This research addresses the limitations of current respiratory diagnostic methods, particularly the need for a readily available, accurate, and real-time tool for assessing respiratory function, especially in the context of Chronic Obstructive Pulmonary Disease (COPD). COPD is a significant global health problem, and current diagnostic methods like spirometry require specialized equipment and trained personnel, limiting accessibility. This study leverages recent advances in flexible sensors and deep learning to develop a novel mechano-acoustic system. Epidermal sensors, utilizing triaxial accelerometers, have shown promise in monitoring physiological signals, often analyzed through frequency domain conversion. This study builds upon this foundation by integrating a wireless sensor with a multimodal deep learning system capable of real-time data processing and analysis. The system aims to overcome limitations of existing methods by providing real-time monitoring, reducing noise interference, and enhancing diagnostic accuracy for respiratory conditions like COPD.

The study reviews previous research on flexible sensors for biomedical applications, highlighting the growing field of epidermal electronics and its use in monitoring various physiological signals. Prior work has demonstrated the value of mechano-acoustic signals from the vocal folds for assessing physiological functions, including respiration rate and cardiac activity. However, these techniques often lack real-time capabilities and are susceptible to environmental noise. The authors cite studies using microphones for acoustic analysis of COPD, noting the limitations of such methods in real-time monitoring and noise susceptibility. They also refer to studies showing the importance of long-term observation in identifying systematic changes in COPD patients' activity.

The study details the design and fabrication of a wireless mechano-acoustic sensor system. The sensor uses a triaxial accelerometer to capture vocal fold vibrations, which are transmitted wirelessly via Bluetooth Low Energy (BLE) to a mobile device and subsequently to a cloud server for analysis. The sensor's design incorporates a medical-grade adhesive for secure and comfortable skin attachment, along with a Polypropylene film for structural integrity and flexibility. The system undergoes rigorous testing, including cyclic bending tests to assess durability, indentation tests to evaluate skin interaction, and adhesion tests to verify long-term stability. Power consumption, temperature stability, and data loss are also carefully monitored and reported. The signal processing pipeline involves several steps, beginning with preprocessing to remove artifacts and noise. Feature extraction utilizes spectrograms, chroma features, mel spectrograms, and mel-frequency cepstral coefficients (MFCCs). These features are then used to train a deep learning model for various classification tasks. The deep learning models, primarily Convolutional Neural Networks (CNNs), are trained and tested using datasets obtained from human subjects. The paper explains the specific architectures, training parameters, and evaluation metrics used for each classification task. The data collection process involved obtaining informed consent from participants and following established ethical guidelines.

The study demonstrates the successful development and validation of a real-time mechano-acoustic system for various classification tasks. The system achieved high accuracy in classifying spoken phrases (95% validation accuracy), distinguishing between users (95% validation accuracy), and classifying gender (high accuracy and recall values comparable to phrase classification). The system also demonstrated the ability to classify COPD severity based on FEV1 values (95% accuracy in distinguishing FEV1 ≥ 60% of predicted value from FEV1 < 60%). The multi-modal deep learning model used for COPD classification integrated features from mel spectrograms, MFCCs, and chroma, achieving better performance than models using individual features. The device showed excellent flexibility, robust adhesion, and minimal temperature increase during prolonged use. Continuous data transmission for extensive periods was demonstrated, showcasing its practical applicability.

The findings of this study demonstrate the potential of integrating flexible sensors and deep learning for accurate and real-time respiratory diagnosis. The high accuracy achieved in various classification tasks, particularly the COPD severity classification, suggests that the system can serve as a valuable tool for early detection and monitoring of respiratory conditions. The system's ability to operate in real-time and its portability through the use of a mobile device provide significant advantages over existing methods. The study's findings contribute to the growing field of wearable healthcare technology, offering a non-invasive and user-friendly approach to respiratory monitoring. The successful integration of multimodal features in COPD classification highlights the importance of considering multiple aspects of respiratory signals for accurate diagnosis. Future studies could explore the generalizability of the system across broader populations and investigate its potential for integration into telehealth platforms.

This research successfully demonstrated a novel real-time mechano-acoustic system for multifunctional classification, including speech, user identification, gender classification, and COPD severity assessment. The system's high accuracy, real-time capabilities, and user-friendly design offer significant advantages over existing methods. Future research could focus on expanding the dataset for improved model generalization, exploring additional applications for the system, and integrating it into clinical settings for broader evaluation.

While the study demonstrates high accuracy in its classification tasks, the sample size, particularly for the COPD classification, is relatively small. Further research with a larger and more diverse population is needed to validate the generalizability of the findings. The study's focus on vocal fold vibrations may limit its applicability to individuals with impaired vocal function. Additional research is necessary to assess the system's performance in diverse environmental conditions and across varying levels of physical activity.