D1.125 - Smartphone-Based Asthma Monitoring Using an Audio-Assisted Mouthpiece and STFT-VGG19 Deep Learning Model

Asthma is a chronic respiratory disease, and the prevalence of allergic diseases hasincreased from 9.4% to 14.5% over the past two decades. Early identification andcontinuous monitoring of asthmatic children are crucial for improving long-term control.Traditional peak expiratory flow meters are often inconvenient to carry and record.Therefore, developing a portable and user-friendly device to monitor asthma status (green-yellow-red zones) is of great importance.

A total of 44 pediatric asthma patients were enrolled. During clinic visits, each child usedboth a conventional peak flow meter and a smartphone-based application, along with anoninvasive auxiliary mouthpiece prototype for airflow and respiratory sound recording. Inaddition to spirometric data, 1,066 respiratory audio samples were collected for modeldevelopment. Short-time Fourier transform (STFT) spectrograms were used as input for aVGG19-based convolutional neural network to classify asthma severity. A self-designedsatisfaction questionnaire evaluated the usability and feasibility of the smartphone app forhome-based asthma self-management.

Use of the mouthpiece significantly improved sound quality and airflow stability: RMS(average sound energy) decreased from 0.0377 to 0.0341 (p=0.041), zero-crossing ratefrom 0.157 to 0.143 (p=0.010), spectral centroid from 1698 Hz to 1574 Hz (p=0.011), andspectral rolloff from 4187 Hz to 3958 Hz (p=0.024), indicating reduced high-frequencynoise and smoother airflow. Material and length analyses showed that TPU’s higherdamping attenuated high-frequency resonance, while PLA preserved details but inducedresonance peaks. However, given the 2-cm short mouthpiece and lower lung capacity inchildren, these material effects were minimal compared to oral structure and blowingpatterns.The STFT-VGG19 model showed robust performance, particularly in distinguishinghealthy (Class 0) and severe (Class 3) categories. Macro-F1 and weighted F1 scores were0.35 and 0.41, respectively. Class-wise analysis indicated strong recall for Class 0 (0.77)and balanced performance for Class 3 (≈0.625 across metrics).

Despite the limited sample size, the proposed system demonstrated promising performance

in multi-class pediatric asthma classification. The combination of a smartphone recorderand short mouthpiece offers both practicality and clinical potential. This approach maysupport early detection and monitoring of pulmonary function in real-world pediatricasthma management.