g., Spotfire, H2O and Sklearn. This approach marks a paradigm shift compared to existing methods, which results in a simplified and more effective asthma detection method. Examining the overall results, vowel pronunciation with its full set of 14 features using the Random Forest (RF) algorithm on the H2O tool yielded the highest accuracy of 92%. Moreover, Support Vector Machine (SVM) and K-Nearest Neighbours (KNN) with correlation-based feature selection (CFS) and customised hyperparameters yielded accuracy of 85% respectively. The proposed method contributes to the advancement of ML theory in healthcare applications, providing a practical framework for the implementation of asthma detection while demonstrating superior performance compared to existing.