This paper proposes an edge-computing-based automated auscultation system for real-time lung sound classification to support nursing activities such as tracheal suctioning. While conventional AI-driven auscultation methods frequently depend on offline analysis, the proposed system performs the entire pipeline — from audio processing to neural network inference — locally on a Sony Spresense edge device. To capture temporal variations, spectrogram images are employed as input features. The system's performance was evaluated through binary, five-class, and seven-class classification tasks. Experimental results indicate that although the system achieves a high accuracy of approximately 95% in offline classification, the performance significantly decreases during real-time inference on the edge device. Furthermore, while the identification of various lung sounds including adventitious sounds was achieved, a substantial decline in accuracy occurred as the number of classification categories increased. These findings demonstrate the feasibility of real-time edge-AI auscultation while simultaneously highlighting the trade-offs between computational resource constraints and classification complexity. This autonomous system establishes a foundation for objective evaluation in home nursing, although further optimization remains necessary to maintain high precision in complex real-time environments.
