Artificial intelligence-based respiratory dyspnea degree quantitative evaluation method and system, and medium

Abstract

The invention discloses a breathing difficulty degree quantitative evaluation method and system based on artificial intelligence and a medium. The method comprises the steps that breathing sound signals, basic information and medical history information of a wearer are obtained; extracting time domain features and frequency domain features in the breath sound signals based on a preset machine learning model, and analyzing and evaluating in combination with the basic information and the medical history information to obtain a current breath condition level; according to the historical breathing data and the current breathing condition grade of the wearer, whether abnormal breathing exists or not is evaluated; if yes, the current breathing condition grade is corrected according to the motion information of the wearer in the period corresponding to the breathing sound signals, and the current dyspnea grade of the wearer is obtained. The breathing sound signals are collected in real time and analyzed and evaluated by the preset machine learning model, and the historical breathing data and the motion information are combined for analysis and correction, so that the accuracy of the evaluated dyspnea grade is improved, and the breathing condition of a wearer in daily life is accurately monitored in real time from multiple dimensions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a method and system for quantitatively evaluating the degree of breathing difficulty based on artificial intelligence and a medium. BACKGROUND

[0002] The rapid growth of industrialization not only promotes the improvement of economic level, but also may cause a series of environmental pollution problems. Although the environmental pollution problem is valued at present, it still inevitably leads to some respiratory diseases, such as asthma, tracheitis, chronic bronchitis, emphysema, etc. The severity of respiratory diseases is usually closely related to the degree of breathing difficulty, and the higher the severity of the disease, the greater the degree of breathing difficulty. Therefore, the change of respiratory diseases can be reflected by monitoring the respiratory condition, especially for the daily life monitoring of patients with mild respiratory symptoms who do not need to be hospitalized for diagnosis and treatment.

[0003] At present, for patients with respiratory diseases who have mild symptoms and do not need to be hospitalized for treatment, or for people who need to monitor the respiratory condition to reflect the physical health condition, there is a lack of an accurate and real-time convenient way to monitor the respiratory condition in daily life. Patients usually fill in such scales as MRC scale, LCADL scale, PFSDQ scale, VAS scale, etc. according to their subjective feelings during diagnosis and treatment or examination, and the degree of breathing difficulty of the patient is reflected by the filled scale. This kind of filling scale method cannot monitor the respiratory condition of the patient in real time in daily life, which may delay the disease, and the accuracy of filling according to subjective feelings is low. Therefore, how to accurately monitor the respiratory condition of people in real time in daily life in order to timely and accurately reflect the degree of breathing difficulty is one of the technical problems to be solved at present. SUMMARY

[0004] The main purpose of the present application is to provide a method and system for quantitatively evaluating the degree of breathing difficulty based on artificial intelligence, and a medium, which aims to solve the technical problem of how to accurately monitor the degree of breathing difficulty of people in real time in daily life in order to timely and accurately reflect the degree of breathing difficulty in the prior art.

[0005] To achieve the above-mentioned purpose, the present application provides a method for quantitatively evaluating the degree of breathing difficulty based on artificial intelligence, which comprises: obtaining the breathing sound signal of the wearer based on the intelligent wearable device, pre-processing the breathing sound signal, and searching for the basic information and medical history information corresponding to the wearer; extracting time domain features and frequency domain features in the respiratory sound signal based on a preset machine learning model, and analyzing and evaluating the basic information, medical history information, time domain features and frequency domain features to obtain a current respiratory condition level corresponding to the wearer; finding historical respiratory data corresponding to the wearer, and evaluating whether the wearer has abnormal breathing according to the historical respiratory data and the current respiratory condition level; If there is abnormal breathing, obtaining motion information of the wearer within a period corresponding to the respiratory sound signal, and correcting the current respiratory condition level according to the motion information to obtain a dyspnea level corresponding to the wearer.

[0006] Preferably, the step of extracting time domain features and frequency domain features in the respiratory sound signal based on a preset machine learning model, and analyzing and evaluating the basic information, medical history information, time domain features and frequency domain features to obtain a current respiratory condition level corresponding to the wearer includes: extracting time domain features and frequency domain features in the respiratory sound signal based on an extraction layer of a preset machine learning model, and analyzing the time domain features and the frequency domain based on a general layer of the preset machine learning model to obtain a general analysis result; analyzing the basic information and the medical history information based on a custom layer of a preset machine learning model to obtain a custom analysis result; performing full connection processing on the general analysis result and the custom analysis result based on a full connection layer of a preset machine learning model to generate the current respiratory condition level.

[0007] Preferably, the step of extracting time domain features and frequency domain features in the respiratory sound signal based on a preset machine learning model includes: obtaining a large number of respiratory sound sample data sets from different persons, each of the respiratory sound sample data sets including a sample respiratory sound signal of a person, and sample time domain features, sample frequency domain features and a sample dyspnea level corresponding to the sample respiratory sound signal; training a preset basic model based on the sample respiratory sound signal in each of the respiratory sound sample data sets, and when a training duration reaches a preset single training duration, obtaining training time domain features and training frequency domain features output by a preset extraction layer in the preset basic model, and a training dyspnea level output by a preset general layer in the preset basic model; generating time domain loss value, frequency domain loss value and level loss value according to the sample time domain features, sample frequency domain features and sample dyspnea level, and the training time domain features, training frequency domain features and training dyspnea level, and generating a loss function value of the preset basic model according to the time domain loss value, frequency domain loss value and level loss value; determine whether the preset basic model meets a preset convergence condition based on the time domain loss value, the frequency domain loss value, the grade loss value, and the loss function value, and if the preset basic model meets the preset convergence condition, generate the preset basic model as a general model, and perform individual training on the general model based on the respiratory sound sample data sets to generate the preset machine learning model.

[0008] Preferably, the step of determining whether the preset basic model meets a preset convergence condition based on the time domain loss value, the frequency domain loss value, the grade loss value, and the loss function value comprises: If the preset basic model does not meet the preset convergence condition, update the correlation coefficients in the preset basic model corresponding to the training time domain features, the training frequency domain features, and the training dyspnea grade based on a preset update formula, the time domain loss value, the frequency domain loss value, the grade loss value, and the loss function value. adjust the time domain weight and the frequency domain weight of the preset extraction layer in the preset basic model and adjust the grade weight of the preset general layer in the preset basic model based on the correlation coefficients, and perform the step of training the preset basic model based on the sample respiratory sound signals in each of the respiratory sound sample data sets on the adjusted preset basic model.

[0009] Preferably, the step of performing individual training on the general model based on the respiratory sound sample data sets to generate the preset machine learning model comprises: generate a plurality of random numbers based on a preset random number algorithm and each of the respiratory sound sample data sets, and select random data sets corresponding to each of the random numbers from each of the respiratory sound sample data sets, and find individual data of the person corresponding to each of the random data sets, the individual data at least including gender, age, weight, height, medical history, family history, living habit information corresponding to breathing, and working condition information. freeze the preset general layer of the general model, and add a custom layer to the general model. generate the gender, the age, the weight, the height, the medical history, the family history, the living habit information, and the working condition information in each of the individual data as a feature vector, and perform individual training on the general model with the frozen preset general layer based on each of the feature vectors to obtain an individual output value corresponding to each of the individual data. determine whether each of the individual output values matches the dyspnea grade corresponding to each of the individual data, and if the individual output values match the dyspnea grade, complete the individual training of the general model, and generate the general model as the preset machine learning model.

[0010] Preferably, the step of judging whether the individual output value matches the dyspnea level corresponding to the individual data comprises: Based on the individual data, a corresponding relationship between the individual output value and the dyspnea level is established, and it is judged whether there is a mismatched corresponding relationship in the corresponding relationship; If there is a mismatched corresponding relationship, the number of mismatched corresponding relationships is counted to generate an abnormal relationship number, and it is judged whether the abnormal relationship number is less than or equal to a preset relationship threshold; If it is less than or equal to the preset relationship threshold, it is judged that the individual output value matches the dyspnea level corresponding to the individual data; If the abnormal relationship number is greater than the preset relationship threshold, it is judged that the individual output value does not match the dyspnea level corresponding to the individual data, and the weight value of the self-defined layer is updated based on the mismatched corresponding relationship.

[0011] Preferably, the step of generating a plurality of random numbers based on a preset random number algorithm and the respiratory sound sample data set comprises: A first number of the respiratory sound sample data set is determined, and a second number of the random numbers is determined according to the first number and a preset ratio relationship; The first number and the second number are calculated based on the preset random number algorithm to obtain a calculation result; It is judged whether the calculation result is greater than the first number, and if it is greater than the first number, the calculation result is subtracted from the first number to obtain a difference operation result as a random number; If the calculation result is less than or equal to the first number, the calculation result is taken as a random number; The number of random numbers is counted, and when the number of random numbers reaches the second number, the generation of the random numbers is completed.

[0012] Preferably, the step of obtaining the respiratory sound signal of the wearer based on the smart wearable device comprises: The device serial number and location information of the smart wearable device are obtained, and a key pair is generated based on a preset key generation algorithm by calculating the device serial number and the location information, the key pair comprising a decryption private key and an encryption public key; The key pair is subjected to a simulation attack to obtain an attack result, and it is judged whether the attack result meets a preset result condition, and if it meets the preset result condition, the encryption public key is sent to the smart wearable device; The ciphertext information encrypted by the smart wearable device based on the encryption key is received, and the ciphertext information is decrypted based on the decryption private key to obtain the respiratory sound signal.

[0013] Further, to achieve the above object, the present application also provides a respiratory difficulty degree quantification evaluation system based on artificial intelligence, which comprises a storage, a processor, a communication bus and a control program stored on the storage: The communication bus is used to realize the connection communication between the processor and the storage. The processor is used to execute the control program to realize the steps of the respiratory difficulty degree quantification evaluation method based on artificial intelligence as described above.

[0014] Further, to achieve the above object, the present application also provides a storage medium, which stores a control program, and the control program is executed by a processor to realize the steps of the respiratory difficulty degree quantification evaluation method based on artificial intelligence as described above.

[0015] The respiratory difficulty degree quantification evaluation method, system and medium based on artificial intelligence have the following advantages: the people who have the need of monitoring the daily respiratory condition wear the intelligent wearable device, the respiratory sound signal of the wearer is acquired and pretreated through the intelligent wearable device, and the basic information and medical history information of the wearer are searched; then the time domain features and frequency domain features in the respiratory sound signal are extracted according to the preset machine learning model formed by a large amount of data training, and the extracted time domain features and frequency domain features are analyzed and evaluated in combination with the basic information and medical history information to obtain the current respiratory condition grade of the wearer; then the historical respiratory data obtained by the past monitoring of the wearer is searched, and whether the wearer has abnormal breathing is evaluated according to the searched historical respiratory data and the current respiratory condition grade. If there is abnormal breathing, the motion information of the wearer in the period corresponding to the respiratory sound signal is acquired, and then the current respiratory condition grade evaluated as abnormal is corrected in combination with the body function condition of the wearer reflected by the motion information to obtain the accurate respiratory difficulty grade of the wearer. In this way, the respiratory sound signal is collected in real time through the intelligent wearable device, the time domain features and frequency domain features are extracted therefrom by the preset machine learning model for analysis and evaluation, and the historical respiratory data and motion information obtained by the past monitoring are analyzed and corrected, which improves the accuracy of the evaluated respiratory difficulty grade, realizes the real-time and accurate monitoring of the respiratory condition of the wearer in daily life from multiple dimensions, and facilitates the timely and accurate reflection of the respiratory difficulty degree of the wearer in daily life. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the first embodiment of the respiratory difficulty degree quantification evaluation method based on artificial intelligence of the present application; Figure 2 The flowchart of the second embodiment of the respiratory difficulty degree quantification evaluation method based on artificial intelligence of the present application; Figure 3 A flowchart of a third embodiment of the method for quantitatively evaluating the degree of dyspnea based on artificial intelligence of the present application; Figure 4 A structural diagram of a hardware operating environment involved in an embodiment of the system for quantitatively evaluating the degree of dyspnea based on artificial intelligence of the present application.

[0017] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0018] It should be understood that the specific embodiments described herein are merely intended to explain the present application and not to limit the present application.

[0019] The present application provides a method for quantitatively evaluating the degree of dyspnea based on artificial intelligence, which will be described below with reference to Figure 1 , Figure 1 A flowchart of a first embodiment of the method for quantitatively evaluating the degree of dyspnea based on artificial intelligence of the present application.

[0020] The embodiments of the method for quantitatively evaluating the degree of dyspnea based on artificial intelligence provided by the present application need to be explained that although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be performed in an order different from that shown herein. Specifically, the method for quantitatively evaluating the degree of dyspnea based on artificial intelligence in the present embodiment includes: Step S10, obtaining the breathing sound signal of the wearer based on the intelligent wearable device, pre-processing the breathing sound signal, and searching for the basic information and medical history information corresponding to the wearer.

[0021] The method for quantitatively evaluating the degree of dyspnea based on artificial intelligence in the present embodiment is applied to a system server. The system server supports communication connection with an intelligent wearable device for collecting breathing sound signals. The intelligent wearable device can be a smart watch, a smart bracelet, a smart ring, etc. People with daily respiratory monitoring needs wear such intelligent wearable devices, collect breathing sound signals reflecting the breathing condition of the wearer through such intelligent wearable devices, and transmit the collected breathing sound signals to the system server for pre-processing. The pre-processing at least includes data cleaning processing and data normalization processing. The data cleaning processing is used to denoise the breathing sound signal and remove abnormal signals and fill in missing values therein. The data normalization processing is used to scale the breathing sound signal to a specific data interval range, for example, 0~1, according to a certain proportion, so as to improve the efficiency and accuracy of signal processing.

[0022] Further, in order to reflect the differences between individuals, the system server also searches for the basic information and medical history information of the wearer. The basic information at least includes individual information such as gender, age, weight, height, and habits and working conditions related to respiratory conditions, such as smoking, alcoholism, and dusty working environment. The medical history information at least includes the past medical history and family medical history of the wearer. The basic information and medical history information can be uploaded by the wearer to the system server through the smart wearable device when the smart wearable device is used for the first time, and stored in the database connected to the system server. During the daily monitoring of the respiratory condition, the information can be obtained from the database.

[0023] In step S20, the time domain features and frequency domain features in the respiratory sound signal are extracted based on the preset machine learning model, and the basic information, medical history information, time domain features and frequency domain features are analyzed and evaluated to obtain the current respiratory condition level corresponding to the wearer.

[0024] Further, the system server has a preset machine learning model trained by a large number of respiratory sound signals. The trained preset machine learning model can effectively extract and analyze the time domain features and frequency domain features in the respiratory sound signal. Therefore, the time domain features and frequency domain features in the preprocessed respiratory sound signal can be extracted by the preset machine learning model. The extracted time domain features at least include features reflecting the amplitude, time characteristics, etc. of the respiratory sound signal, and the energy of each frame of the extracted signal to reflect the intensity of the signal, and / or the frequency of the signal waveform crossing zero to reflect the frequency characteristics of the signal, and / or the mean and variance of the signal to reflect the central tendency and dispersion of the signal, etc. The extraction of the frequency domain features is to convert the respiratory sound signal from the time domain to the frequency domain by Fourier transform, and extract features reflecting the frequency distribution and spectral characteristics of the signal; for example, by extracting the frequency power of the signal to reflect the energy distribution of different frequency components, and extracting the energy in a specific frequency band, such as low frequency band energy and high frequency band energy, to reflect the frequency distribution of the signal.

[0025] Further, in order to accurately reflect the different respiratory conditions of the wearer due to individual differences, the preset machine learning model combines the basic information and medical history information of the wearer to analyze and evaluate the extracted time domain features and frequency domain features to obtain the current respiratory condition level of the wearer. The current respiratory condition level reflects whether the wearer is currently in a difficult breathing state and the degree of difficulty breathing. The trained preset machine learning model can evaluate multiple levels of different degrees of difficulty breathing, such as 0 level representing normal non-difficult breathing, 1 level representing mild difficulty breathing, 2 level representing moderate difficulty breathing, 3 level representing severe difficulty breathing, and 4 level representing extremely severe difficulty breathing, etc.

[0026] Specifically, the step of extracting the time domain features and the frequency domain features in the respiratory sound signal based on the preset machine learning model, and analyzing and evaluating the basic information, the medical history information, the time domain features and the frequency domain features to obtain the current respiratory condition level corresponding to the wearer includes: In step S21, the extraction layer of the preset machine learning model extracts the time domain features and the frequency domain features in the respiratory sound signal, and analyzes the time domain features and the frequency domain based on the general layer of the preset machine learning model to obtain a general analysis result. In step S22, the custom layer of the preset machine learning model analyzes the basic information and the medical history information to obtain a custom analysis result. In step S23, the fully connected layer of the preset machine learning model fully connects the general analysis result and the custom analysis result to generate the current respiratory condition level.

[0027] Further, the preset machine learning model is trained to have an extraction layer, a general layer, a custom layer and a fully connected layer, and the data information processed by different layers is different. Specifically, the respiratory sound signal is first analyzed by the extraction layer to extract the time domain features and the frequency domain features therein. Then, the extracted time domain features and frequency domain features are taken as the input of the general layer, and the general analysis result is obtained by processing of the general layer. The analysis result is a manifestation of the respiratory condition of the respiratory sound signal itself, and does not reflect the individual differences between different wearers. In addition, the basic information and the medical history information are analyzed by the custom layer to obtain a custom analysis result. The custom analysis result reflects the respiratory condition that the wearer with various basic information and medical history information should have. Then, the fully connected layer combines the general analysis result of the general layer and the custom analysis result of the custom layer for fully connected processing to generate the current respiratory condition level of the wearer. The current respiratory condition level combines the information of the wearer's age, gender, gender, weight, height, living habit, working condition, past medical history, genetic history, etc., and accurately reflects the current respiratory condition of the wearer.

[0028] In step S30, historical respiratory data corresponding to the wearer are searched, and whether the wearer has an abnormal respiration is evaluated according to the historical respiratory data and the current respiratory condition level.

[0029] Further, in order to reflect the change of the wearer's breathing condition over time, historical breathing data of the wearer is also searched. The historical breathing data is a plurality of continuous historical breathing condition levels and respective corresponding historical breathing sound signals generated adjacent to the generation of the current breathing condition level, for example, ten continuous historical breathing condition levels and corresponding historical breathing sound signals generated before the current breathing condition level. Through the historical breathing data and the breathing change condition reflected by the current breathing condition level, it is evaluated whether the wearer has abnormal breathing. If the current breathing condition level is not much different from the historical breathing data, there is no sudden change in the breathing condition, and it is evaluated that the wearer does not have abnormal breathing. On the contrary, if the current breathing condition level is greatly different from the historical breathing data, there is a sudden change in the breathing condition, and it is evaluated that the wearer has abnormal breathing.

[0030] In addition, in order to avoid the case of determining error caused by continuous abnormal breathing, a warning mark can be set. Specifically, for the case that the current breathing condition level is not much different from the previous ten historical breathing data, it is possible that the eleventh historical breathing condition level is abnormal, which lasts for eleven historical breathing condition levels until the current, that is, the wearer has a prolonged breathing abnormality. At this time, instead of directly determining that the wearer does not have abnormal breathing according to the current breathing condition level being not much different from the historical breathing data, it is identified whether there is a warning mark that has not been eliminated. If there is a warning mark that has not been eliminated, it means that the wearer has already appeared breathing abnormality before, and the breathing abnormality lasts until the current, so it is evaluated that the wearer has abnormal breathing, and the related warning information is continuously output to remind the wearer of the abnormal breathing. When the wearer's breathing is monitored to return to its corresponding normal state, the warning mark is eliminated, and the next round of breathing abnormality evaluation is performed.

[0031] In step S40, if there is abnormal breathing, the motion information of the wearer in the period corresponding to the breathing sound signal is obtained, and the current breathing condition level is corrected according to the motion information to obtain the breathing difficulty level corresponding to the wearer.

[0032] Understandably, the abnormal breathing of the wearer can be caused by exercise, i.e. the wearer has performed strenuous exercise resulting in rapid breathing and breathing difficulty, rather than respiratory diseases. In order to avoid the influence of such factors on the accuracy of the assessment of the degree of breathing difficulty, the embodiment further acquires the motion information of the wearer in the period corresponding to the breathing sound signal after assessing that the wearer has abnormal breathing. Specifically, the motion information can be acquired by monitoring the heart rate of the wearer in the period corresponding to the breathing sound signal through the smart wearable device. The motion information reflects whether the wearer has performed exercise in the period and the intensity of the exercise. The greater the intensity, the more likely to cause abnormal breathing. Therefore, the current breathing condition level assessed as abnormal can be corrected through the motion information to obtain the breathing difficulty level accurately reflecting the abnormality of the breathing of the wearer.

[0033] Specifically, the breathing condition levels corresponding to various heart rate intervals can be pre-set. If the current breathing condition level matches the heart rate interval corresponding to the motion information, it indicates that the current breathing condition assessed as abnormal is caused by exercise, so the current breathing condition level of the wearer is corrected and adjusted to normal, or the abnormal level is adjusted lower, e.g. from severe to moderate, in combination with the basic information and medical history information of the wearer. The level of the current breathing condition level after the correction and adjustment, i.e. the breathing difficulty level, accurately reflects the current breathing condition of the wearer. For the correction and adjustment in combination with the basic information and medical history information, the basic information and medical history information can be adjusted in the breathing condition level corresponding to the heart rate interval corresponding to the motion information. For the situation that the current breathing condition level does not match the heart rate interval corresponding to the motion information, and the situation that the motion information reflects that the wearer has not performed exercise, it indicates that the wearer indeed has abnormal breathing at present, and the current breathing condition level assessed as abnormal accurately reflects the abnormality of the breathing of the wearer, so the current breathing condition level does not need to be corrected and adjusted, and can be directly taken as the breathing difficulty level corresponding to the wearer.

[0034] The artificial intelligence-based dyspnea degree quantification evaluation method of the embodiment has a population with a demand for monitoring daily respiratory conditions wearing a smart wearable device, acquires and pre-processes the respiratory sound signals of the wearer through the smart wearable device, and simultaneously searches for the basic information and medical history information of the wearer; then, according to the preset machine learning model formed by training a large amount of data in advance, the time domain features and frequency domain features in the respiratory sound signals are extracted, and the extracted time domain features and frequency domain features are analyzed and evaluated in combination with the basic information and medical history information to obtain the current respiratory condition level of the wearer; then, the historical respiratory data obtained by the wearer in the past is searched, and whether the wearer has abnormal breathing is evaluated according to the searched historical respiratory data and the current respiratory condition level. If there is abnormal breathing, the motion information of the wearer in the period corresponding to the respiratory sound signals is acquired, and then the current respiratory condition level evaluated as abnormal is corrected in combination with the body function condition of the wearer reflected by the motion information to obtain the accurate dyspnea level of the wearer. In this way, the time domain features and frequency domain features in the respiratory sound signals collected in real time by the smart wearable device are extracted by the preset machine learning model for analysis and evaluation, and the historical monitoring respiratory data and motion information are analyzed and corrected, which improves the accuracy of the evaluated dyspnea level, realizes real-time and accurate monitoring of the respiratory condition of the wearer in daily life from multiple dimensions, and facilitates timely and accurate reflection of the dyspnea degree of the wearer in daily life.

[0035] Further, please refer to Figure 2 Based on the first embodiment of the artificial intelligence-based dyspnea degree quantification evaluation method of the present application, the second embodiment of the artificial intelligence-based dyspnea degree quantification evaluation method of the present application is proposed.

[0036] The difference between the second embodiment of the artificial intelligence-based dyspnea degree quantification evaluation method and the first embodiment of the artificial intelligence-based dyspnea degree quantification evaluation method is that the step of extracting the time domain features and frequency domain features in the respiratory sound signals based on the preset machine learning model includes: Step S50, a large number of respiratory sound sample data groups derived from different persons are acquired, each of the respiratory sound sample data groups includes a sample respiratory sound signal of a person, and sample time domain features, sample frequency domain features and sample dyspnea levels corresponding to the sample respiratory sound signal; Step S60, a preset basic model is trained based on the sample respiratory sound signals in each of the respiratory sound sample data groups, and when the training duration reaches a preset single training duration, the training time domain features and training frequency domain features output by a preset extraction layer in the preset basic model, and the training dyspnea levels output by a preset general layer in the preset basic model are acquired; Step S70, according to the sample time domain feature, sample frequency domain feature and sample dyspnea level, and the training time domain feature, training frequency domain feature and training dyspnea level, generate time domain loss value, frequency domain loss value and level loss value, and generate the loss function value of the preset basic model according to the time domain loss value, frequency domain loss value and level loss value; Step S80, based on the time domain loss value, frequency domain loss value, level loss value and loss function value, judge whether the preset basic model meets the preset convergence condition, if the preset convergence condition is met, the preset basic model is generated as a general model, and the general model is individualized trained according to the respiratory sound sample data set, and the preset machine learning model is generated.

[0037] Further, for the training of the preset machine learning model, a large number of respiratory sound sample data sets are acquired, and the acquired respiratory sound sample data sets are derived from different persons to reflect the individual differences of different persons. And each group of acquired respiratory sound sample data set contains a sample respiratory sound signal of a person, and the sample time domain feature, sample frequency domain feature and sample dyspnea level corresponding to the sample respiratory sound signal. At the same time, a preset basic model containing a preset extraction layer and a preset general layer is set in advance, and the preset basic model is trained by the sample respiratory sound signal in each acquired respiratory sound sample data set. The training time domain feature and training frequency domain feature of each sample respiratory sound signal are extracted by the preset extraction layer of the preset basic model, and the extracted training time domain feature and training frequency domain feature are taken as the input of the preset general layer in the preset basic model, and the training dyspnea feature is obtained by the analysis and evaluation of the preset general layer.

[0038] Further, the training duration is counted, and the counted training duration is compared with a preset single training duration to determine whether the training duration reaches the preset single training duration. If the preset single training duration is reached, the training time-domain feature and the training frequency-domain feature output by the preset extraction layer in the preset basic model are extracted, and the training dyspnea level output by the preset general layer is obtained. Then, a time-domain loss value is generated according to the difference between the sample time-domain features in each sample data set and the training time-domain feature, a frequency-domain loss value is generated according to the difference between the sample frequency-domain features and the training frequency-domain feature, and a level loss value is generated according to the difference between the sample dyspnea levels and the training dyspnea level. The time-domain loss value, the frequency-domain loss value and the level loss value can be calculated according to a preset loss function, for example, a cross-entropy loss function or a mean square error function. Moreover, the sample time-domain features, the sample frequency-domain features and the sample dyspnea levels corresponding to the sample respiratory sound signals reflect the accurate characteristics of the sample respiratory sound signals in the time domain, the frequency domain and the dyspnea level. The differences between the training time-domain feature, the training frequency-domain feature and the training dyspnea level and each of them reflect the accuracy of the extracted time-domain feature and the frequency-domain feature of the preset basic model, and the accuracy of the evaluated dyspnea level. The greater the loss value, the greater the corresponding error, and the less accurate the extracted time-domain feature, the frequency-domain feature or the evaluated dyspnea level of the preset basic model.

[0039] Further, in order to reflect the overall performance of the preset basic model, the time-domain loss value, the frequency-domain loss value and the level loss value are generated as the loss function value of the preset basic model. According to the influence of the time-domain feature, the frequency-domain feature and the dyspnea level on the overall performance of the model, the respective weights are set, the time-domain loss value, the frequency-domain loss value and the level loss value are multiplied by the respective weights, and the sum of the operation results is generated as the loss function value reflecting the overall performance of the preset basic model.

[0040] Further, according to the generated time domain loss value, frequency domain loss value, grade loss value and loss function value, it is judged whether the preset basic model meets a preset convergence condition, the preset convergence condition being an index representing that the preset basic model converges. Wherein, the preset basic model converges means that the preset basic model can accurately and stably extract the time domain feature, the frequency domain feature, and accurately and stably evaluate the dyspnea grade, and the corresponding index is that the continuously generated loss values are all smaller. Therefore, the preset loss threshold value representing that the respective loss value is smaller is preset, and it is judged whether the time domain loss value is smaller than the corresponding preset loss threshold value, whether the frequency domain loss value is smaller than the corresponding preset loss threshold value, whether the grade loss value is smaller than the corresponding preset loss threshold value, and whether the loss function value is smaller than the corresponding preset loss threshold value.

[0041] Further, if the time domain loss value, the frequency domain loss value, the grade loss value and the loss function value are all smaller than the respective corresponding preset loss threshold value, it is judged whether the number of times of being smaller than the respective corresponding preset loss threshold value continuously reaches a preset number of times, and if the preset number of times is continuously reached, it means that the evaluation accuracy of the preset basic model is high and stable. Therefore, it is determined that the preset basic model meets the preset convergence condition. On the contrary, if any one of the time domain loss value, the frequency domain loss value, the grade loss value and the loss function value is not smaller than the corresponding preset loss threshold value, or any one of them does not continuously reach the preset number of times, it is determined that the preset basic model does not meet the preset convergence condition, and the preset basic model needs to be iteratively trained. Specifically, the step of judging whether the preset basic model meets the preset convergence condition based on the time domain loss value, the frequency domain loss value, the grade loss value and the loss function value includes: Step S81, if the preset basic model does not meet the preset convergence condition, the correlation coefficients in the preset basic model corresponding to the training time domain feature, the training frequency domain feature and the training dyspnea grade are updated based on a preset update formula, the time domain loss value, the frequency domain loss value, the grade loss value and the loss function value. Step S82, the time domain weight and the frequency domain weight of the preset extraction layer in the preset basic model are adjusted based on the correlation coefficients, the grade weight of the preset general layer in the preset basic model is adjusted, and the preset basic model after adjustment is used to execute the step of training the preset basic model based on the sample respiratory sound signals in each respiratory sound sample data group.

[0042] Furthermore, if the preset base model is determined not to have reached the preset convergence condition, the time-domain loss value, frequency-domain loss value, level loss value, and loss function value are calculated according to the preset update formula. This calculation updates the correlation coefficients in the preset base model corresponding to training time-domain features, training frequency-domain features, and training dyspnea levels. These correlation coefficients reflect the relationship between the accuracy of extracting training time-domain features, the accuracy of extracting training frequency-domain features, and the accuracy of assessing training dyspnea levels in the preset base model. The higher the accuracy of the extracted training time-domain and training frequency-domain features, the more accurate the assessed training dyspnea level. Specifically... Time Domain The loss values ​​are reflected by frequency domain loss, level loss, and loss function. The calculation formula can be found in the following formula (1).

[0043] (1); Where k represents the updated correlation coefficient, ep represents the loss function value, ew, eu, and ex represent the time-domain loss, frequency-domain loss, and rank loss values, respectively, ew0, eu0, and ex0 represent the previous time-domain loss, previous frequency-domain loss, and previous rank loss values ​​generated during the previous training, respectively, and k0 represents the correlation coefficient before the update. The loss function value ep is negatively correlated with the accuracy of the dyspnea level; the larger the loss function value, the lower the accuracy, and vice versa. Furthermore, testing has shown that converting it to the reciprocal of the natural logarithm more accurately reflects the negative correlation characteristic. middle The ratio of the difference between two consecutive time-domain loss values ​​to the grade loss value indirectly reflects the positive correlation between the change in the time-domain loss value and the accuracy of the dyspnea grade assessment. A smaller difference indicates a smaller change, meaning the preset base model's extraction of time-domain features is closer to convergence, resulting in more accurate extracted time-domain features and thus improved accuracy of dyspnea grade assessment. Conversely, the ratio of the time-domain loss value (ew) to the grade loss value (ex) directly reflects the positive correlation between the time-domain loss value and the accuracy of the dyspnea grade assessment. A smaller time-domain loss value indicates more accurate extracted time-domain features, further improving the accuracy of dyspnea grade assessment. Similarly, middle The difference value of the continuous two frequency domain loss values, the ratio of the frequency domain loss value to the grade loss value, indirectly reflects the positive correlation between the change of the frequency domain loss value and the accuracy of the respiratory difficulty grade, the smaller the difference value, the smaller the change, the closer the extraction of the preset basic model on the frequency domain feature to convergence, the more accurate the extracted frequency domain feature, thereby improving the accuracy of the respiratory difficulty grade evaluation; while the ratio of the frequency domain loss value eu to the grade loss value ex directly reflects the positive correlation between the frequency domain loss value and the accuracy of the respiratory difficulty grade, the smaller the frequency domain loss value, the more accurate the extracted frequency domain feature, which will improve the accuracy of the respiratory difficulty grade evaluation.

[0044] Further, the time domain weight, the frequency domain weight of the extraction layer in the preset basic model, and the grade weight of the general layer are adjusted according to the updated correlation coefficient, and then the sample respiratory sound signals in each respiratory sound sample data set are used to iteratively train the adjusted preset basic model until the preset basic model meets the preset convergence condition. In this way, the correlation coefficient corresponding to the training time domain feature, the training frequency domain feature and the training respiratory difficulty grade in the preset basic model is reflected by the time domain loss value, the frequency domain loss value, the grade loss value and the loss function value, and is updated by the preset update formula set in advance, so that the updated correlation coefficient more accurately reflects the correlation between the time domain feature accuracy, the frequency domain feature accuracy and the respiratory difficulty grade accuracy, which is beneficial to improve the accuracy of the weight adjusted according to the updated correlation coefficient.

[0045] Further, for the case where it is determined that the preset basic model meets the preset convergence condition, it is indicated that the preset basic model can accurately and stably evaluate the respiratory condition reflected by the respiratory sound signal alone, so as to generate the preset basic model as a general model. At the same time, in order to reflect the difference in respiratory conditions before different individuals, the general model is individualized trained in the embodiment, and a preset machine learning model which can evaluate the respiratory condition combined with the basic information and medical history information of the individual is generated by individualized training. Specifically, the step of individualizing training the general model according to the respiratory sound sample data set to generate the preset machine learning model comprises: Step S83, generating a plurality of random numbers based on a preset random number algorithm and each of the respiratory sound sample data sets, and screening a random data set corresponding to each of the random numbers from each of the respiratory sound sample data sets, and finding individual data of the person corresponding to each of the random data sets, the individual data at least including gender, age, weight, height, past medical history, family medical history, and life habit information and working condition information corresponding to breathing; Further, in order to improve the efficiency of model training, the embodiment can extract part of the data set from the respiratory sound sample data set for training. At the same time, in order to ensure the accuracy of the training, part of the data set is randomly extracted by using a random number. Wherein, a preset random number algorithm is set in advance, and a plurality of random numbers are generated by combining each respiratory sound sample data set with the preset random number algorithm. Specifically, the step of generating a plurality of random numbers based on the preset random number algorithm and each respiratory sound sample data set comprises: Step S831, determining the first number of each respiratory sound sample data set, and determining the second number of the random number according to the first number and the preset ratio relationship; Step S832, calculating the first number and the second number based on the preset random number algorithm to obtain a calculation result; Step S833, determining whether the calculation result is greater than the first number, if greater than the first number, then the calculation result and the first number are subtracted to obtain a difference operation result as a random number; Step S834, if the calculation result is less than or equal to the first number, then the calculation result is taken as a random number; Step S835, counting the number of random numbers, and when the number of random numbers reaches the second number, the generation of each random number is completed.

[0046] Further, in order to ensure that the number of data sets extracted from the respiratory sound sample data set is not too small to affect the accuracy of the model, and not too high to affect the efficiency of model training, a reasonable preset ratio relationship is set by testing in advance. The number of respiratory sample data sets is determined as the first number, and the number of random numbers required to be extracted is determined as the second number according to the first number and the preset ratio relationship. Further, the first number and the second number are calculated by the preset random number algorithm to obtain a plurality of calculation results, and the random numbers are generated according to the plurality of calculation results. Wherein, the formula of the preset random number algorithm can be seen in the following formula (2) ; Wherein, Ki represents the i-th calculation result, p represents the first number, q represents the second number, % represents the remainder operation, Ki-1 represents the i-1-th calculation result, i=1, 2, 3…q. By combining the first number and the second number to generate a random number, the correlation between the random number and the number of respiratory sound sample data sets is stronger, and the number of generated random numbers is more reasonable, which is beneficial to the balance of model training efficiency and accuracy.

[0047] In addition, it should be noted that the random number represents the arrangement position of each data group in the respiratory sound sample data group, and therefore the size of the random number cannot exceed the first quantity of the respiratory sound sample data group. Therefore, for each calculation result generated by the preset random number algorithm, the calculation result is compared with the first quantity to determine whether the calculation result is greater than the first quantity. If the calculation result is greater than the first quantity, a difference value operation is performed between the calculation result and the first quantity to obtain a difference value operation result as the random number. If it is determined through comparison that the calculation result is not greater than the first quantity, the calculation result is directly taken as the random number. At the same time, the quantity of the generated random numbers is counted to determine whether the second quantity is reached. If the second quantity is reached, the generation of each random number is completed, and otherwise the generation is continued.

[0048] In order to better illustrate the generation method of the random number, a specific embodiment is described herein. In the specific embodiment, the first quantity of the respiratory sound sample data group is 100, the preset ratio is 1:10, the second quantity of the random number is determined to be 10, and then the calculation result generated by the above formula (2) is 2, 3, 27, 29, 64, 88, 101, 3, 4, and 13, respectively. The seventh random number 101 is greater than the first quantity 100, and therefore the difference value 1 between the two is generated as the seventh random number, so that the random numbers are 2, 3, 27, 29, 64, 88, 1, 3, 4, and 13.

[0049] Further, after each random number is generated, the data group corresponding to each random number is searched from each respiratory sound sample data group as a random data group. For example, the random numbers 2, 3, 27, 29, 64, 88, 1, 3, 4, and 13 are searched from each respiratory sound sample data group to find the data group arranged at the 2nd, 3rd, 27th, 29th, 64th, 88th, 1st, 3rd, 4th, and 13th positions, respectively, as the random data group. At the same time, the individual data of the corresponding person is searched for each random data group. The individual data includes gender, age, weight, height, and individual basic information such as habit information and working condition information corresponding to the respiratory habit, such as smoking, alcoholism, dust working environment, and medical history information such as medical history and family history.

[0050] Step S84, freezing the preset general layer of the general model, and adding a custom layer to the general model; Step S85, generating the gender, age, weight, height, medical history, family history, habit information, and working condition information in each individual data as a feature vector, and individually training the general model with the frozen preset general layer based on each feature vector to obtain an individual output value corresponding to each individual data; Step S86, judging whether the individual output value matches the dyspnea level corresponding to the individual data, if matched, completing the individualization training of the general model, and generating the general model as the preset machine learning model.

[0051] Further, the general layer of the general model is frozen, and a new custom layer is added. The freezing of the general layer is a process of adjusting the weight value, bias and other parameters of the general layer to an unchangeable state, so as to avoid the change of the above-mentioned parameters of the general layer in the individualization training process of the general model, which affects the evaluation performance of the general model on the respiratory sound signal which has been trained. The addition of the custom layer can be to set a standby layer in advance, and in the individualization training process, the standby layer is activated to add the custom layer.

[0052] Further, the gender, age, weight, height, medical history, family history, living habit information and working condition information in each individual data are respectively converted into a single vector, and each single vector is combined to form a feature vector, and then the general model with the frozen general layer is individualized trained according to each feature vector. The general model processes each feature vector to generate an individual output value corresponding to each individual data. The individual output value is the respiratory condition of the wearer with individual data evaluated by the general model. At the same time, each individual data is pre-set with a dyspnea level representing the corresponding respiratory condition. For each individual data, it is judged whether the corresponding individual output value matches the corresponding dyspnea level, if matched, it means that the individual output value evaluated by the general model is accurate, so as to generate the general model as the preset machine learning model. Otherwise, if not matched, the general model is continuously trained according to each feature vector until the generated individual output value matches the dyspnea level.

[0053] Further, the matching between each individual output value and each dyspnea level can be complete matching, or a certain proportion can be set as the matching standard, and the number of individual output values and dyspnea levels that match the proportion is determined as matching, and the number that does not reach the proportion is determined as not matching. Specifically, the step of judging whether the individual output value matches the dyspnea level corresponding to the individual data includes: Step S861, based on each individual data, establishing a corresponding relationship between each individual output value and each dyspnea level, and judging whether there is a mismatched corresponding relationship in each corresponding relationship; Step S862, if there is a mismatched corresponding relationship, counting the number of mismatched corresponding relationships to generate an abnormal relationship number, and judging whether the abnormal relationship number is less than or equal to a preset relationship threshold; Step S863, if less than or equal to the preset relationship threshold, judging whether the individual output value corresponding to each individual data and the dyspnea level match. Step S864, if the number of abnormal relationships is greater than the preset relationship threshold, judging that the individual output value corresponding to each individual data and the dyspnea level do not match, and updating the weight value of the self-defined layer based on the corresponding relationship that does not match.

[0054] Further, for each individual data, the individual output value generated by processing the individual data through the general model is established in correspondence with the corresponding dyspnea level, and for each corresponding relationship, it is judged whether the individual output value and the dyspnea level match. Whether the corresponding relationship exists is determined by the judgment of whether the corresponding relationship matches. As long as there is any one corresponding relationship in which the individual output value and the corresponding dyspnea level do not match in all corresponding relationships, it can be determined that there is a corresponding relationship that does not match in each corresponding relationship. Conversely, if the individual output value in all corresponding relationships matches the corresponding dyspnea level, it is determined that there is no corresponding relationship that does not match in each corresponding relationship. For the case where there is a corresponding relationship that does not match, the number of corresponding relationships that do not match in all corresponding relationships is counted to obtain the number of abnormal relationships.

[0055] Further, in order to reflect the performance of the general model through the number of corresponding relationships that do not match, a preset relationship threshold is set in advance. The number of abnormal relationships is compared with the preset relationship threshold to determine whether the number of abnormal relationships is less than or equal to the preset relationship threshold. If it is less than or equal to the preset relationship threshold, it means that the number of corresponding relationships that do not match is small, so that the individual output value corresponding to each individual data and the dyspnea level match, indicating that the self-defined layer of the general model has good processing performance, and it can be generated as a preset machine learning model. Conversely, if it is determined by comparison that the number of abnormal relationships is greater than the preset relationship threshold, it means that the number of corresponding relationships that do not match is large, so that the individual output value corresponding to each individual data and the dyspnea level do not match, indicating that the processing performance of the self-defined layer of the general model is poor, and it needs to be continuously trained until the number of corresponding relationships that do not match is reduced, and the individual output value corresponding to each individual data and the dyspnea level match.

[0056] In the case of continuous training, in order to improve the training efficiency and make the self-defined layer of the general model quickly obtain good processing performance, the embodiment updates the weight value of the self-defined layer according to the corresponding relationship that does not match. The specific updating formula can be seen in the following formula (3).

[0057] (3) ; wherein, represents the updated weight value, n represents the number of corresponding relationships, n1 represents the number of abnormal relationships, yi represents the individual output value in the i-th unmatched corresponding relationship, and yi0 represents the dyspnea level in the i-th unmatched corresponding relationship, represents the weight value before updating, yt represents a change trend value of the historical cumulative error generated by the individualized training, and st represents a weight offset factor corresponding to the change trend value. The change trend value of the historical cumulative error can be generated by the size change trend of the historical cumulative errors of the previous training adjacent to the current individualized training, for example, the historical cumulative errors of the previous 3 times. If the first error difference between the historical cumulative error of the previous 1 time and the historical cumulative error of the previous 2 times becomes smaller relative to the second error difference between the historical cumulative error of the previous 2 times and the historical cumulative error of the previous 3 times, it indicates that the size change trend of the historical cumulative error is smaller, and the updating of the weight value of the self-defined layer is effective. At this time, the updating can be continued in the updated manner, and the updating can be strengthened by the weight offset factor to quickly reduce the processing error of the self-defined layer of the general model. On the contrary, if the first error difference becomes larger relative to the second error difference, it indicates that the size change trend of the historical cumulative error is larger, and the updating of the weight value of the self-defined layer is biased. At this time, the updating can be weakened by the weight offset factor. For the strengthening or weakening effect of the weight offset factor, a corresponding relationship between the change trend value and the weight offset factor can be established in advance. After the change trend value is determined by each historical cumulative error, the corresponding weight offset factor is determined by the corresponding relationship, and then the weight value of the self-defined layer is updated by the weight offset factor, so that the self-defined layer quickly obtains accurate weight value and improves the training efficiency.

[0058] In the embodiment, the preset machine learning model is divided into general model training and individualized training. The individualized training is performed on part of the data groups selected from the respiratory sound sample data groups used for the general model training, thereby improving the training efficiency of the preset machine learning model, and enabling the preset machine learning model to process the respiratory sound signal in combination with the individualized data, and improving the processing accuracy.

[0059] Further, referring to Figure 3 , based on the first and second embodiments of the artificial intelligence-based dyspnea degree quantification evaluation method, the third embodiment of the artificial intelligence-based dyspnea degree quantification evaluation method is proposed.

[0060] The third embodiment of the artificial intelligence-based dyspnea degree quantification evaluation method is different from the first and second embodiments of the artificial intelligence-based dyspnea degree quantification evaluation method in that the step of acquiring the respiratory sound signal of the wearer based on the intelligent wearable device comprises: In step S11, the device serial number and the location information of the smart wearable device are acquired, and a key pair including a decryption private key and an encryption public key is generated based on a preset key generation algorithm. In step S12, the key pair is subjected to a simulation attack, an attack result is obtained, and it is determined whether the attack result meets a preset result condition. If the attack result meets the preset result condition, the encryption public key is sent to the smart wearable device. In step S13, the ciphertext information encrypted by the smart wearable device based on the encryption key is received, and the ciphertext information is decrypted based on the decryption private key to obtain the respiratory sound signal.

[0061] Understandably, the respiratory sound signal of the wearer belongs to the personal privacy data of the wearer, and in order to avoid the leakage of such data, the respiratory sound signal of the wearer is acquired in an encrypted manner in the embodiment. Specifically, the device serial number and the location information of the smart wearable device are acquired by the system server before the respiratory sound signal of the wearer is acquired based on the smart wearable device, and the acquisition can be achieved by requesting to read the hardware information and the real-time positioning information of the smart wearable device. Moreover, a preset key generation algorithm is pre-set in the system server, and the device serial number and the location information acquired are calculated by the preset key generation algorithm to generate a key pair including a decryption private key and an encryption public key. The specific formula of the preset key generation algorithm can be referred to as formula (4) as follows. (4) ; Wherein, K1 and K2 represent the encryption public key and the decryption private key in the key pair respectively, ID represents the device serial number, La and Lo represent the latitude value and the longitude value in the location information respectively, and e and b represent preset formula parameters which can be generated by pre-experiment. The key pair is generated from the device serial number representing the uniqueness of the smart wearable device and the real-time location information representing the position change of the smart wearable device, so that the key pair has the uniqueness related to the device serial number of the smart wearable device, and also has the dynamic characteristic represented by the real-time location information, and the security strength of the key pair is better.

[0062] Further, in order to further ensure the security of the key pair, the embodiment also simulates an attack on the generated key pair, predicts or infers the key pair through the simulated attack, and takes the predicted or inferred key pair as an attack result. Then, it is judged whether the attack result meets a preset result condition, which is a condition indicating an attack failure. For example, the attack result is inconsistent with the key pair, and the number of inconsistencies reaches a set number of times, such as 5 times, 8 times, etc. If it is determined that the attack result meets the preset result condition, it means that the attack results of the set number of times are all incorrect, and the security of the key pair is high. At this time, the system server sends the encryption public key in the key pair to the smart wearable device, and the smart wearable device encrypts the collected respiratory sound signals according to the encryption public key to generate ciphertext information. Then, the smart wearable device transmits the ciphertext information to the system server based on a communication protocol between the smart wearable device and the system server. The system server receives the ciphertext information and decrypts it according to the decryption private key in the key pair to obtain the respiratory sound signals of the wearer.

[0063] Further, if it is determined that the attack result does not meet the preset result condition, it means that the key pair is correctly predicted or inferred through the simulated attack, and the security of the key pair is low. At this time, the preset key generation algorithm is updated according to a preset update mechanism, such as increasing the number of encryptions based on formula (4) to update. Then, a new key pair is generated based on the updated preset key generation algorithm to perform a simulated attack until the attack result of the simulated attack meets the preset result condition.

[0064] The embodiment generates a key pair through a preset key generation algorithm to encrypt and transmit the respiratory sound signals, in order to avoid leakage of the wearer's private information. The generated key pair is not only unique to the device serial number of the smart wearable device, but also reflects dynamic characteristics through real-time location information of the smart wearable device. At the same time, the security is verified through a simulated attack, which ensures the security strength of the key pair from multiple aspects, thereby ensuring the security of the respiratory sound signal transmission.

[0065] In addition, the embodiment of the present application also provides a respiratory difficulty degree quantification evaluation system based on artificial intelligence. Please refer to Figure 4 , Figure 4 The structure diagram of the hardware running environment of the device involved in the embodiment of the respiratory difficulty degree quantification evaluation system based on artificial intelligence of the present application.

[0066] As Figure 4As shown, the artificial intelligence-based dyspnea degree quantification evaluation system can include a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a storage 1005. The communication bus 1002 is used to realize the connection communication between the components. The user interface 1003 can include a display screen (Display) and an input unit such as a keyboard (Keyboard). The optional user interface 1003 can also include a standard wired interface and a wireless interface. The network interface 1004 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The storage 1005 can be a high-speed RAM storage or a stable storage (non-volatile memory) such as a magnetic disk storage. The storage 1005 can also be a storage device independent of the aforementioned processor 1001.

[0067] Those skilled in the art can understand that Figure 4 The hardware structure of the artificial intelligence-based dyspnea degree quantification evaluation system shown in the figure does not constitute a limitation on the artificial intelligence-based dyspnea degree quantification evaluation system, and can include more or fewer components than the figure, or combine certain components, or different component arrangements.

[0068] As Figure 4 As shown, the storage 1005 as a storage medium can include an operating system, a network communication module, a user interface module, and a control program. The operating system is a program that manages and controls the artificial intelligence-based dyspnea degree quantification evaluation system and software resources, supports the running of the network communication module, the user interface module, the control program, and other programs or software; the network communication module is used to manage and control the network interface 1004; and the user interface module is used to manage and control the user interface 1003.

[0069] In Figure 4 In the hardware structure of the artificial intelligence-based dyspnea degree quantification evaluation system shown in the figure, the network interface 1004 is mainly used to connect other system servers and communicate data with other system servers; the user interface 1003 is mainly used to connect a client (user end) and communicate data with the client; and the processor 1001 can call the control program stored in the storage 1005 and perform the following operations: Obtaining the breathing sound signal of the wearer based on the intelligent wearable device, preprocessing the breathing sound signal, and searching for the basic information and medical history information corresponding to the wearer; Extracting time domain features and frequency domain features in the breathing sound signal based on a preset machine learning model, and analyzing and evaluating the basic information, medical history information, time domain features, and frequency domain features to obtain a current breathing condition level corresponding to the wearer. searching historical breathing data corresponding to the wearer, and evaluating whether the wearer has abnormal breathing according to the historical breathing data and the current breathing condition level; if the wearer has abnormal breathing, obtaining motion information of the wearer in a period corresponding to the respiratory sound signal, and correcting the current breathing condition level according to the motion information to obtain a dyspnea level corresponding to the wearer.

[0070] Further, the step of extracting the time domain features and the frequency domain features in the respiratory sound signal based on the preset machine learning model and analyzing and evaluating the basic information, the medical history information, the time domain features and the frequency domain features to obtain the current breathing condition level corresponding to the wearer comprises: extracting the time domain features and the frequency domain features in the respiratory sound signal based on an extraction layer of a preset machine learning model, and analyzing the time domain features and the frequency domain based on a general layer of the preset machine learning model to obtain a general analysis result; analyzing the basic information and the medical history information based on a self-defined layer of the preset machine learning model to obtain a self-defined analysis result; performing full connection processing on the general analysis result and the self-defined analysis result based on a full connection layer of the preset machine learning model to generate the current breathing condition level.

[0071] Further, before the step of extracting the time domain features and the frequency domain features in the respiratory sound signal based on the preset machine learning model, the processor 1001 can call the control program stored in the storage 1005 and perform the following operations: obtain a large number of respiratory sound sample data sets from different persons, each of the respiratory sound sample data sets comprising a sample respiratory sound signal of a person, and sample time domain features, sample frequency domain features and a sample dyspnea level corresponding to the sample respiratory sound signal; train a preset basic model based on the sample respiratory sound signal in each of the respiratory sound sample data sets, and when the training duration reaches a preset single training duration, obtain training time domain features and training frequency domain features output by a preset extraction layer in the preset basic model, and a training dyspnea level output by a preset general layer in the preset basic model; generate a time domain loss value, a frequency domain loss value and a level loss value according to the sample time domain features, the sample frequency domain features and the sample dyspnea level, and the training time domain features, the training frequency domain features and the training dyspnea level, and generate a loss function value of the preset basic model according to the time domain loss value, the frequency domain loss value and the level loss value; determine whether the preset base model meets a preset convergence condition based on the time domain loss value, the frequency domain loss value, the grade loss value, and the loss function value, and if the preset base model meets the preset convergence condition, generate the preset base model as a general model and perform individual training on the general model based on the respiratory sound sample data sets to generate the preset machine learning model.

[0072] Further, after the step of determining whether the preset base model meets a preset convergence condition based on the time domain loss value, the frequency domain loss value, the grade loss value, and the loss function value, the processor 1001 can call the control program stored in the storage 1005 and perform the following operations: If the preset base model does not meet the preset convergence condition, update the correlation coefficients in the preset base model corresponding to the training time domain features, the training frequency domain features, and the training dyspnea grades based on a preset update formula, the time domain loss value, the frequency domain loss value, the grade loss value, and the loss function value; adjust the time domain weight and the frequency domain weight of the preset extraction layer in the preset base model and adjust the grade weight of the preset general layer in the preset base model based on the correlation coefficients, and perform the step of training the preset base model based on the sample respiratory sound signals in each of the respiratory sound sample data sets on the adjusted preset base model

[0073] Further, the processor 1001 can call the control program stored in the storage 1005 and perform the following operations in the step of performing individual training on the general model based on the respiratory sound sample data sets to generate the preset machine learning model: generate a plurality of random numbers based on a preset random number algorithm and each of the respiratory sound sample data sets, and select random data sets corresponding to each of the random numbers from each of the respiratory sound sample data sets, and find individual data of the person corresponding to each of the random data sets, the individual data at least including gender, age, weight, height, medical history, family history, living habit information corresponding to breathing, and working condition information; freeze the preset general layer of the general model, and add a custom layer to the general model; generate the gender, age, weight, height, medical history, family history, living habit information, and working condition information in each of the individual data as a feature vector, and perform individual training on the general model with the frozen preset general layer based on each of the feature vectors to obtain an individual output value corresponding to each of the individual data; determine whether the individual output value matches the dyspnea level corresponding to the individual data, and if so, complete the individualization training of the general model, and generate the general model as the preset machine learning model.

[0074] Further, the step of determining whether the individual output value matches the dyspnea level corresponding to the individual data, the processor 1001 can call the control program stored in the storage 1005, and perform the following operations: Based on each of the individual data, establish a corresponding relationship between each of the individual output values and each of the dyspnea levels, and determine whether there is a mismatched corresponding relationship in each of the corresponding relationships; If there is a mismatched corresponding relationship, count the number of mismatched corresponding relationships, generate an abnormal relationship number, and determine whether the abnormal relationship number is less than or equal to a preset relationship threshold; If less than or equal to the preset relationship threshold, it is determined that the individual output value matches the dyspnea level corresponding to the individual data; If the abnormal relationship number is greater than the preset relationship threshold, it is determined that the individual output value does not match the dyspnea level corresponding to the individual data, and the weight value of the self-defined layer is updated based on the mismatched corresponding relationship.

[0075] Further, the step of generating a plurality of random numbers based on the preset random number algorithm and each of the respiratory sound sample data sets comprises: Determine a first number of each of the respiratory sound sample data sets, and determine a second number of the random numbers according to the first number and a preset ratio relationship; Calculate the first number and the second number based on the preset random number algorithm to obtain a calculation result; Determine whether the calculation result is greater than the first number, if greater than the first number, difference operation is performed on the calculation result and the first number to obtain a difference operation result as a random number; If the calculation result is less than or equal to the first number, the calculation result is taken as a random number; Count the number of random numbers, and complete the generation of each of the random numbers when the number of random numbers reaches the second number.

[0076] Further, the step of obtaining the respiratory sound signal of the wearer based on the smart wearable device comprises: Obtain the device serial number and location information of the smart wearable device, and generate a key pair based on a preset key generation algorithm to calculate the device serial number and the location information, the key pair including a decryption private key and an encryption public key; Analog attack is performed on the key pair to obtain an attack result, and it is judged whether the attack result meets a preset result condition, and if yes, the encryption public key is sent to the smart wearable device; The ciphertext information encrypted by the smart wearable device based on the encryption key is received, and the ciphertext information is decrypted based on the decryption private key to obtain the respiratory sound signal.

[0077] The specific implementation of the artificial intelligence-based respiratory difficulty degree quantification evaluation system of the present application is basically the same as the above-mentioned artificial intelligence-based respiratory difficulty degree quantification evaluation method, and will not be repeated here.

[0078] The present application also provides a storage medium. The storage medium stores a control program, and the control program is executed by a processor to realize the steps of the above-mentioned artificial intelligence-based respiratory difficulty degree quantification evaluation method.

[0079] The present application also provides a storage medium. The storage medium stores a control program, and the control program is executed by a processor to realize the steps of the above-mentioned artificial intelligence-based respiratory difficulty degree quantification evaluation method.

[0080] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative and not limiting, and those skilled in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims. Any equivalent structure or equivalent flow conversion made by using the content of the present application specification and drawings, or directly or indirectly applied to other related technical fields, all belong to the protection of the present application.

Claims

1. A method for quantitatively assessing the degree of dyspnea based on artificial intelligence, characterized in that, The quantitative assessment method for the degree of dyspnea includes: The device acquires the wearer's breathing sound signal based on a smart wearable device, preprocesses the breathing sound signal, and searches for the wearer's basic information and medical history information. The time-domain and frequency-domain features of the respiratory sound signal are extracted based on a preset machine learning model, and the basic information, medical history information, time-domain features and frequency-domain features are analyzed and evaluated to obtain the current respiratory status level corresponding to the wearer. Find the historical breathing data corresponding to the wearer, and assess whether the wearer has abnormal breathing based on the historical breathing data and the current breathing status level; If abnormal breathing is present, the wearer's motion information during the period corresponding to the breathing sound signal is obtained, and the current breathing status level is corrected according to the motion information to obtain the breathing difficulty level corresponding to the wearer.

2. The method for quantitatively assessing the degree of dyspnea as described in claim 1, characterized in that, The step of extracting the time-domain and frequency-domain features from the respiratory sound signal based on a preset machine learning model, and analyzing and evaluating the basic information, medical history information, time-domain features, and frequency-domain features to obtain the current respiratory status level corresponding to the wearer includes: The extraction layer based on the preset machine learning model extracts the time-domain and frequency-domain features from the respiratory sound signal, and the general layer based on the preset machine learning model analyzes the time-domain features and the frequency-domain features to obtain general analysis results; A custom layer based on a preset machine learning model analyzes the basic information and the medical history information to obtain custom analysis results. A fully connected layer based on a preset machine learning model performs fully connected processing on the general analysis results and the custom analysis results to generate the current respiratory status level.

3. The method for quantitatively assessing the degree of dyspnea as described in claim 1, characterized in that, Prior to the step of extracting the time-domain and frequency-domain features from the respiratory sound signal based on a preset machine learning model, the following steps are included: A large number of respiratory sound sample data sets from different people are obtained. Each respiratory sound sample data set includes a sample respiratory sound signal of one person, as well as the sample time domain characteristics, sample frequency domain characteristics, and sample respiratory difficulty level corresponding to the sample respiratory sound signal. The preset basic model is trained based on the sample breathing sound signals in each of the breathing sound sample data groups. When the training time reaches the preset single training time, the training time domain features and training frequency domain features output by the preset extraction layer in the preset basic model, as well as the training breathing difficulty level output by the preset general layer in the preset basic model, are obtained. Based on the sample time-domain features, sample frequency-domain features, and sample dyspnea level, as well as the training time-domain features, training frequency-domain features, and training dyspnea level, a time-domain loss value, a frequency-domain loss value, and a level loss value are generated, and the loss function value of the preset base model is generated based on the time-domain loss value, frequency-domain loss value, and level loss value. Based on the time-domain loss value, frequency-domain loss value, rank loss value, and loss function value, it is determined whether the preset basic model meets the preset convergence condition. If the preset convergence condition is met, the preset basic model is generated into a general model, and the general model is individually trained according to the breathing sound sample data group to generate the preset machine learning model.

4. The method for quantitatively assessing the degree of dyspnea as described in claim 3, characterized in that, The step of determining whether the preset basic model satisfies the preset convergence condition based on the time-domain loss value, frequency-domain loss value, rank loss value, and loss function value includes: If the preset basic model does not meet the preset convergence condition, the correlation coefficients in the preset basic model corresponding to the training time-domain features, training frequency-domain features and training dyspnea level are updated based on the preset update formula, time-domain loss value, frequency-domain loss value, level loss value and loss function value. Based on the correlation coefficient, adjust the time domain weights and frequency domain weights of the preset extraction layer in the preset basic model, and adjust the level weights of the preset general layer in the preset basic model. Then, for the adjusted preset basic model, perform the step of training the preset basic model based on the sample breathing sound signals in each of the breathing sound sample data groups.

5. The method for quantitatively assessing the degree of dyspnea as described in claim 3, characterized in that, The step of training the general model individually based on the respiratory sound sample data set to generate the preset machine learning model includes: Multiple random numbers are generated based on a preset random number algorithm and each of the breathing sound sample data groups. Random data groups corresponding to each of the random numbers are selected from each of the breathing sound sample data groups. Individual data of the person corresponding to each of the random data groups is found. The individual data includes at least gender, age, weight, height, past medical history, family medical history, living habits information and work status information related to breathing. Freeze the preset general layer of the general model and add a custom layer to the general model; The gender, age, weight, height, past medical history, family medical history, lifestyle information, and work status information in each of the individual data items are generated as feature vectors, and the general model of the frozen preset general layer is trained individually based on each of the feature vectors to obtain the individual output value corresponding to each of the individual data items. Determine whether the output value of each individual matches the dyspnea level corresponding to the individual data. If they match, complete the individualized training of the general model and generate the general model as the preset machine learning model.

6. The method for quantitatively assessing the degree of dyspnea as described in claim 5, characterized in that, The step of determining whether the output value of each individual matches the dyspnea level corresponding to the individual data includes: Based on the individual data, establish a correspondence between the individual output value and the dyspnea level, and determine whether there are any mismatched correspondences among the correspondences; If there are mismatched correspondences, the number of mismatched correspondences is counted to generate an abnormal relationship count, and it is determined whether the abnormal relationship count is less than or equal to a preset relationship threshold. If the value is less than or equal to the preset relationship threshold, then it is determined that the output value of each individual matches the dyspnea level corresponding to the data of each individual. If the number of abnormal relationships is greater than a preset relationship threshold, it is determined that the output value of each individual does not match the dyspnea level corresponding to the individual data, and the weight value of the custom layer is updated based on the mismatched correspondence.

7. The method for quantitatively assessing the degree of dyspnea as described in claim 5, characterized in that, The step of generating multiple random numbers based on a preset random number algorithm and each of the breathing sound sample data groups includes: A first number of each of the breathing sound sample data groups is determined, and a second number of the random numbers is determined based on the first number and a preset ratio relationship; The first quantity and the second quantity are calculated based on the preset random number algorithm to obtain the calculation result; Determine whether the calculation result is greater than the first quantity. If it is greater than the first quantity, perform a difference operation between the calculation result and the first quantity to obtain the difference operation result as a random number. If the calculation result is less than or equal to the first quantity, then the calculation result is used as a random number; The number of random numbers is counted, and when the number of random numbers reaches a second number, the generation of each random number is completed.

8. The method for quantitatively assessing the degree of dyspnea as described in any one of claims 1-7, characterized in that, The steps for acquiring the wearer's breathing sound signal based on a smart wearable device include: The device serial number and location information of the smart wearable device are obtained, and a key pair is generated by calculating the device serial number and the location information based on a preset key generation algorithm. The key pair includes a decryption private key and an encryption public key. A simulated attack is performed on the key pair to obtain the attack result, and it is determined whether the attack result meets the preset result conditions. If the preset result conditions are met, the encryption public key is sent to the smart wearable device. The device receives ciphertext information encrypted by the smart wearable device based on the encryption key, and decrypts the ciphertext information based on the decryption private key to obtain the breathing sound signal.

9. A quantitative assessment system for the degree of dyspnea based on artificial intelligence, characterized in that, The quantitative assessment system for the degree of dyspnea includes a storage device, a processor, a communication bus, and a control program stored in the storage device. The communication bus is used to enable communication between the processor and the memory. The processor is used to execute the control program to implement the steps of the artificial intelligence-based quantitative assessment method for respiratory distress as described in any one of claims 1-8.

10. A medium, characterized in that, The medium is a readable storage medium, on which a control program is stored. When the control program is executed by a processor, it implements the steps of the artificial intelligence-based quantitative assessment method for the degree of respiratory distress as described in any one of claims 1-8.

Images