Combining Voiceprint Feature Extraction With Deep Learning Based Classifiers for Early Chronic Obstructive Pulmonary Disease Screening

Chronic obstructive pulmonary disease (COPD) is a prevalent chronic illness and a leading cause of global mortality, characterised by irreversible and progressive airflow obstruction. Common symptoms include chronic cough, sputum production, and dyspnoea. However, early‐stage COPD is frequently overlooked by both patients and clinicians, resulting in delayed diagnosis and missed opportunities for timely treatment. To address this issue, this study proposes an early COPD screener based on voiceprint signal feature extraction and a deep learning (DL)‐based classifier. Respiratory‐related voiceprint signals, including normal breathing, coughing, crackles, and wheezes, are acquired and preprocessed using digital filtering techniques, and then a Mel‐frequency cepstral coefficient (MFCC)‐based feature extractor is employed to capture both short‐term spectral characteristics (static information) and temporal energy–frequency variations (dynamic information). Dynamic information is further enhanced through first‐order (Δ) and second‐order (Δ–Δ) derivatives of the Mel‐scale features. Subsequently, these features are transformed into two‐ dimensional (2D) colour‐coded representations to effectively characterise the respiratory dynamics information in spectro‐temporal voiceprint patterns, thereby enhancing the discriminative capability between normal and abnormal breathing conditions. In pattern recognition, two DL‐based classifiers, namely, ‘Δ + Δ–Δ + one‐dimensional (1D) CNN (Convolutional Neural Network)’ and ‘Δ + Δ–Δ + 2D CNN’, are trained for automatic classification of voiceprint signals from healthy subjects and COPD patients using the Kaggle Respiratory Sound Database (KRSD). A 10‐fold cross‐validation with patient‐level separation is employed to evaluate the proposed classifiers’ performances, ensuring that there is no data leakage between the training and validation datasets. The experimental results indicated that the proposed classifiers have promising performance for early COPD detection, with the following approximate average metrics: Precision (%) of 100.0%, Recall (%) of 99.7%, F1 Score of 0.9990, and Accuracy (%) of 99.80%, respectively.