Utilizing a transformer-based deep learning architecture, the system classified up to six distinct breathing patterns, achieving a validation accuracy of 93.41%. These findings highlight the potential of wearable AI systems for continuous healthcare monitoring.
Challenges in Respiratory Assessment
Traditional pulmonary assessments rely on mouthpiece spirometry, which requires clinical visits and patient cooperation, making it unsuitable for continuous home monitoring. This reliance can lead to undetected gradual changes in breathing patterns until serious conditions arise, such as sleep apnea or chronic obstructive pulmonary disease (COPD).
To address these limitations, researchers have focused on wearable, non-invasive respiratory monitoring systems. Earlier sensors often relied on single sensing methods, such as strain, acoustic, or resistive bands, which were vulnerable to motion artifacts and anatomical variability. This created a need for lightweight, wireless multi-sensor platforms capable of accurately distinguishing respiratory motion from external noise.
Methodology: Novel Dual-Sensor Design
To address the limitations of single-sensor respiratory systems, researchers developed a multi-sensor wearable patch based on an ESP32 C3 microcontroller that combines a six-axis IMU with an analog resistive flex sensor. The flex sensor measured localized chest wall deformation, while the IMU recorded thoracic acceleration. Data transmission was performed via Bluetooth Low Energy (BLE), and signal processing converted raw sensor signals into physical units and applied filtering to reduce motion-related interference.
An adaptive windowing algorithm segmented physiological signals according to detected breathing cycles, preserving complete inhalation and exhalation sequences. Sensor fusion methods utilized accelerometer and gyroscope data to identify body position in real-time.
Three machine learning architectures were evaluated for breathing pattern recognition: aused accelerometer and gyroscope data to estimate body position in real time Transformer model, a hybrid Convolutional Neural Network Long Short-Term Memory (CNN-LSTM) model, and a Histogram Gradient Boosting (HGB) classifier. Different model configurations, including pure, simple, and complex variants, were evaluated to distinguish overlapping breathing patterns.
Furthermore, the system was tested on 20 healthy adult participants, including 13 males and 7 females. The wearable patch was attached to the chest using hypoallergenic adhesive tape, while a high-precision spirometer provided reference respiratory measurements. Participants performed breathing tasks in various body positions, allowing the models to account for body posture-related changes.
Performance Evaluation and Insights
The Transformer model trained with focal loss delivered the highest performance. Using combined IMU and flex sensor data, the model achieved a validation accuracy of 93.41% and a mean area under the curve (AUC) of 0.9919 for the three-class classification task.
On unseen test data, the Transformer maintained an accuracy of 93.06%, outperforming the CNN-LSTM model at 89.58% and the HGB model at 83.33%. The integration of IMU and flex sensor signals improved accuracy by up to 20% compared to flex-only systems.
When the classification task expanded to six categories, the complex Transformer model achieved a holdout accuracy of 78.57%, while the CNN-LSTM model reached 74.03%, and the HGB model achieved 58.12%. Per-class F1 scores showed strong performance for coughing and shallow breathing, with scores of 0.8824 and 0.9189, respectively.
However, confusion arose between deep breathing and yawning, as both produced similar thoracic expansion patterns. Feature importance analysis identified signal entropy and gyroscope z-axis measurements as the most influential across the predictive models.
Applications for Home-Based Digital Healthcare
This dual-sensor wearable patch system can accurately identify different respiratory states without needing clinical supervision, thereby supporting remote monitoring of conditions like sleep apnea, asthma, and chronic bronchitis. Its ability to detect coughing patterns with high precision also makes it suitable for continuous cough frequency monitoring.
Beyond clinical applications, the system could be integrated into consumer health technologies. Real-time respiratory data may help sports scientists evaluate breathing efficiency during physical training. In contrast, the detection of shallow breathing and breath-holding could support wearable stress monitoring and biofeedback systems. The wireless design makes the platform suitable for long-term use outside hospital environments.
Saving this for later? Download a PDF here.
Future Directions in Wearable Respiratory Technology
In summary, this study demonstrates that combining multimodal sensor fusion with deep learning provides a reliable method for non-invasive respiratory classification. By addressing the limitations of traditional spirometry and the noise sensitivity of single-sensor systems, it shows that thoracic rotational movement data improves classification accuracy. The successful conversion of the Transformer models into web-compatible ONNX formats indicates the potential for efficient deployment of AI-based wearable diagnostics.
Future work should focus on testing the system in real-world environments and in patients with diagnosed respiratory disorders to evaluate clinical use. Incorporating larger participant groups and advanced validation methods will further enhance the robustness of the system. Overall, these developments could support the creation of generalized, patient-independent wearable diagnostic systems for continuous digital healthcare monitoring.
Journal Reference
Comeau, C., &. et al. (2026). AI-enabled wireless wearable breathing sensor for breathing pattern recognition. Sci Rep. DOI: 10.1038/s41598-026-49343-z. https://www.nature.com/articles/s41598-026-49343-z
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.