Diagnostic Methods Of The Future: AI Pioneers COVID-19's Diagnostic Revolution
3 main points
✔️ Combining the strengths of magnetic respiration sensing technology (MRST) and machine learning (ML), COVID-19 proposes an approach to real-time monitoring and diagnosis
✔️COVID-19Feature extraction from complex breathing patterns in patients
Classifies patients and healthy subjects with >95% accuracy
Real-Time Magnetic Tracking and Diagnosis of COVID-19 via Machine Learning
writtenby Dang Nguyen, Phat K. Huynh, Vinh Duc An Bui, Kee Young Hwang, Nityanand Jain, Chau Nguyen, Le Huu Nhat Minh, Le Van Truong, Xuan Thanh Nguyen, Dinh Hoang Nguyen, Le Tien Dung, Trung Q. Le, Manh-Huong Phan
(Submitted on 1 Nov 2023)
Comments: Published on arxiv.
Subjects: Machine Learning (cs.LG); Instrumentation and Detectors (physics.ins-det); Medical Physics (physics.med-ph)
The images used in this article are from the paper, the introductory slides, or were created based on them.
Breathing is an essential part of human life, and breathing patterns are an important indicator of a person's health status. Abnormal breathing patterns are also often suggestive of respiratory diseases such as chronic obstructive pulmonary disease (COPD), obstructive sleep apnea (OSA), pneumonia, cystic fibrosis, asthma, and novel coronavirus infection (COVID-19).Among other things, symptoms of novel coronavirus infectionmay includeelevated respiratory rate,decreasedsingle ventilation rate (the amount ofair that enters and exits the airways and lungs with each breath), and irregular respiratory rhythm. Therefore, accurate and rapid assessment of respiratory patterns is essential for the diagnosis and management of novel coronavirus infection (COVID-19) and its variants.
In this paper, we integrate Magnetic Respiratory Sensing Technology (MRST) and Machine Learning (ML) to create and validate a diagnostic platform for real-time tracking and diagnosis of novel coronavirus infection (COVID-19) and other respiratory diseases. We have created and validated a diagnostic platform to track and diagnose novel coronavirus infections (COVID-19) and other respiratory diseases in real time.
Magnetic Breath Detection
Fig 1. below provides an overview of the study and a schematic of the diagnosis and surveillance system for novel coronavirus infections.
This sequence of processes begins with a magnetic respiration monitoring system (Fig. 1a). This system uses a Hall effect sensor to detect changes in the magnetic field produced by a small permanent magnet attached to a person's chest. These changes are caused by respiratory motion. This ensures that numerous breathing patterns can be tracked noninvasively and accurately.
Next, a breath test protocol was employed in the collection of respiratory data to establish a baseline for the study (Fig. 1b). This protocol includes three different breathing styles: normal breathing, breath holding, and deep breathing. The inclusion of these different breathing styles ensures that a wide range of respiratory activity is captured and provides a robust data set for subsequent analysis.
The resulting respiratory data are further refined through specialized algorithms dedicated to signal processing and feature extraction (Fig. 1c). This analysis phase is critical because it helps identify and isolate diagnostic patterns within the respiratory data.
Finally, the extracted features are entered into an ML model (Fig. 1d). This ML model is tailored to identify and diagnose novel coronavirus infection (COVID-19) based on specified respiratory markers.
Subjects and data collection
The study was conducted over a three-month period from July 2021 to October 2021. This was a time when Vietnam was implementing its "Zero COVID" policy, which required the mandatory intensive isolation of all patients with novel coronavirus infection in hospitals or medical camps.
Throughout its three-month duration, the project involved 33 patients with novel coronavirus infection from a medical camp in Binh Tan, Ho Chi Minh City, and 37 healthy participants living in Ho Chi Minh City.
Summary of data collected
From Table 1. below, we can see that, on average, patients with novel coronavirus infection are slightly older than healthy subjects. The standard deviations, however, indicate a wide range of ages in both groups, highlighting the diversity of the data set. In addition, the gender distribution of both groups is nearly balanced, minimizing the potential for gender-centric bias in the survey results.
Of note here is the fact that the average length of hospital stay for patients with novel coronavirus infection was approximately six days, which reveals the severity of illness in this group and its possible manifestation in their breathing patterns. The inclusion of physiological parameters such as temperature, blood pressure, and BMI, as well as symptoms of novel coronavirus infection, provides a comprehensive snapshot of the participants' health status.
Here we delve into a detailed feature extraction analysis based on the collated respiratory signal data. The goal is to highlight the striking differences in breathing patterns between healthy subjects and patients with novel coronavirus infection. Fig. 2. below outlines the peak detection and power spectral density (PSD) analysis of the respiratory signal for both groups during three breathing tests: normal breathing, breath holding, and deep breathing.
The data reveal clear differences in peak patterns and associated spectral content between healthy and affected subjects (Figs. 2a-2l). We also recorded four representative features from both the time and frequency domains: mean respiratory rate (RR), mean prominence (Prom), normalized power spectral density (NPSD), and dominant frequency (Freq), and compared the mean values of these features over three different breathing conditions, as shown in Figs. .2m-2o shows the results of comparing the averaged values of these features across the three different breathing conditions. It is clear that there is a marked difference in these features when comparing normal subjects and COVID-19 subjects, suggesting that the disease has a marked effect on respiratory dynamics.
Fig. 3 thenshows the results of encapsulating the dynamic features of the respiration signalusing RQA, an advanced nonlinear analysis technique.
The RQA results shown in Fig. 3 highlight the characteristic recurrence plots for the two groups under the three respiratory conditions (Figs. 3a-3f). These plots vividly display specific regions, called "fingerprint" regions, that uniquely characterize the complex respiratory dynamics seen in COVID-19 patients. These regions are also marked by three critically important RQA indices: determinacy (DET), entropy (ENT), and laminarity (LAM). The 3D scatter plots (Figs. 3g-3i) of these features for both groups show well-separated clusters, especially for breath holding and deep breathing. This suggests that these RQA indicators have strong discriminative power. Furthermore, Figures 3j-3l compare the six RQA indices between the two groups. The results of comparing these RQA indices show important and characteristic data on the breathing behavior of healthy subjects and COVID-19 patients. However, it is important to recognize that not all of the indices show significant differences across each respiratory condition, reflecting the complex effects of COVID-19 on respiratory dynamics. Therefore, quantifying the statistical significance of these features will be essential for the judicious selection of study features in the ML model.
Feature Selection and Machine Learning Models
In this study, following feature extraction and detailed analysis of the respiratory signal, the most relevant features are narrowed down through a feature selection phase. This allows us to focus on salient attributes and effectively reduce the dimensionality of the data set. This process is visualized in Fig. 4, which shows the results of each of the three breathing tests: normal breathing, breath holding, and deep breathing.
Figures 4a-c show the results of prioritizing features from four feature groups using the Kolmogorov-Smirnov (KS) statistic: time domain, frequency domain, peak analysis, and RQA. For normal breathing, the most prominent features across these groups were Flux mean, NPSD, Prom std, and ENT. In breath holding, Flux BF std, Mean freqAll, Prom AF mean, and ENT were the most discriminating features. For deep breathing, peak2peakAF, PSDAll mean, Prom AF std, and LMAX were the most suggestive features. These insights provided important guidance for distinguishing between healthy subjects and patients with COVID-19 and were reflected in the preliminary feature extraction and analysis.
The results of visualizing the complex feature space using the t-SNE method of manifold learning are shown in Figs. 4d-f. These 3D plot results highlight the distinction between healthy and COVID-19 subjects within the transformed feature domain, highlighting the discriminative ability of the selected features.
After feature selection, the ML models were trained on the COVID-19 cases and evaluated for their effectiveness in distinguishing between COVID-19 cases and healthy controls. The resulting confusion matrix (Fig. 4g-i) is a snapshot of each model's classification ability, showing remarkable accuracy in distinguishing the two cohorts. In addition, Fig. 4j-l shows the results of a five-part cross-validation and a plot of the receiver operating characteristic (ROC) curve. Here, the area under the curve (AUC) quantifies the classification ability of each model. Notably, for normal breathing, the Fine Gaussian SVM model came out on top, recording a sensitivity of 99%, specificity of 94.1%, and average ROC curve area (AUC) of 0.954 in a 5-fold cross-validation. For breath holding, the Bagged Trees model had a slightly lower sensitivity of 94.1% and specificity of 90%, but its AUC value was 0.962. For deep breathing, the Coarse Gaussian SVM model performed best with a sensitivity of 99%, specificity of 94.1%, and AUC value of 0.956.
These evaluations across multiple breath tests and ML models demonstrate the rigor of feature selection in this study and the subsequent model's ability to discriminate between COVID-19 patients and healthy subjects with high accuracy.
This paper proposes a pioneering approach to real-time monitoring and diagnosis of COVID-19 and its variants by combining the strengths of magnetic respiration sensing technology (MRST) and machine learning (ML). This study highlights the effectiveness of respiratory signal features in distinguishing COVID-19 patients from healthy individuals and is a significant contribution to the global effort to combat the pandemic.
This research also represents the tremendous potential of interdisciplinary efforts that draw on health science, engineering, and data science to improve the quality of rapidly growing digital health solutions and address global health challenges, and will improve the outlook for health care monitoring and diagnostic tools in the future. The project is expected to improve the outlook for health care monitoring and diagnostic tools in the future.
Categories related to this article