Breathalytics: Deep Learning-Based Prediction of Abnormality in Respiratory Sounds using Continuous Wavelet Transform

Analyzing respiratory sounds is vital for diagnosing respiratory conditions, requiring evaluation across time and frequency domains. When diagnosing respiratory disorders, detecting lung sound events is a crucial initial step. This study presents a new method that combines continuous wavelet transform (CWT) with deep learning to predict abnormal respiratory sounds. Using the HV Lung V1 dataset, we denoise respiratory recordings, segmenting them into 63,944 augmented 3-second segments to identify abnormal sound occurrences within the entire recording. We then employ the Analytic Morlet (Gabor) wavelet to extract CWT images, capturing detailed time-frequency representations. The rationale for using CWT lies in its ability to provide detailed respiratory sound representations, crucial for distinguishing abnormal respiratory sounds (e.g., Crackles, Rhonchi, Stridor, and Wheeze) from healthy lung sounds. Our approach involves training and comparing DenseNet and EfficientNet variants, prominent deep learning architectures, to predict abnormal respiratory sounds. By integrating CWT-derived features, we enhance both detection accuracy and efficiency, addressing the need for reliable, remote, and continuous real-time respiratory sound prediction models. Experimental results show the effectiveness of our approach in capturing detailed differences between healthy lung sounds and various abnormal respiratory sounds. This research contributes to advancing respiratory sound analysis and healthcare diagnostics, ultimately improving patient outcomes, and enabling proactive healthcare interventions, especially in scenarios requiring remote monitoring or limited access to healthcare facilities.