Laboratory Electronic Nose, Miniaturized Electronic Nose in the Cabin, and Primary Screening Method and System for Diabetes

Abstract

A laboratory electronic nose, a miniaturized in-cabin electronic nose, and a method for initial screening of diabetes belong to the technical field of diabetes detection. The present invention comprises: 1. Establishing a laboratory electronic nose for machine learning model training and electronic nose miniaturization, including: preparing a laboratory electronic nose; collecting cabin odor samples; preprocessing odor sample data; extracting and selecting features from the data to obtain the sensor array model used in the miniaturized electronic nose; classifying the data to obtain a machine learning model; and establishing a cloud initial screening system and cloud platform. 2. Setting up a miniaturized in-cabin electronic nose to perform data detection for diabetes before starting the car and while driving; using the miniaturized in-cabin electronic nose for initial screening of diabetes. The present invention can achieve painless, efficient, and low-cost detection of diabetes, and leverages the advantages of cloud resources to achieve painless and efficient detection of diabetes, thereby preventing diabetic patients from traffic accidents caused by diabetes and ensuring road traffic safety.

Description

Technical Field

[0001] The present invention belongs to the technical field of diabetes detection, and in particular relates to a laboratory electronic nose, a miniaturized electronic nose in a vehicle cabin, and a diabetes initial screening method and system. Background Art

[0002] As people's quality of life improves, their daily routines have undergone dramatic changes, impacting their health to a certain extent. Poor dietary habits can contribute to the development of diabetes, a metabolic disease characterized by high blood sugar levels. Long-term high blood sugar levels can lead to chronic damage and dysfunction of tissues such as the eyes, kidneys, and heart. Diabetes and its related diseases have become a global health concern, garnering widespread attention.

[0003] Traffic accidents caused by abnormal blood sugar levels due to diabetes are common, and diabetic drivers are a higher proportion of motor vehicle accidents. The primary reason why diabetic drivers are more dangerous is the impact of diabetes complications, such as drowsiness, visual impairment, decreased perception due to peripheral neuropathy, and cognitive impairment caused by cerebrovascular disease, which significantly affect driving ability. People with diabetes are also more likely to develop foot ulcers, which can impair their ability to pedal and prevent them from controlling the accelerator and brakes steadily and promptly. Furthermore, people with diabetes are prone to hypoglycemia. Because diabetics rely on medication to lower their blood sugar levels for a long time, excessive doses of insulin or other hypoglycemic medications can easily lead to low blood sugar levels, resulting in hypoglycemia. Hypoglycemia can cause symptoms such as headaches, blurred vision, decreased consciousness, and even coma. This can slow a driver's reaction time and impair cognitive function, impacting their ability to drive properly.

[0004] Normal blood sugar levels ensure traffic safety. Some countries have implemented laws and regulations limiting the duration and scope of driver's licenses, restricting diabetic drivers, especially those driving large trucks or public passenger vehicles. Therefore, monitoring the diabetes status of vehicle drivers is very necessary.

[0005] Traditional medical testing methods require patients to prick their fingers to obtain blood samples, and then measure the glucose level in the blood to diagnose diabetes. This is painful for patients. Furthermore, blood glucose monitoring requires frequent testing and is cumbersome, seriously affecting patients' normal lives and creating numerous traffic safety risks.

[0006] In recent years, breath analysis has been widely used due to its non-invasive, real-time nature and effectiveness across a wide range of patients and diseases. Studies have shown that different diseases have distinct volatile organic compound signatures, and by analyzing the concentration of biomarkers in breath, we can achieve preliminary detection of diseases. Currently, techniques such as gas chromatography-mass spectrometry and ion mobility spectrometry are commonly used to analyze the components in the breath of subjects. However, the detection instruments for these methods have shortcomings such as high cost, complex operation, lack of portability, and low efficiency, which limit their large-scale application. The electronic nose is an instrument that mimics the sense of smell of mammals and can rapidly detect and identify complex gases. It has the advantages of low cost, high efficiency, and non-invasiveness, and has received widespread attention in the detection of gas volatile biomarkers. These technologies integrate artificial neural networks to improve existing clinical disease detection methods.

[0007] Studies have shown that diabetic patients have abnormal concentrations of acetone and some volatile organic compounds in their breath. Therefore, based on the relationship between blood glucose and breath acetone concentrations, a corresponding miniaturized in-vehicle electronic nose system was designed. Combined with traditional pattern recognition methods, a blood glucose monitoring system was developed. The machine learning model used for this testing was further deployed to the cloud, where data was processed and successfully differentiated between diabetic patients and healthy subjects. Finally, the diagnostic results were transmitted to a smart mobile terminal for initial disease screening. The disease status was then linked to vehicle start-up, enabling real-time, accurate, and shared patient diagnosis, ensuring patients do not drive and improving traffic safety. Summary of the Invention

[0008] The purpose of this invention is to provide a method for non-invasive and rapid initial diabetes screening using a miniaturized electronic nose deployed in a vehicle cabin based on cloud deployment. This method can solve the problems of existing diabetes detection technology, such as high instrument cost, poor portability, low detection efficiency, and pain for patients. At the same time, it realizes anonymous, non-invasive, rapid detection and large-scale application, and includes the following contents:

[0009] The system of the present invention for performing preliminary screening of diabetes using a laboratory electronic nose and a miniaturized electronic nose is embodied by the following specific operation process:

[0010] The laboratory electronic nose and the miniaturized electronic nose are composed of an air inlet pipe I (1), an array of 3≤N≤32 gas sensors I (2), a sensor chamber I (3), a connecting pipe I (4), an air outlet pipe I (5), an air pump I (6), a conditioning circuit board I (7), a data acquisition card (8) and a USB output interface (9), wherein the sensor chamber I (3) and the air pump I (6) are arranged front to back, the inlet of the sensor chamber I (3) is connected to the rear end of the air inlet pipe I (1); the outlet of the sensor chamber I (3) is connected to the inlet of the air pump I (6) via the connecting pipe I (4), and the outlet of the air pump I (6) is connected to the The air outlet pipe I (5) is connected; N gas sensor arrays I (2) are uniformly surrounded and fixed on the inner wall of the sensor chamber I (3), and are connected to the data acquisition card (8) through the conditioning circuit board I (7), and the data acquisition card (8) is provided with a USB output interface (9); the miniaturized electronic nose (B) is composed of a wireless data acquisition card (17), a conditioning circuit board II (16), an air pump II (15), an air outlet pipe II (14), a connecting pipe II (13), a sensor chamber II (12), 3≤n≤32 gas sensor arrays II (11) and an air inlet pipe II (10), wherein the air pump II (15) and the sensor chamber II

[0011] (12) are arranged front and back, the inlet of the sensor chamber II (12) is connected to the rear end of the air inlet pipe II (10), the outlet of the sensor chamber II (12) is connected to the inlet of the air pump II (15) through the connecting pipe II (13); the outlet of the air pump II (15) is connected to the outlet pipe II (14); n gas sensor arrays II (11) are uniformly surrounded and fixed in the sensor chamber II (12), and are connected to the wireless data acquisition card (17) through the conditioning circuit board II (16);

[0012] 1) Connect the inlet of the laboratory electronic nose (A) through the air inlet pipe I (1) to the air bag storing the gas to be tested, and the detected gas is discharged through the air outlet pipe I (5); the data collected by the data acquisition card (8) is connected to the computer terminal through the USB output interface (9);

[0013] 2) Classification and acquisition of odor information of diabetic patients and healthy passengers, including the following steps:

[0014] 2.1) Based on the actual situation in the car, the odor types are divided into: "exhaled breath of diabetic patients" and "exhaled breath of healthy volunteers", recorded as G1 and G2 respectively;

[0015] 2.2) Breath samples from healthy volunteers were collected in the morning without eating. The healthy volunteers provided blood samples containing blood glucose level information, and the corresponding blood glucose values were estimated by a physician using the glucose oxidase method. Exhaled breath G1 from diabetic patients was collected in a 1-liter gas collection bag; exhaled breath G2 from healthy volunteers was collected in a 1-liter gas collection bag.

[0016] 2.3) The sampling frequency of the laboratory electronic nose (A) was set to 100 Hz. Each sampling session was divided into two phases: odor sample data collection and sensor array cleaning. The sampling time was set to T1, and the sensor array cleaning time was set to T2. The total sampling time was recorded as T.

[0017] 2.4) Connect the G1 gas collection bag to the inlet of the air inlet pipe I (1) of the laboratory electronic nose (A), and draw the gas to be tested into the sensor chamber I (3) where the sensor is installed through an air pump with a flow rate of 1.2 L / min. Then, use the laboratory electronic nose (A) to sample and obtain odor sample data; connect the G2 gas collection bag to the inlet of the air inlet pipe I (1) of the laboratory electronic nose (A), and obtain odor sample data; the odor sample data obtained by G1 and G2 are recorded as G1 sample and G2 sample, respectively;

[0018] 3) Using the odor sample data collected in step 2.4), the particle swarm algorithm is used to optimize the sensor array to reduce the number of sensors and achieve miniaturization of the electronic nose, including the following steps:

[0019] 3.1) For the raw data obtained by the N gas sensor array I (2), the discrete wavelet transform method of the db1-based wavelet is used to reconstruct the original breath odor sample data and remove the signal noise. The preprocessed signal is expressed as:

[0020] DWT(u,w)= <x(t),ψ u,w (T)>

[0021]

[0022] Where: x(t) is the original signal, u is the scale parameter, w is the shift parameter, ψ is the mother wavelet; the processed data is denoted as di, where: i = 1, 2, 3...100T1, and the preprocessed sample data are denoted as g1 samples and g2 samples respectively;

[0023] 3.2) Two methods are used to extract features from the preprocessed and denoised signals. The first method extracts features directly from the raw sensor response, while the second method extracts features after Fourier transform. Using the first method, five features are extracted from each sensor response signal: mean, standard deviation, kurtosis, skewness, and waveform factor.

[0024] The formula for calculating the average value is:

[0025] Standard deviation calculation formula:

[0026] Kurtosis index calculation formula:

[0027] Skewness calculation formula:

[0028] Form factor calculation formula:

[0029] 3.3) Use three ranking algorithms: XGBoost, LightGBM, and ExtraTrees to rank the feature importance and select the most promising features to form a feature subset;

[0030] 3.4) Using particle swarm optimization for feature selection, including the following steps:

[0031] 3.4.1) Set the population size v and randomly initialize v particles to form the initialization population X = {X1, X2, ..., Xv};

[0032] 3.4.2) Set the number of iterations and other parameters, and set the current number of iterations T = 1;

[0033] 3.4.3) Update the velocity vector of each individual:

[0034] V id t+1 =w*V id t +c1*r 1i *(P id t -X id t )+c2*r 2i *(P gd t -X id t )

[0035] Where w represents the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random numbers uniformly distributed on [0,1], and V id t and X id t represents the velocity and position of the particle in d dimension at time t, P id t and P gd t Represents the local optimal position vector Pbesti of the individual particle in d dimension at time t and the global optimal position vector Gbest of the population;

[0036] 3.4.4) Update the position vector of each individual:

[0037] X id t+1 =X idt +V id t+1

[0038] 3.4.5) Update local and global position vectors: Update the individual's local optimal position vector Pbesti and the population's global optimal position vector Gbest. Use the fitness function to find the optimal position Pbesti of each particle after iteration i, and compare it with other particles to obtain the global optimal position Gbest.

[0039] 3.4.6) Perform termination condition judgment. If the number of iterations reaches the number of iterations T max , then output Gbest; if not satisfied, return to step 3.4.3) and continue iterating;

[0040] 3.5) After performing noise reduction, feature extraction, feature sorting, and feature selection on the collected odor sample data in steps 3.2), 3.3, and 3.4), determine n gas sensors to be included in the sensor array of the miniaturized electronic nose (B) based on the perspective of higher diabetes screening accuracy;

[0041] 4) using the random forest algorithm to perform model training on the n-sensor miniaturized electronic nose selected in step 3);

[0042] 5) Building a cloud-based initial screening system; combining the machine learning model obtained in the above steps with cloud deployment, deploying a cloud-based initial screening system consisting of a model training end and a data prediction end; deploying a miniaturized electronic nose (B) in the hospital and in the vehicle's cabin, respectively; using the electronic nose deployed in the hospital to collect respiratory gas samples and disease conditions from patients, transmitting them to the model training end for data processing and model training; using the in-vehicle electronic nose to obtain odor sample data from the vehicle cabin, and transmitting the data to the cloud-based data prediction end; the data prediction end uses the trained model to predict diabetes based on the odor sample data, and sends the prediction results to the bound person's mobile phone and the built-in software on the vehicle's central control screen for health reminders;

[0043] 6) Construct the overall architecture of the cloud platform; the cloud platform includes a data sharing part, a data training part, and a model usage part. In the data sharing part, after obtaining permission authentication, the client will encrypt the data request command and transmit it to the cloud server. The cloud server decrypts it and sends the required data to the client for data sharing. In the data training part, relevant patient data is uploaded and the recognition accuracy of the model is improved through desensitized storage and model training of the uploaded data. The vehicle-mounted electronic nose detection terminal uses the model usage part. After uploading the data to the detection terminal, the trained model is used for detection and judgment. Finally, the detection results are sent to the smart mobile terminal to complete the real-time notification of the detection results.

[0044] 7) Mounting a miniaturized electronic nose (B) on a car to obtain cabin gas data for diabetes prediction includes the following steps:

[0045] 7.1) The inlet of the miniaturized electronic nose (B) is placed at the driver's seat in the vehicle cabin, and the detected gas is discharged through the outlet pipe II (14); the odor sample data collected by the wireless data acquisition card (17) is directly uploaded to the cloud server;

[0046] 7.2) After the driver enters the vehicle and fastens his seat belt, the driver's exhaled breath is collected through the intake pipe II (10) placed on the driver's seat in the vehicle cabin. The odor sample data is collected by the wireless data acquisition card (17) and uploaded to the cloud server; the data is processed in the cloud, and the transmitted data is judged using a machine learning model. Finally, the diagnosis result is encrypted and transmitted to the corresponding smart mobile terminal, completing the initial diabetes screening and result notification; in the cabin, if the driver does not detect diabetes, the car starts normally; if diabetes is detected, a reminder is issued and the car cannot be started temporarily.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The method proposed in this invention uses a miniaturized in-cabin electronic nose and a cloud-based machine learning algorithm for initial diabetes screening. This method can accurately distinguish between "exhaled breath of diabetic patients" and "exhaled breath of healthy volunteers," enabling large-scale, real-time, painless, low-cost, and efficient initial screening of diabetic patients.

[0049] 2. Miniaturizing the electronic nose system can reduce equipment costs and power consumption, making it easier to install it in the vehicle cabin.

[0050] 3. Use cloud processing platforms to improve the initial screening rate of diabetic patients and maximize the reliability of cloud platform services and the utilization of data resources.

[0051] 4. By diagnosing the condition of diabetic drivers and issuing warnings to them, traffic accidents caused by diabetes and its symptoms can be avoided to a great extent, and road traffic safety can be effectively guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a schematic diagram of the structure of the laboratory electronic nose A;

[0053] Figure 2 Schematic diagram of the structure of the miniaturized electronic nose B;

[0054] Figure 3 Flowchart of the method for initial diabetes screening using a cloud-based miniaturized electronic nose in a vehicle cabin;

[0055] Figure 4 It is a data processing flow chart;

[0056] Figure 5 This is the flow chart of the cloud-based primary screening system;

[0057] Figure 6 This is the overall structure diagram of the cloud processing platform;

[0058] Figure 7 A flowchart for the initial screening of diabetes for cabin occupants;

[0059] Among them: A. Laboratory electronic nose B. Miniaturized electronic nose 1. Inlet pipe I 2. Gas sensor array I 3. Sensor chamber I 4. Connecting pipe I 5. Outlet pipe I 6. Air pump I 7. Conditioning circuit board I 8. Data acquisition card 9. USB output interface 10. Inlet pipe II 11. Gas sensor array II 12. Sensor chamber II 13. Connecting pipe II 14. Outlet pipe II 15. Air pump II 16. Conditioning circuit board II 17. Wireless data acquisition card DETAILED DESCRIPTION

[0060] The present invention will be described below with reference to the accompanying drawings.

[0061] The system of the present invention for performing preliminary screening of diabetes using a laboratory electronic nose and a miniaturized electronic nose is embodied by the following specific operation process:

[0062] The laboratory electronic nose (A) is composed of an air inlet pipe I (1), an array of 32 gas sensors I (2), a sensor chamber I (3), a connecting pipe I (4), an air outlet pipe I (5), an air pump I (6), a conditioning circuit board I (7), a data acquisition card (8) and a USB output interface (9), wherein the sensor chamber I (3) and the air pump I (6) are arranged front and back, and the inlet of the sensor chamber I (3) is connected to the rear end of the air inlet pipe I (1); the outlet of the sensor chamber I (3) is connected to the inlet of the air pump I (6) via the connecting pipe I (4), and the air pump I

[0063] (6) The outlet is connected to the gas outlet pipe I (5); 32 gas sensor arrays I (2) are evenly surrounded and fixed in the sensor chamber I

[0064] (3) inner wall, and connected to the data acquisition card (8) through the conditioning circuit board I (7), and the data acquisition card (8) is provided with a USB output interface (9); the 32 gas sensor models are TGS2612, TGS2611, TGS2620, TGS2603, TGS2602, TGS2610, TGS2600, GSBT11, MS1100, MP135, MP901, MP-9, MP-3B, MP-4, MP-5, MP-2, MP 503, MP801, MP905, MP402, WSP1110, WSP2110, WSP7110, MP-7, TGS-2612, TGS-2611, TGS-2620, MP-3B, MP702, TGS2610, TGS2600, TGS2618-COO; the miniaturized electronic nose (B) is composed of a wireless data acquisition card (17), a conditioning circuit board II (16), an air pump II (15), an air outlet pipe II (14), a connecting pipe II

[0065] (13), a sensor chamber II (12), 14 gas sensor arrays II (11) and an air inlet pipe II (10), wherein an air pump II (15) and the sensor chamber II (12) are arranged front to back, an inlet of the sensor chamber II (12) is connected to the rear end of the air inlet pipe II (10), an outlet of the sensor chamber II (12) is connected to the inlet of the air pump II (15) via a connecting pipe II (13); an outlet of the air pump II (15) is connected to the outlet pipe II

[0066] (14) connection; 14 gas sensor arrays II (11) are evenly surrounded and fixed in the sensor chamber II (12), and are connected to the wireless data acquisition card (17) through the conditioning circuit board II (16).

[0067] 1) Connect the inlet of the laboratory electronic nose (A) through the air inlet pipe I (1) to the air bag storing the gas to be tested, and the detected gas is discharged through the air outlet pipe I (5); the data collected by the data acquisition card (8) is connected to the computer terminal through the USB output interface (9);

[0068] 2) Classification and acquisition of odor information of diabetic patients and healthy passengers, including the following steps:

[0069] 2.1) Based on the actual situation in the car, the odor types are divided into: "exhaled breath of diabetic patients" and "exhaled breath of healthy volunteers", recorded as G1 and G2 respectively;

[0070] 2.2) Breath samples from healthy volunteers were collected in the morning without eating. The healthy volunteers provided blood samples containing information about their blood glucose levels, and a physician estimated their corresponding blood glucose levels using the glucose oxidase method. A total of 495 breath samples (G1) were collected from diabetic patients using a 1-liter gas collection bag. A total of 374 breath samples (G2) were collected from healthy volunteers using a 1-liter gas collection bag, for a total of 869 samples.

[0071] 2.3) The sampling frequency of the laboratory electronic nose (A) was set to 100 Hz. Each sampling session was divided into two phases: odor sample data collection and sensor array cleaning. The sampling time was set to 60 s, and the sensor array cleaning time was set to 180 s, for a total sampling time of 240 s.

[0072] 2.4) Connect the G1 gas collection bag to the inlet of the air inlet pipe I (1) of the laboratory electronic nose (A), and draw the gas to be tested into the sensor chamber I (3) where the sensor is installed through an air pump with a flow rate of 1.2 L / min. Then, use the laboratory electronic nose (A) to sample and obtain odor sample data; connect the G2 gas collection bag to the inlet of the air inlet pipe I (1) of the laboratory electronic nose (A), and obtain odor sample data; the odor sample data obtained by G1 and G2 are recorded as G1 sample and G2 sample, respectively;

[0073] 3) Using the odor sample data collected in step 2.4), the particle swarm algorithm is used to optimize the sensor array to reduce the number of sensors and achieve miniaturization of the electronic nose, including the following steps:

[0074] 3.1) The original data obtained by the 32 gas sensor arrays I (2) are reconstructed using the discrete wavelet transform method of the db1-based wavelet to remove the signal noise. The preprocessed signal is expressed as:

[0075] DWT(u,w)= <x(t),ψ u,w (T)>

[0076]

[0077] Where: x(t) is the original signal, u is the scale parameter, w is the shift parameter, ψ is the mother wavelet; the processed data is denoted as di, where: i = 1, 2, 3...6000, and the preprocessed sample data are denoted as g1 sample and g2 sample respectively;

[0078] 3.2) Two methods are used to extract features from the preprocessed and denoised signals. The first method extracts features directly from the raw sensor response, while the second method extracts features after Fourier transform. Using the first method, five features are extracted from each sensor response signal: mean, standard deviation, kurtosis, skewness, and waveform factor.

[0079] The formula for calculating the average value is:

[0080] Standard deviation calculation formula:

[0081] Kurtosis index calculation formula:

[0082] Skewness calculation formula:

[0083] Form factor calculation formula:

[0084] 3.3) Use three ranking algorithms: XGBoost, LightGBM, and ExtraTrees to rank the feature importance and select the most promising features to form a feature subset;

[0085] 3.4) Using particle swarm optimization for feature selection, including the following steps:

[0086] 3.4.1) Set the population size to 50 and randomly initialize 50 particles to form the initial population X = {X1, X2, ..., X50};

[0087] 3.4.2) Set the parameters: set the maximum number of iterations to 200, the feature threshold to 0.6, the inertia weight w to 0.9, the learning factors c1 and c2 to c1 = c2 = 2, r1 and r2 to random probability values between [0, 1], and set the current number of iterations T to 1;

[0088] 3.4.3) Update the velocity vector of each individual:

[0089] V id t+1 =w*V id t +c1*r 1i *(P id t -X id t )+c2*r 2i *(P gd t -X id t )

[0090] Where w represents the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random numbers uniformly distributed on [0,1], and V id t and X id t represents the velocity and position of the particle in d dimension at time t, P id t and P gd t Represents the local optimal position vector Pbesti of the individual particle in d dimension at time t and the global optimal position vector Gbest of the population;

[0091] 3.4.4) Update the position vector of each individual:

[0092] X id t+1 =X id t +V id t+1

[0093] 3.4.5) Update local and global position vectors: Update the individual's local optimal position vector Pbesti and the population's global optimal position vector Gbest. Use the fitness function to find the optimal position Pbesti of each particle after iteration i, and compare it with other particles to obtain the global optimal position Gbest.

[0094] 3.4.6) Perform termination condition judgment. If the number of iterations reaches the number of iterations T max , then output Gbest; if not satisfied, return to step 3.4.3) and continue iterating;

[0095] 3.5) After performing noise reduction, feature extraction, feature sorting, and feature selection on the collected odor sample data in steps 3.2), 3.3, and 3.4), sensors are selected from the perspective of higher diabetes screening accuracy. The 14 gas sensors that will be included in the sensor array of the miniaturized electronic nose (B) are determined. The models of the 14 sensors are TGS2602, TGS2610, TGS2600, MS1100, MP135, MP-3B, MP503, MP801, MP905, MP402, MP-7, TGS-2611, MP-3B, and MP702.

[0096] 4) Using the random forest algorithm to train the model of the 14 miniaturized electronic nose sensors selected in step 3);

[0097] 5) Building a cloud-based initial screening system; combining the machine learning model obtained in the above steps with cloud deployment, deploying a cloud-based initial screening system consisting of a model training end and a data prediction end; deploying a miniaturized electronic nose (B) in the hospital and in the vehicle's cabin, respectively; using the electronic nose deployed in the hospital to collect respiratory gas samples and disease conditions from patients, transmitting them to the model training end for data processing and model training; using the in-vehicle electronic nose to obtain odor sample data from the vehicle cabin, and transmitting the data to the cloud-based data prediction end; the data prediction end uses the trained model to predict diabetes based on the odor sample data, and sends the prediction results to the bound person's mobile phone and the built-in software on the vehicle's central control screen for health reminders;

[0098] 6) Construct the overall architecture of the cloud platform; the cloud platform includes a data sharing part, a data training part, and a model usage part. In the data sharing part, after obtaining permission authentication, the client will encrypt the data request command and transmit it to the cloud server. The cloud server decrypts it and sends the required data to the client for data sharing. In the data training part, relevant patient data is uploaded and the recognition accuracy of the model is improved through desensitized storage and model training of the uploaded data. The vehicle-mounted electronic nose detection terminal uses the model usage part. After uploading the data to the detection terminal, the trained model is used for detection and judgment. Finally, the detection results are sent to the smart mobile terminal to complete the real-time notification of the detection results.

[0099] 7) Mounting a miniaturized electronic nose (B) on a car to obtain cabin gas data for diabetes prediction includes the following steps:

[0100] 7.1) The inlet of the miniaturized electronic nose (B) is placed at the driver's seat in the vehicle cabin, and the detected gas is discharged through the outlet pipe II (14); the odor sample data collected by the wireless data acquisition card (17) is directly uploaded to the cloud server;

[0101] 7.2) After the driver enters the vehicle and fastens his seat belt, the driver's exhaled breath is collected through the intake pipe II (10) placed on the driver's seat in the vehicle cabin. The odor sample data is collected by the wireless data acquisition card (17) and uploaded to the cloud server; the data is processed in the cloud, and the transmitted data is judged using a machine learning model. Finally, the diagnosis result is encrypted and transmitted to the corresponding smart mobile terminal, completing the initial diabetes screening and result notification; in the cabin, if the driver does not detect diabetes, the car starts normally; if diabetes is detected, a reminder is issued and the car cannot be started temporarily.

[0102] 8) The accuracy of the initial screening system was evaluated using a random forest algorithm. 30 randomized data sets were selected as the test set and 70% as the training set. The algorithm was run 30 times and the average accuracy, classification precision, recall, and F1 score were calculated. The results are shown in the table below. The results show that the diabetes initial screening system achieved an accuracy of 92.12%, while reducing the number of sensors from 32 to 14. While maintaining accuracy, this system further miniaturized the testing equipment and reduced energy consumption, facilitating in-vehicle diabetes initial screening.

[0103] Accuracy (%) Classification accuracy (%) Recall rate (%) F1 score (%) 92.12 93.19 92.93 93.02 .

Claims

1. A system for primary screening of diabetes using a laboratory electronic nose and a miniaturized in-vehicle electronic nose, characterized by: It is reflected by the following specific operation process: The laboratory electronic nose and the miniaturized electronic nose are composed of an air inlet pipe I (1), 3≤N≤32 gas sensor arrays I (2), a sensor chamber I (3), a connecting pipe I (4), an air outlet pipe I (5), an air pump I (6), a conditioning circuit board I (7), a data acquisition card (8) and a USB output interface (9), wherein the sensor chamber I (3) and the air pump I (6) are arranged front to back, and the inlet of the sensor chamber I (3) is connected to the rear end of the air inlet pipe I (1); the outlet of the sensor chamber I (3) is connected to the inlet of the air pump I (6) via the connecting pipe I (4), and the outlet of the air pump I (6) is connected to the air outlet pipe I (5); N gas sensor arrays I (2) are uniformly surrounded and fixed on the inner wall of the sensor chamber I (3), and are connected to the data acquisition card (8) via the conditioning circuit board I (7), and the data acquisition card (8) is connected to the data acquisition card (8). A USB output interface (9) is provided; the miniaturized electronic nose (B) is composed of a wireless data acquisition card (17), a conditioning circuit board II (16), an air pump II (15), an air outlet pipe II (14), a connecting pipe II (13), a sensor chamber II (12), 3≤n≤32 gas sensor arrays II (11) and an air inlet pipe II (10), wherein the air pump II (15) and the sensor chamber II (12) are arranged front to back, the inlet of the sensor chamber II (12) is connected to the rear end of the air inlet pipe II (10), the outlet of the sensor chamber II (12) is connected to the inlet of the air pump II (15) via the connecting pipe II (13); the outlet of the air pump II (15) is connected to the air outlet pipe II (14); n gas sensor arrays II (11) are uniformly surrounded and fixed in the sensor chamber II (12), and are connected to the wireless data acquisition card (17) via the conditioning circuit board II (16); 1) Connect the inlet of the laboratory electronic nose (A) through the air inlet pipe I (1) to the air bag storing the gas to be tested, and the detected gas is discharged through the air outlet pipe I (5); the data collected by the data acquisition card (8) is connected to the computer terminal through the USB output interface (9); 2) Classification and acquisition of odor information of diabetic patients and healthy passengers, including the following steps: 2.1) Based on the actual situation in the car, the odor types were divided into: "exhaled breath of diabetic patients" and "exhaled breath of healthy volunteers", recorded as G1 and G2 respectively; 2.2) Breath samples from healthy volunteers were collected in the morning without eating. The healthy volunteers provided blood samples containing blood glucose level information, and their corresponding blood glucose values were estimated using the glucose oxidase method. The exhaled breath G1 from the diabetic patients was collected in a 1L gas collection bag; the exhaled breath G2 from the healthy volunteers was collected in a 1L gas collection bag. 2.3) The sampling frequency of the laboratory electronic nose (A) was set to 100 Hz. Each sampling session was divided into two phases: odor sample data collection and sensor array cleaning. The sampling time was set to T1, and the sensor array cleaning time was set to T2. The total sampling time was recorded as T. 2.4) Connect the G1 gas collection bag to the inlet of the air inlet pipe I (1) of the laboratory electronic nose (A), and draw the gas to be tested into the sensor chamber I (3) where the sensor is installed through an air pump with a flow rate of 1.2 L / min. Then, use the laboratory electronic nose (A) to sample and obtain odor sample data; connect the G2 gas collection bag to the inlet of the air inlet pipe I (1) of the laboratory electronic nose (A), and obtain odor sample data; the odor sample data obtained by G1 and G2 are recorded as G1 sample and G2 sample, respectively; 3) Using the odor sample data collected in step 2.4), the particle swarm algorithm is used to optimize the sensor array to reduce the number of sensors and achieve miniaturization of the electronic nose, including the following steps: 3.1) For the raw data obtained by the N gas sensor array I (2), the discrete wavelet transform method of the db1-based wavelet is used to reconstruct the original breath odor sample data and remove the signal noise. The preprocessed signal is expressed as: DWT(u,w)=<x(t),ψ u,w (T)> Where: x(t) is the original signal, u is the scale parameter, w is the shift parameter, ψ is the mother wavelet; the processed data is denoted as di, where: i = 1, 2, 3...100T1, and the preprocessed sample data are denoted as g1 samples and g2 samples respectively; 3.2) Two methods are used to extract features from the preprocessed and denoised signals. The first method extracts features directly from the raw sensor response, while the second method extracts features after Fourier transform. Using the first method, five features are extracted from each sensor response signal: mean, standard deviation, kurtosis, skewness, and waveform factor. The formula for calculating the average value is: Standard deviation calculation formula: Kurtosis index calculation formula: Skewness calculation formula: Form factor calculation formula: 3.3) Use three ranking algorithms: XGBoost, LightGBM, and ExtraTrees to rank the feature importance and select the most promising features to form a feature subset; 3.4) Using particle swarm optimization for feature selection, including the following steps: 3.4.1) Set the population size v and randomly initialize v particles to form the initialization population X = {X1, X2, ..., Xv}; 3.4.2) Set the number of iterations and set the current number of iterations T = 1; 3.4.3) Update the velocity vector of each individual: V id t+1 =w*V id t +c1*r 1i *(P id t -X id t )+c2*r 2i *(P gd t -X id t ) Where w represents the inertia weight, c1 and c2 are acceleration constants, r1 and r2 are random numbers uniformly distributed on [0,1], and V id t and X id t represents the velocity and position of the particle in d dimension at time t, P id t and P gd t Represents the local optimal position vector Pbesti of the individual particle in d dimension at time t and the global optimal position vector Gbest of the population; 3.4.4) Update the position vector of each individual: X id t+1 =X id t +V id t+1 3.4.5) Update local and global position vectors: Update the individual's local optimal position vector Pbesti and the population's global optimal position vector Gbest. Use the fitness function to find the optimal position Pbesti of each particle after i iterations, and compare it with other particles to obtain the global optimal position Gbest. 3.4.6) Perform termination condition judgment. If the number of iterations reaches the number of iterations T max , then output Gbest; if not satisfied, return to step 3.4.3) and continue iterating; 3.5) After performing noise reduction, feature extraction, feature sorting, and feature selection on the collected odor sample data in steps 3.2), 3.3, and 3.4), determine n gas sensors to be included in the sensor array of the miniaturized electronic nose (B) based on the perspective of higher diabetes screening accuracy; 4) using the random forest algorithm to train the model of the n miniaturized electronic nose sensors selected in step 3); 5) Building a cloud-based initial screening system; combining the machine learning model obtained in the above steps with cloud deployment, deploying a cloud-based initial screening system consisting of a model training end and a data prediction end; deploying a miniaturized electronic nose (B) in the hospital and in the vehicle's cabin, respectively; using the electronic nose deployed in the hospital to collect respiratory gas samples and disease conditions from patients, transmitting them to the model training end for data processing and model training; using the in-vehicle electronic nose to obtain odor sample data from the vehicle cabin, and transmitting the data to the cloud-based data prediction end; the data prediction end uses the trained model to predict diabetes based on the odor sample data, and sends the prediction results to the bound person's mobile phone and the built-in software on the vehicle's central control screen for health reminders; 6) Build the overall architecture of the cloud platform; the cloud platform includes data sharing, data training, and model usage. In the data sharing part, after obtaining permission authentication, the client will encrypt the data request command and send it to the cloud server. The cloud server will decrypt it and send the required data to the client for data sharing. In the data training part, relevant patient data is uploaded, and the recognition accuracy of the model is improved through desensitized storage and model training of the uploaded data; The vehicle-mounted electronic nose detection terminal uses the model usage part. After uploading the data to the detection terminal, it uses the trained model to perform detection and judgment, and finally sends the detection results to the smart mobile terminal to complete the real-time notification of the detection results. 7) Mounting a miniaturized electronic nose (B) on a car to obtain cabin gas data for diabetes prediction includes the following steps: 7.1) The inlet of the miniaturized electronic nose (B) is placed at the driver's seat in the vehicle cabin, and the detected gas is discharged through the outlet pipe II (14); the odor sample data collected by the wireless data acquisition card (17) is directly uploaded to the cloud server; 7.2) After the driver enters the vehicle and fastens his seat belt, the driver's exhaled breath is collected through the intake pipe II (10) placed on the driver's seat in the vehicle cabin. The odor sample data is collected by the wireless data acquisition card (17) and uploaded to the cloud server; the data is processed in the cloud, and the transmitted data is judged using a machine learning model. Finally, the diagnosis result is encrypted and transmitted to the corresponding smart mobile terminal, completing the initial diabetes screening and result notification; in the cabin, if the driver does not detect diabetes, the car starts normally; if diabetes is detected, a reminder is issued and the car cannot be started temporarily.

Images