Minimal-preprocessing paradigm for robust COVID-19 detection from cough audio using deep learning

Abstract The COVID-19 pandemic has highlighted the need for rapid, accessible, and non-invasive screening tools to complement conventional diagnostic methods. While audio-based analysis of cough signals has emerged as a promising approach, most existing studies rely on complex preprocessing pipelines aimed at noise suppression and signal enhancement. However, the impact of such preprocessing on model generalisation remains insufficiently understood. This study challenges the prevailing assumption that extensive preprocessing is necessary for robust audio-based diagnosis. We propose a systematic evaluation of preprocessing strategies, investigating whether minimal preprocessing can preserve diagnostically relevant acoustic information and improve model performance. Cough recordings were transformed into time–frequency representations using short-time Fourier transform (STFT) and Mel-frequency cepstral coefficients (MFCCs) and classified using multiple convolutional neural network (CNN) architectures, including ResNet, Inception, and MobileNet variants. Experimental results demonstrate that minimal preprocessing using Fast Fourier Transform (FFT) without additional filtering consistently outperforms conventional noise-reduction techniques. The best-performing model, ResNet50V2, achieved a testing accuracy of 86.79%, with balanced precision, sensitivity, specificity, and F1-score. Furthermore, the combination of STFT and MFCC features provided complementary information, improving classification robustness. These findings indicate that preserving full spectral information is critical for effective deep learning–based audio diagnosis, and that aggressive preprocessing may inadvertently suppress subtle pathological signatures. This work establishes a minimal-preprocessing paradigm for cough-based COVID-19 detection and provides practical insights for developing robust, scalable, and non-invasive diagnostic systems.