Advanced Machine Learning Techniques for Accurate Diagnosis of Respiratory Diseases Using Vocal Biomarkers

Respiratory diseases are the leading morbidity in the world. The correct and speedy diagnostic tools are crucially required for efficient management. This paper examines the power of vocal biomarkers in correctly diagnosing respiratory conditions using advanced machine-learning techniques. We have utilized a benchmark dataset and used sophisticated preprocessing techniques to reduce noise and improve the quality of the signals. We extracted diverse numerical features including time domain, statistical, chroma, harmonic, percussive, entropy, peak-based, and rhythm. All these features were compiled structurally in a data frame for further analysis. We adopted a two-step strategy to deal with class imbalance. First, data augmentation techniques such as time stretch, pitch shift, and random crop were applied before feature extraction to increase the sample size; still, class imbalance persisted. Hence, at the later stage of feature extraction, we introduced a hybrid approach that combines random oversampling and undersampling techniques to make classes balanced, thus ensuring accuracy by measuring and adjusting class weights. The performance was evaluated against several classification traditional algorithms and neural networks. All classifiers and neural network architectures returned out-standing results; the ExtraTrees classifier stood out at 99.47%. This paper focuses on the effectiveness of using vocal biomarkers and machine learning models for the non-invasive but accurate diagnosis of respiratory conditions and opens up ways for further smart healthcare solutions.