Source-linked AI summary
Abnormal respiratory patterns classifier may contribute to large-scale screening of people infected with COVID-19 in an accurate and unobtrusive manner
Yunlu Wang, Menghan Hu, Qingli Li, Xiao-Ping Zhang, Guangtao Zhai, Nan Yao
TL;DR
The paper addresses remote, unobtrusive abnormal respiratory-pattern detection despite scarce real-world training data. It combines a Respiratory Simulation Model with BI-AT-GRU and reports excellent validation performance that outperforms existing state-of-the-art models.
Problem
Remote, unobtrusive respiratory-pattern classification is important, but real-world respiratory data are scarce for training deep models.
Method
The paper uses a depth camera, a Respiratory Simulation Model to generate abundant training data, and a BI-AT-GRU network for abnormal respiratory-pattern classification.
Results
The classifier was validated on 605 real-world depth-camera samples and reportedly achieved excellent performance, outperforming existing state-of-the-art models.
Takeaways & Limitations
The classifier and modeling ideas have potential for extension to large-scale applications such as public places, sleep scenarios, and office environments.
Abstract
from arXiv · showhide
Research significance: The extended version of this paper has been accepted by IEEE Internet of Things journal (DOI: 10.1109/JIOT.2020.2991456), please cite the journal version. During the epidemic prevention and control period, our study can be helpful in prognosis, diagnosis and screening for the patients infected with COVID-19 (the novel coronavirus) based on breathing characteristics. According to the latest clinical research, the respiratory pattern of COVID-19 is different from the respiratory patterns of flu and the common cold. One significant symptom that occurs in the COVID-19 is Tachypnea. People infected with COVID-19 have more rapid respiration. Our study can be utilized to distinguish various respiratory patterns and our device can be preliminarily put to practical use. Demo videos of this method working in situations of one subject and two subjects can be downloaded online. Research details: Accurate detection of the unexpected abnormal respiratory pattern of people in a remote and unobtrusive manner has great significance. In this work, we innovatively capitalize on depth camera and deep learning to achieve this goal. The challenges in this task are twofold: the amount of real-world data is not enough for training to get the deep model; and the intra-class variation of different types of respiratory patterns is large and the outer-class variation is small. In this paper, considering the characteristics of actual respiratory signals, a novel and efficient Respiratory Simulation Model (RSM) is first proposed to fill the gap between the large amount of training data and scarce real-world data. The proposed deep model and the modeling ideas have the great potential to be extended to large scale applications such as public places, sleep scenario, and office environment.
1. INTRODUCTION
The paper targets remote, unobtrusive detection and classification of abnormal respiratory patterns, addressing limitations in contact measurement, existing sensing methods, and available training data.
- Abnormal respiratory patterns can provide clues for disease diagnosis, but people may not notice these patterns themselves.A remote, unobtrusive detector could support earlier identification of affected individuals.
- Contact measurement devices are heavy, expensive, and inconvenient, motivating non-contact respiratory measurement.
- Depth cameras avoid some radar and thermal-imaging drawbacks by measuring displacement rather than pixel-intensity or ambient-heat changes.
- Prior work often extracted signals or detected abnormalities without specifically classifying respiratory patterns, limiting the available clinical clues.
- Real-world respiratory-pattern datasets are limited, while generic deep-learning architectures are not optimized for respiratory-pattern characteristics.
- The paper proposes an AT-BI-GRU classifier and Respiratory Simulation Model to generate training data, validated on 605 depth-camera measurements for remote, unobtrusive pattern measurement.
2. MODELLING PROCEDURE
The modelling procedure combines simulated respiratory waveforms, depth-camera measurements, and BI-AT-GRU classification. RSM models signal variability, while bidirectional and attentional mechanisms capture temporal context and informative waveform points.
- The procedure generates simulated respiratory data, acquires depth-camera data, establishes BI-AT-GRU, and performs comparative experiments.
- Respiratory Simulation Model: RSM approximates respiratory signals with variable depth, rate, longitudinal deviation, and oblique deviation, adding Gaussian white noise for realism.Changing these parameters and combining partial waveforms through breakpoints produces six respiratory patterns.
- Measuring respiratory signal by depth camera: Depth-camera measurement extracts averaged depth values from chest, abdomen, and shoulder regions, then smooths and min-max normalizes the respiratory signals.The three regions address cases where a single region may not work and increase data obtained per measurement.
- BI-AT-GRU for respiratory patterns classification: BI-AT-GRU trains on RSM-generated data and tests on depth-camera data, using recurrent modelling for sequential respiratory waveforms.The architecture contains input, BI-GRU, attention, and output layers.
- BI-AT-GRU for respiratory patterns classification: Bidirectional processing captures forward and reverse temporal information, while attention weights waveform points to form a representation used for classification.The design targets reverse-order information and diagnostically informative turning points in combined patterns such as Central-Apnea.
3. EXPERIMENT AND RESULTS
The BI-AT-GRU was trained on simulated respiratory data and tested on 605 real-world depth-camera samples. It achieved the strongest overall performance among the compared models, while errors mainly involved confusing Cheyne-Stokes with Central-Apnea.
- Experimental setup: 120,000 simulated samples covered six respiratory patterns, while 605 real-world depth-camera samples formed the test set.The training set contained 20,000 randomly generated samples per pattern; erroneous real-world patterns were eliminated.
- Experimental setup: The test set included 108 groups of Eupnea, Bradypnea, and Tachypnea; 97 groups of Cheyne-Stokes and Central-Apnea; and 87 groups of Biots.
- Comparative results: BI-AT-GRU achieved higher accuracy, precision, recall, and F1 than BI-AT-LSTM, GRU, and LSTM on the same real-world test set.The comparison used the same 120,000-sample training set and 605-sample test set for all four models.
- Error analysis: Classification errors mainly involved predicting Cheyne-Stokes as Central-Apnea.The paper attributes this confusion to similar waveform patterns, amplitude sensitivity during normalization, and subject movement.
- Error analysis: BI-AT-GRU had the lowest error rate, while bidirectional and attentional mechanisms improved performance over the corresponding basic networks.
- Confusion matrices: The confusion matrices map real labels on the x axis and predicted labels on the y axis across six respiratory-pattern classes.The classes are Eupnea, Bradypnea, Tachypnea, Biots, Cheyne-Stokes, and Central-Apnea.
4. CONCLUSION
The conclusion presents BI-AT-GRU as a respiratory-pattern classifier trained with simulated data and validated against real-world depth-camera measurements. It reports strong performance relative to existing state-of-the-art models and potential for large-scale extension.
- BI-AT-GRU classifies respiratory patterns using abundant training data generated by a novel Respiratory Simulation Model.
- The classifier shows excellent validation performance and outperforms existing state-of-the-art models.
- The classifier has potential for extension to large-scale applications.