rhBNN+ Comprehensive Detections and Analyses of the Human Body Temperatures and Sounds by the Same Smart Mask ^†

Abstract

Scholars have recently used various methods to study the human body’s characteristics, propose improvements, and rely on medical data. This research introduces a new real−time data collection and analysis system comprising two components: a smart mask for data collection and a processing unit for analysis and storage. The medical mask incorporates four sensors: three heat sensors to measure the nose tip and exhaled air temperature from the nostrils and mouth, along with a microphone for audio data. Collected data is instantly transmitted to the processing unit. Temperature readings are labeled and stored for analysis, while audio data is classified into breathing, coughing, and speaking categories using the Support Vector Machine (SVM) algorithm. The SVM’s performance is further optimized using the Lion Algorithm. Evaluation results indicate satisfactory accuracy for temperature data collected from the mask. The SVM achieved 84.3% accuracy for sound data classification, while the Lion algorithm−optimized SVM achieved an even higher accuracy of 93.6%. Future work aims to establish a reliable data collection tool that will serve as the foundation for developing an intelligent human body analysis model called the real human body neural network (rhBNN+).

1. Introduction

The remarkable progress in medical technology has allowed doctors to provide better diagnoses and treatments. Nowadays, Artificial Intelligence (AI) is used in medical research to classify and predict diseases. As a main element in AI, the more data collected, the more efficient the system. Several scholars worked on the modeling of human body temperature. Mondaet al. proposed a low−cost, lightweight remote temperature monitoring solution with 98% accuracy [1]. AirPoP Co. developed a sensor array−integrated mask for detecting breathing parameters [2]. The human thermal database developed by JSC collected relevant data from literature and experiments on human subjects from 1953 to 2004 [3]. Concerning sound databases, it has been researched for urban noise monitoring systems in large cities, environmental sound classification for bird migration research, and sound event detection for emergency sound in construction sites [4,5,6]. However, these datasets are not suitable for medical use.

In this paper, we propose a new data collection system for both body temperatures and sounds. The proposed system collects temperatures from three different sources: the nose tip, nostril exhaled gas, and oral exhaled gas. Next, the collected data is transferred in real−time to the processing unit for labeling and storage. Using the Support Vector Machine (SVM), the sound data is then classified into breathing, coughing, and speaking. The accuracy of the SVM is improved using the Lion algorithm. For the evaluation, the collected temperature from the smart mask shows high accuracy when compared to traditionally measured temperatures. In addition, the improved SVM using the Lion algorithms exhibits an accuracy of 93.6%.

This research paper presents the following key contributions:

• A novel smart mask design is proposed, enabling the collection of human body temperatures and sounds data;
• Optimized SVM models are employed to detect and classify specific sound events, including breathing, coughing, and speaking.

2. Environment Setup and Data Collection

2.1. Smart Mask System

This section introduces the hardware setup for the proposed data collection and analysis. It discusses the hardware used for temperature and sound collection, information analysis, and data storage. The hardware is divided into two parts: the data collection terminal and the data analysis and storage terminal. Concerning the data collection, Figure 1a shows the hardware details of the data collection terminal. Three TMP117 temperature sensors are located in the nose tip, nostrils, and mouth. In addition, the sound is acquired using an MP34DT05−A sound sensor. The collected information is transmitted to the STM32WB55RG main control chip via FPC connectors. Concerning the data transmission to the analysis/storage terminal, which is responsible for displaying the data on the screen, executing the data analysis algorithm, and storing the data on the memory card. More hardware details of the analysis/storage terminal are shown in Figure 1b.

2.2. Data Collection

This section presents the integration of the data collection terminal into a face mask. Figure 2 displays installment results for KN95 and surgical masks. The temperature sensors are positioned near the nose, inside the nostril, and near the oral cavity to ensure accurate data collection. The mask simultaneously collects temperature and sound data and transmits it, along with time information, to the analysis and storage terminal at a 1-s frequency. In addition, human axillary temperature and ambient temperature were measured using an electronic thermometer and the TMP117EVM GUI, respectively.

3. Analysis of Body Temperatures

To evaluate the accuracy of the collected temperatures, the TMP117EVM GUI is simultaneously used to measure the ambient temperature as a reference. Figure 3 displays three iterations of temperature collection with a total duration of 120 s. Results show a maximum deviation of 0.2 degrees Celsius. Considering the TMP117′s accuracy of ±0.1 degrees Celsius, the body temperature data collected using our proposed mask is considered accurate. Figure 4 shows the collected temperature in the morning, noon, and evening.

4. Analysis of Body Sounds

4.1. Data Annotation

The analysis of the raw sound data collected from one microphone poses a challenge. As a solution, we use a classification approach. Initially, we used the open−source audio software Audacity to label our audio data. Subsequently, we preprocess the labeled data and use a Support Vector Machine (SVM) for the classification task. The ability of SVM to handle high−dimensional data makes it suitable for our data. By training the SVM model on the preprocessed audio data, we can classify incoming sound samples into specific classes.

4.2. SVM

The Support Vector Machine (SVM) is a supervised machine learning algorithm. It has gained popularity due to its ability to handle complex datasets and provide robust generalization capabilities. SVM is inherently capable of nonlinear classification and regression tasks through the use of kernel functions. By applying a nonlinear kernel, SVM can map the data into a higher−dimensional space, where it becomes linearly separable. This enables SVM to handle complex decision boundaries and capture intricate relationships within the data. Equation (1) demonstrates the conversion of the original sample coordinates into mapping vectors within the three−dimensional space. This transformation is crucial for accurately determining the optimal classification plane.

$$\begin{array}{l}\min \left[\frac{1}{2}{\displaystyle \sum }_{i=1}^n{\displaystyle \sum }_{j=1}^n{\lambda }_i{\lambda }_j{y}_i{y}_j\left(\theta \left(x_i\right)\cdot \theta \left(x_j\right)\right)-{\displaystyle \sum }_{j=1}^n{\lambda }_i\right]\hfill \\ \mathrm{s}.\mathrm{t}. {\displaystyle \sum }_{i=1}^n{\lambda }_i{y}_i=0,{\lambda }_i\ge 0,C-{\lambda }_i-{\mu }_i=0\hfill \end{array}$$

(1)

In the equation: μ, λ—Lagrange Multiplier;

ϕ(x_i), ϕ(x_j)—The high-dimensional vector converted from the original sample coordinates.

Lion Optimization Algorithm (LOA)

The Lion Optimization Algorithm (LOA) is a nature−inspired algorithm that mimics the hunting behavior and social structure of lions to maximize their hunting success. In this case, the LOA is used on the SVM in order to optimize its parameters.

The LOA starts with the random generation of lions (potential solutions), the evaluation of each lion, and updating the lion pride to reach the optimal solution. Next, several lions are randomly selected to reproduce offspring by crossing over the selected lions’ parameters and mutations. The new offspring go through the fitness evaluation and pride updating with a selected set of best−performing lions to approach the optimal solution. We repeat this process until a specific condition is met, such as the number of iterations or convergence.

5. Model Training and Prediction

In this section, we train the SVM to classify audio into three classes: coughing, breathing, and speaking. As shown in Figure 5, we use the Fast Fourier Transform (FFT) to extract features from the audio signal. FFT helps to illustrate the variations in amplitude, frequency, and intensity for each sound class. For the classification task, we employ three linear kernel SVM models.

The training process uses three SVM models, each model aims to classify one of the three classes. We use manually labeled data for training and testing. The One−vs.−Rest method is deployed with the fitcsvm function to train the SVM models. Next, the trained SVM models are used to predict the test dataset and set the accuracy scores. For each test sample, we create a score vector of 3 to store the positive confidence scores of the three models. If the position of the greatest value in the score vector matches the true class of the test sample, that class’s accuracy is incremented by one. After the loop, the accuracy for each class is calculated and displayed. Finally, the class with the highest accuracy is identified.

6. Discussion

In this study, we employ the FFT as a pre−processing procedure. Next, we use SVM alongside LOA to classify the sound into three classes: coughing, breathing, and speaking. LOA is also implemented to optimize SVM’s hyper−parameter selection. The experiment results show that the SVM achieves an accuracy of 84.3%. However, the SVM optimized by LOA has an accuracy of 93.6%. These results demonstrate a substantial improvement in sound classification accuracy. As a result, we recommend the adoption of the LOA for hyper−parameter optimization to enhance sound classification performance.

7. Conclusions

This research presents a new system for the collection, analysis, and storage of body temperature and sound data. The proposed system incorporates a smart mask with heat and sound sensors for data collection. Additionally, a dedicated processing unit is integrated for the analysis and storage of the collected data. The smart mask uses three temperature sensors in different locations to ensure accurate measurements and one microphone for audio. The data is transmitted to the processing unit in real time, where it undergoes labeling and storage procedures. The collected sound data is classified into three distinct categories: breathing, coughing, and speaking, using SVM, which is enhanced with the Lion Optimization Algorithm. The evaluation demonstrates the accuracy of the collected temperature. Furthermore, the implementation of the Lion algorithm significantly enhances the accuracy of the SVM, achieving an impressive rate of 93.6%.

In terms of future work, we suggest exploring deep−learning algorithms for sound classification. Moreover, the ultimate objective of our proposed data collection system is to provide the necessary inputs and data for the development of an intelligent human body analysis model called the Real Human Body Neural Network (rhBNN⁺).