Detecting Patient and Healthy People's Personalized Breathing Patterns with Few-Shot Learning

Analysis of respiratory sounds, such as coughing and breathing, has emerged as a promising non-invasive approach for the early detection of pulmonary conditions, including COVID-19 and chronic obstructive pulmonary disease (COPD), as well as for managing treatment plans. In this work, we explore audio-based classification of respiratory conditions in terms of breathing patterns using the few-shot learning approach that requires only a few samples to develop models. We experimented with three publicly available datasets of audio recordings of breathing patterns of healthy people and patients with COVID-19 or COPD. Through a detailed evaluation using three types of audio features commonly employed for audio event classification, with varying embedding dimensions and shot numbers, we found that Mel-Frequency Cepstral Coefficients (MFCCs) with 20 embedding dimensions can achieve an average accuracy of around 85% using only 10 shots when classifying the breathing patterns obtained from the three datasets. These findings highlight the potential for developing audio-based screening tools that require only a few samples, which can be utilized for public health diagnostics.