Health diagnostic system and method for analyzing animal health
Through a machine learning-based health diagnosis system, using signal reception, noise reduction and analysis modules, the problem of difficult to effectively analyze animal organ sounds in the prior art is solved, and high-accurate health problem diagnosis is achieved, suitable for mobile devices .
Patent Information
- Application Number
- CN202311152673.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-08-09
- Filing Date
- 2023-09-07
- Publication Date
- 2025-02-18
AI Technical Summary
The prior art is difficult to effectively analyze the sound produced by organs in animals to diagnose health problems, especially when noise interference is high.
A health diagnosis system based on machine learning is adopted, including a signal receiving module, a signal noise reduction module and a health diagnosis and analysis module. The signal receiving module receives sound recording from the organ, the signal noise reduction module uses discrete wavelet transformation and resampling technology to reduce noise signals, and the health diagnostic analysis module uses a neural network to analyze the noise-reduced signal to identify normal and abnormal sounds and diagnose health problems.
Improves the accuracy of predicting health problems, reduces noise interference, can operate effectively on mobile devices, and is suitable for rapid analysis of animal health.
Smart Images

Figure CN119480062A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to a health diagnosis system and a method for analyzing animal health, and in particular, but not exclusively, to a health diagnosis system based on machine learning. Background Art
[0002] For many years, the stethoscope has been an essential tool for medical professionals. Doctors, veterinarians, and other medical practitioners use stethoscopes to listen to various sounds in the body during physical examinations. Doctors and veterinarians can use stethoscopes to listen to the heart and lungs of a patient (who can be a human or non-human patient), which can help detect abnormal sounds, murmurs, arrhythmias, and breathing problems. Stethoscopes can also be used to monitor blood flow and blood pressure in arteries and veins.
[0003] A stethoscope can also be used to listen to bowel sounds in the abdomen to detect digestive problems. It can also be used to detect fluid accumulation in the lungs or around the heart, as well as to detect sounds in the gastrointestinal tract to detect problems such as constipation, diarrhea, or bloating. To facilitate recording of these sounds produced by organs in the body, a digital stethoscope can be used if necessary. Summary of the invention
[0004] According to a first aspect of the present disclosure, a health diagnosis system is provided, comprising: a signal receiving module for receiving a sound recording generated by an organ in an animal's body when the organ performs a predetermined function for a predetermined period of time; a signal noise reduction module for reducing noise signals in the recording; and a health diagnosis analysis module for analyzing the health problems of the animal based on normal sounds and occasional sounds generated by the organ in the noise-reduced recording.
[0005] According to the first aspect, the incidental sounds include murmurs and arrhythmia sound patterns produced by the animal's heart or abnormal sounds produced by other organs of the animal.
[0006] According to a first aspect, the health diagnosis analysis module comprises a neural network for classifying the records so as to mark at least one health indicator associated with the health problem.
[0007] According to the first aspect, the noise reduction module is used to reduce interference caused by the noise signal in the record, thereby improving the prediction accuracy of the health problem provided by the health diagnosis analysis module.
[0008] According to a first aspect, the noise reduction module is used to transform the record using discrete wavelet transform (DWT) to reduce noise signals in the record.
[0009] According to the first aspect, after decomposing the signal in the recording, no threshold coefficient is performed, but a resampling process is performed to obtain the noise-reduced signal.
[0010] According to a first aspect, the noise-reduced signal is resampled to 48 KHz, and the frequency range of the signal less than 300 Hz in the recording is retained.
[0011] According to a first aspect, the neural network comprises a lightweight neural network for running on a computing device having a mobile or lightweight processor.
[0012] According to the first aspect, the signal receiving module is further used to verify normal sounds produced by one or more organs in the animal's body for further processing.
[0013] According to a first aspect, the signal receiving module comprises a signal identifying module for verifying the normal sound by identifying the presence of the sound in the recording received by the signal receiving module.
[0014] According to a first aspect, the signal noise reduction module and the health diagnosis analysis module are used to process the sound recording after successful verification of normal sound in the recording.
[0015] According to a first aspect, the recording comprises one or more segments extracted from the recording received by the signal receiving module.
[0016] According to the first aspect, the signal receiving module is further configured to, after successfully confirming the normal sound, extend the sound recording for further processing.
[0017] According to the first aspect, the successful verification is indicated by a positive classification of two consecutive segments, each of the two consecutive segments comprising the normal sound.
[0018] According to a first aspect, the health problem includes risk of valvular heart disease (VHD).
[0019] According to a first aspect, the neural network comprises a downstream phonocardiogram (PCG) classification network for VHD screening.
[0020] According to a first aspect, the downstream PCG classification network is trained based on an upstream self-supervised learning network for PCG classification and a transfer learning process.
[0021] According to a first aspect, the health diagnostic analysis module is used to tag the phonocardiogram associated with the recording by the signal receiving module, thereby facilitating the downstream PCG classification network to tag the at least one health indicator associated with the health problem.
[0022] According to the first aspect, the sound recording includes heartbeat sounds, breathing sounds, lung sounds or bowel movement sounds.
[0023] According to a first aspect, the signal receiving module comprises a microphone for producing said recording of sounds produced by one or more organs in the animal.
[0024] According to a first aspect, the system can be deployed and hosted in a consumer-grade smartphone, digital device, computer, or cloud.
[0025] According to a second aspect of the present disclosure, a method for analyzing animal health is provided, comprising the following steps: receiving a recording of sounds produced by an organ in the animal's body when the organ performs a predetermined function for a predetermined period of time; reducing noise signals in the recording; and analyzing the animal's health problems based on normal sounds and occasional sounds produced by the organ in the noise-reduced recording.
[0026] According to a second aspect, the incidental sounds include murmurs and arrhythmia sound patterns produced by the animal's heart or abnormal sounds produced by other organs of the animal.
[0027] According to a second aspect, the step of analyzing the health problem of the animal based on the normal sounds and murmurs and arrhythmia sound patterns in the noise-reduced recording includes: classifying the recording through a neural network to mark at least one health indicator associated with the health problem.
[0028] According to a second aspect, the step of reducing the noise signal in the recording is performed to reduce interference caused by the noise signal in the recording in order to increase the accuracy of prediction of the health problem being analyzed.
[0029] According to the second aspect, the method further comprises the following steps: prior to the step of reducing the noise signal in the recording and the step of analyzing the health problems of the animal based on the normal sounds and incidental sounds produced by the organs in the noise-reduced recording, verifying the normal sounds produced by one or more organs in the animal's body by identifying the presence of the sounds in the received recording for further processing.
[0030] According to a second aspect, the step of reducing the noise signal in the recording comprises: transforming the recording using a DWT to reduce the noise signal in the recording.
[0031] According to a second aspect, after decomposing the signal in the recording, the step of transforming the recording using DWT includes: not performing a threshold coefficient, but performing a resampling process to obtain the noise-reduced signal
[0032] According to a second aspect, the noise-reduced signal is resampled to 48 KHz, and the frequency range of the signal less than 300 Hz in the recording is retained.
[0033] According to a second aspect, the neural network comprises a lightweight neural network for running on a computing device having a mobile or lightweight processor.
[0034] According to a second aspect, the health problem comprises a risk of VHD.
[0035] According to a second aspect, the neural network comprises a PCG classification network for VHD screening.
[0036] According to a second aspect, the downstream PCG classification network is trained based on an upstream self-supervised learning network for PCG classification and a transfer learning process.
[0037] According to the second aspect, the step of analyzing the health problems of the animal based on the normal sounds and occasional sounds produced by the organs in the noise-reduced recording includes: marking the phonocardiogram associated with the recording received by the signal receiving module, thereby facilitating the downstream PCG classification network to mark at least one health indicator associated with the health problem.
[0038] According to a second aspect, the sound recording includes heartbeat sounds, breathing sounds, lung sounds, or bowel movement sounds.
[0039] According to a second aspect, the method further comprises generating said recording of the sounds produced by one or more organs in said animal recorded by a microphone. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Embodiments of the present disclosure will be described below by way of example with reference to the accompanying drawings, in which:
[0041] Figure 1 is a schematic diagram of a computer server according to an embodiment of the present disclosure, the computer server is used to implement a health diagnosis system;
[0042] Figure 2 is a block diagram of a health diagnosis system according to an embodiment of the present disclosure;
[0043] Figure 3 is a flow chart showing a process of denoising a received recording;
[0044] Figure 4Figure 2 shows a diagram of the upstream model (a) and downstream model (b) architectures; after the upstream model converges in all six tasks, the parameters in the first convolutional layer from the upstream model are transferred to the downstream model and frozen.
[0045] Figure 5 Figure 2 is a diagram showing the visualization of the six transformations applied to PCG data. After applying the transformations, pseudo labels are automatically generated.
[0046] Figure 6 It shows that a portion of random length at the beginning of the original signal is clipped and appended to the tail of the signal. The perturbed signal starts at a random position in the cardiac cycle, rather than at the beginning of S1. The purpose of the perturbation is to further test the robustness of the classification model.
[0047] Figure 7 A diagram showing the process of developing a mobile app prototype. The diagram also includes screenshots of the app. Users can press the corresponding buttons to fine-tune the pre-trained model or perform inference locally. The time required to complete a specific task is recorded and presented on the interface.
[0048] Fig. 8A A block diagram of the aggregation results. Regardless of whether the signal starts from S1, the model can perform better when pre-trained with unlabeled data upstream.
[0049] Figure 8B is another block diagram showing the aggregated results of different attempts.
[0050] Fig. 9 A graph showing a visualization of the training history of the upstream model shows that the model will eventually converge on all six tasks. Task 1 is the easiest to learn, while tasks 3 and 6 are relatively tricky.
[0051] Fig.10 is the confusion matrix of the median cross validation results from the model without SSL on the edited dataset (a) and the median cross validation results from the model with SSL on the edited dataset (b).
[0052] Fig.11 A line graph showing the aggregated results of the two model attempts with increasing noise levels. The model with SSL is more robust when Gaussian noise is injected into the signal. b Performance of the two methods when training for different numbers of epochs.
[0053] Fig. 12A It is the category activation map of the actual category and predicted category of sample No.17. The sample is misclassified by the model without SSL, but correctly classified by the model with SSL.
[0054] Fig. 12BIt is the class activation map of sample No. 46. DETAILED DESCRIPTION
[0055] The inventors believe that Cardiovascular Disease (CVD) encompasses the diseases that are the leading causes of illness and death worldwide. Valvular heart disease (VHD) is a type of CVD in which one or more heart valves have functional or structural defects that affect blood flow in the heart.
[0056] Undetected and untreated VHD can lead to worsening cardiac function and death. VHD remains a major health burden in developed countries. Detection of abnormal heart sounds (heart murmurs) could be a simple strategy for population-based VHD screening. With the widespread availability of mobile smartphones, phonocardiogram (PCG) data can be collected and analyzed on these edge devices for VHD screening.
[0057] For example, an interpretable PCG in the database contains four identifiable heart sound phases (S1, systolic, S2, and diastolic). The presence of different murmur patterns in the sound segments can indicate VHD. Heart sounds from the dataset can be classified as normal and abnormal. Alternatively, the PCG dataset can contain five categories (mitral valve prolapse, mitral valve stenosis, mitral valve regurgitation, aortic valve regurgitation, and normal heart sounds), and each category can have 200 clean PCG samples.
[0058] In addition, PCG data can be applied to the heart sound classification model based on machine learning for VHD screening. Heart sound classification models can be divided into artificial neural network-based, support vector machine-based, hidden Markov model-based and clustering-based methods.
[0059] refer to Figure 1 , showing an embodiment of the present disclosure. The embodiment provides a health diagnosis system, including: a signal receiving module for receiving a sound recording generated by an organ in an animal's body when the organ performs a predetermined function for a predetermined period of time; a signal noise reduction module for reducing noise signals in the recording; and a health diagnosis analysis module for analyzing the animal's health problems based on normal sounds and occasional sounds generated by the organ in the noise-reduced recording, such as murmurs produced by the heart. For clarity, the term "animal" should be biologically defined and therefore includes humans as well as non-human animals, such as but not limited to pets, cattle or exotic wildlife.
[0060] In this exemplary embodiment, the interface and the processor are implemented by a computer with an appropriate user interface. The computer can be implemented by any computing architecture, including: a portable computer, a tablet computer, a stand-alone personal computer (PC), a smart device, an Internet of Things (IoT) device, an edge computing device, a client / server architecture, a "dumb" terminal / host architecture, a cloud computing-based architecture, or any other appropriate architecture. The computing device can be appropriately programmed to implement the present disclosure.
[0061] The system can be used to provide information related to the health of a patient (e.g., a human) based on analysis of biomedical sounds produced by certain organs (e.g., the patient's heart). For example, a recording of sound clips mixed with "normal" heartbeats and heart murmurs can be provided to a system for determining patient health issues. A heart murmur is an abnormal sound detected during a heartbeat when blood flow is not smooth. These abnormal sounds, i.e., murmurs or arrhythmia sound patterns, can be caused by various factors, including valve problems, congenital heart defects, or infections. However, not all murmurs are harmful. Some murmurs may not require medical attention. Therefore, accurately labeling murmurs can help practitioners take necessary measures to identify different murmurs.
[0062] It can be vital to have a health care provider evaluate the severity and cause of the murmur to determine if the murmur is benign or an indicator of a more serious underlying condition. Medical options may be determined based on the cause of the murmur and the severity of the condition. In some cases, medication or surgery may be necessary to correct the underlying problem and prevent further complications.
[0063] For example, health care providers may also use other diagnostic tools such as echocardiograms, electrocardiograms (ECGs), or chest X-rays to verify health concerns. These tests can help visualize the structure and function of the heart and provide more information about the underlying cause of a murmur or arrhythmia sound pattern.
[0064] Without departing from the spirit of the present disclosure, sound recordings produced by any animal species other than humans can also be processed by the system of the present disclosure, and can also help veterinary practitioners quickly analyze any health problems of an animal. In addition, sounds produced by other organs and multiple organs can also be analyzed using the system.
[0065] like Figure 1, a schematic diagram of a computer system or computer server 100 is shown. The computer system or computer server 100 is used to implement an example embodiment of a system. The system includes computer executable instructions, which, when executed, are used to perform a method for analyzing animal health. In this embodiment, the system includes a server 100. The server 100 includes appropriate components required to receive, store and execute appropriate computer instructions. These components may include: a processing unit 102, which includes a central processing unit (CPU), a math co-processing unit (Math Co-processing Unit, Math Processor), a graphics processing unit (Graphics Processing Unit, GPU) or a tensor processing unit (TPU) for tensor or multi-dimensional array calculation or manipulation operations; a read-only memory (ROM) 104; a random-access memory (RAM) 106; an input / output (I / O) device (e.g., a disk drive 108); an input device 110 (e.g., an Ethernet port, a USB port, etc.); a display 112, such as a liquid crystal display, a light-emitting display, or any other suitable display; and a communication link 114. The server 100 may include instructions that may be stored in the ROM 104, the RAM 106, or the disk drive 108 and executed by the processing unit 102. A plurality of communication links 114 may be provided that may be variously connected to one or more computing devices, such as servers, personal computers, terminals, wireless or handheld computing devices, IoT devices, smart devices, and edge computing devices. At least one of the plurality of communication links may be connected to an external computing network via a telephone line or other type of communication link.
[0066] The server 100 may include storage devices such as a disk drive 108, which may include a solid-state drive, a hard disk drive, an optical drive, a tape drive, or a remote or cloud-based storage device. The server 100 may use a single disk drive or multiple disk drives, or a remote storage service 120. The server 100 may also have a suitable operating system 116 resident on the disk drive or in ROM of the server 100.
[0067] The computer or computing device may also provide the necessary computing power to operate or interface with a machine learning network such as a neural network to provide various functions and outputs. The neural network may be implemented locally, or may be accessed or partially accessed through a server or cloud-based service. The machine learning network may also be untrained, partially trained, or fully trained, and / or may also be retrained, modified, or updated over time.
[0068] refer to Figure 2 , an embodiment of a health diagnostic system 200 is provided. In this embodiment, a server 100 is provided as part of a system. The system is used to receive a recording of sounds 202 or sound clips generated by an organ (e.g., a patient's heart), and after processing the sound clips, generate a report listing the patient's health issues based on the recorded patterns of periodic sounds and murmurs. This helps medical personnel follow up with other diagnostic tests, if necessary, to determine whether the patient suffers from any CVD, such as heart valve disease.
[0069] In this example, the system 200 includes a signal receiving module 204, such as a microphone for detecting sound waves that can be further recorded in a memory device, or it can be a software module for reading sound clips in a pre-recorded form or a digital file in an electronic form. For example, a recording of a patient's heartbeat can be provided to the system for analysis.
[0070] If one does not wish to be bound by theory, the heart sound recording may be classified as "uncertain", meaning that the noise in the recording is too loud to be interpreted. The recording may therefore affect the performance or accuracy of the analysis. In this example, the health diagnostic system 200 includes a signal noise reduction module 206 for reducing the noise signal in the recording 202, so that the health diagnostic analysis module 208 can analyze health problems based on a recording with a clearer signal pattern. For example, a digital signal processor (DSP) may be used to process noisy sound clips so that when an organ in the animal's body performs a predetermined function for a predetermined period, the noise signal is clearly separated from the sound produced by the organ. Alternatively, a machine learning or neural network processing engine may be used to perform noise reduction processing.
[0071] In addition, the noise reduction module 206 can reduce interference caused by noise signals in the recording, thereby improving the accuracy of prediction of health problems provided by the health diagnosis analysis module. By reducing the noise of the original sound clips provided to the system 200, the post-processing resources required to operate the health diagnosis analysis module 208, such as computing, storage and transmission resources, can be reduced. Preferably, the health diagnosis analysis module 208 includes a neural network for classifying the recording so as to mark at least one health indicator associated with the health problem, and the neural network can be a lightweight neural network for running on a computing device with a mobile or lightweight processor, such as a processor of a smart phone, a tablet computer, a handheld computing device, etc. Preferably, the noise reduction module 206 can also output the identified audio track 216 after noise reduction, for example, for further processing by the health diagnosis analysis module, or the output audio track 214 can be saved as a data file for subsequent use or analysis.
[0072] Preferably, the noise reduction module 206 can reduce noise in the cardiophone (heart sound) recording (or other type of biomedical sound recording) by using a wavelet-based technique, such as a discrete wavelet transform (DWT), to reduce noise in the cardiophone (heart sound) recording (or other type of biomedical sound recording). For example, thresholding can be applied to the coefficients output from the decomposition and the parameter selections on the DWT and thresholding techniques can be compared, although other methods can also be used to implement the noise reduction module 206, including using a machine learning network that has been trained to recognize certain types of signals for filtering or selection.
[0073] Alternatively, it is not necessary to perform threshold processing on the coefficients after decomposition 310. For example, referring to Figure 3 , all detailed coefficients 312 can be reset to zero, and the approximate coefficients 314 can be kept at the highest level. Then, during the resampling process, the signal can be reconstructed based on the approximate coefficients 314. With this improved method, excluding the threshold process and its replacement by simply erasing the high-frequency signal will significantly reduce the time complexity. In this way, the algorithm can be executed in a device with limited computing resources.
[0074] Preferably, providing a noise reduction signal can ensure that the noise reduction module 206 can reduce interference caused by the noise signal in the record, thereby improving the prediction accuracy of the health problem provided by the health diagnosis analysis module.
[0075] In order to further improve the accuracy of the system, refer to Figure 2, the signal receiving module includes a signal identification module 212 for "screening out" invalid signals or records for further processing. Preferably, the signal receiving module can further verify the sounds produced by one or more organs in the animal body for further processing, for example, by identifying the presence of sounds produced by one or more organs in the records received using the signal identification module. In some exemplary embodiments of the present disclosure, the normal sound is a heart sound, which is a periodic sound at a high frequency, and preferably, by identifying the periodic pattern and the frequency of the input periodic sound, the signal identification can immediately indicate whether the heartbeat is successfully captured to verify the input sound. On the other hand, abnormal sounds (or extraneous sounds) may include (but are not limited to) systolic or diastolic murmurs in heart sounds, as well as rales, wheezes, and dry rales.
[0076] Preferably, the signal noise reduction module and the health diagnostic analysis module are used to process the sound recording after successful verification of normal sounds in the recording. For example, the signal receiving module may choose to transmit the recorded signal for further processing only after successfully verifying the presence of the heart sound in the recording, and preferably selects a clear recording of the heartbeat that can be processed by a lightweight processor in the smartphone. Otherwise, the system will provide an error message to notify the user to reposition the microphone to the correct position in order to capture a recording that can be processed by the noise reduction module and the health diagnostic analysis module implemented by using the smartphone processor after successfully confirming the sound. Advantageously, the accuracy of the analysis can be improved by ensuring that recordings that may be blurred or captured at incorrect locations on the patient's body are not further processed or analyzed.
[0077] By performing an additional signal recognition process, an immediate indication of whether the heartbeat was successfully captured can be provided, such as through a built-in microphone in a smartphone, so that the person performing the self-auscultation can be notified, and when valid information is confirmed to be involved, the recording of PCG data can be started. Preferably, a small footprint keyword detection system widely used in virtual voice assistants can be adopted, and the classifier can be constructed in a similar manner to make the system accurate, responsive, and able to operate with low power consumption. Reference Figure 2 The signal receiving module is used to output the identified audio track 214 for further processing, such as further processing by a signal noise reduction or health diagnosis analysis module, or the output audio track 214 can be saved as a data file for subsequent use or analysis.
[0078] Preferably, the input recording may include one or more one-second segments extracted from the recording received by the signal receiving module, and the signal receiving module may further extend the sound recording for further processing, such as classification of disease or risk, upon successful verification of the sounds produced by one or more organs in the input recording.
[0079] For example, the heart sound recording can be windowed into 1 second segments non-overlappingly, and the vast majority of the segments will contain heartbeats. This is because the normal range of resting heart rate is 60-100bpm. A 1 second window time interval will also speed up the process of determining whether a heartbeat is captured compared to longer time intervals. In addition, the input segments can be downsampled to 1KHz to ensure that the heart sound frequency range less than 300Hz is included and to reduce the complexity of the input data to the CNN. The heart sound segments are scaled to a range between 0 and 1.
[0080] Preferably, the recognition module can be trained based on a CNN tuned with labeled data. A heart sound dataset is trained, where negative labeled samples include unrecognized heart sound recordings. In addition, white noise with small amplitude and blank recordings is generated to simulate the situation where the smartphone is placed in a quiet environment and the microphone does not capture the heartbeat. Positive labeled data including normal and abnormal heart sound recordings with good quality can also be included in the training process.
[0081] In an exemplary embodiment, the CNN is constructed to be as lightweight as possible so that it will subsequently occupy as little computing resources and storage space as possible in a mobile device. The input layer of the CNN is constructed to accept a one-dimensional heart sound signal containing 1000 samples. Then, the signal can be processed with three convolution-max pooling layers. The kernel sizes of the three convolutional layers are 100, 50 and 10 respectively, and the kernel size of the maximum pooling layer is 4. The number of filters of the three convolutional layers is 16, 8 and 4 respectively. Then, the last pooling layer is followed by three fully connected dense layers with decreasing number of units. Finally, an output layer with only one node representing the binary classification result is added to the tail of the network. The Relu activation function is used in all hidden layers, and the sigmoid function is used in the output layer. The training loss is calculated as the binary cross entropy between the actual output and the expected output. The network is trained using the Adam optimizer, and 10% of the training data is randomly split into a validation set to stop the training process early and save the model with the highest validation accuracy.
[0082] Optionally, in an embodiment of the heartbeat identifier, a simple fault tolerance mechanism may be employed. For example, a positive classification of two consecutive 1 second segments each containing sounds produced by a body organ indicates successful verification. Thus, the recording of heart sounds will begin after two consecutive 1 second segments are positively classified to avoid the situation where a noise similar to a heartbeat starts the recording process by verifying the periodicity of the sound. Similarly, the recording process will stop after two consecutive 1 second segments are negatively classified to ensure that the process is not interrupted by bradycardia.
[0083] In an exemplary embodiment, a lightweight one-dimensional convolutional neural network (CNN) architecture is used to analyze health issues in animals based on sounds produced by body organs and incidental sounds (e.g., murmurs, arrhythmic sound patterns, or irregular heartbeats) in noise-reduced recordings. An exemplary one-dimensional CNN is constructed by following the following strategy to minimize the size of the network while maintaining accuracy. After the analysis is completed, outputs such as labeled recordings or even automatically generated health reports 210 can be obtained.
[0084] First, complex components in the neural network are avoided, including residual blocks, extractors, and enhancers. These components are designed to achieve certain purposes. For example, residual blocks were originally proposed to solve gradient-related problems in the back-propagation stage of very deep neural networks. This component has been shown to be helpful for tasks that require large models to handle complex and high-dimensional data. However, a relatively shallow network with only basic CNN layers is sufficient to classify PCG data as a periodic signal, especially when the data is clean.
[0085] Second, small convolutional filters are applied to all convolutional layers. As suggested by SqueezeNet, by using small filters within a certain number of filter budgets, the number of parameters can be greatly reduced.
[0086] Third, the features output by the convolutional layer are max-pooled so that fewer parameters will be sufficient to process these features in the subsequent fully connected dense layer. According to the inventor's design, the machine learning classifier can achieve high accuracy on a feature vector of about 40 in length, which can be extracted from the original PCG data. Before feeding to the dense layer, the outputs of the convolutional and pooling layers are fully compressed to reduce the need for parameters.
[0087] In a preferred embodiment, the neural network includes a downstream PCG classification network that can be used to screen for the risk or possibility of valvular heart disease (VHD). For example, when the health diagnosis analysis module receives a noise-reduced record associated with a phonocardiogram being generated, the PCG classification network can mark the phonocardiogram received by the signal receiving module, thereby facilitating the downstream PCG classification network to mark at least one health indicator associated with a health problem.
[0088] For example, refer to Figures 4 to 6, showing an exemplary embodiment of the operation of the system. A user can record the sounds of his chest, which include sounds produced by the heart, and upload the recorded sound clips to the system for VHD screening, for example, through an application installed on a smartphone or tablet device. Preferably, the screening process can be performed locally by a built-in processor of the smartphone or tablet device. The built-in processor of the smartphone or tablet device can be a lightweight processor or a processor with relatively limited resources (e.g., processing power and / or bandwidth), rather than a processor in a computer server. Therefore, a preferred solution is to shrink a neural network that can be specifically designed to run on a mobile device or edge device.
[0089] Preferably, the downstream PCG classification network is trained based on the upstream self-supervised learning network for PCG classification and transfer learning. Figure 4 At the beginning of the model, the input layer of the network is used to accept a one-dimensional PCG signal containing 2000 samples. The signal is then processed with three convolution-max pooling layers. The kernel size in all three convolutional layers is 20, while the kernel size in all three pooling layers is 4. The number of filters in each of the three convolutional layers is 16, 8, and 4, respectively. The final pooling layer is followed by a fully connected dense layer with 128 units. Finally, the output layer with five nodes outputs the results of 5-class classification through the SoftMax activation function. The Tanh activation function is used in all hidden layers. The training loss is calculated as the categorical cross entropy between the actual output and the target output. The ADAM algorithm can be used in the network optimization process. In this example, the entire model contains only 20,193 parameters.
[0090] The forward propagation of a one-dimensional CNN can be symbolically represented as:
[0091] Y=F(X|Θ)=f n (…f2(f1(X∣θ1)∣θ2)∣θ n ) (1)
[0092] In this equation, Y represents the prediction of CNN, X represents the input one-dimensional PCG data, and f n Represents the calculation of the nth layer, where θ th Represents the learnable parameters on this layer. In addition, the operation of the convolutional layer can be described as:
[0093]
[0094] represents the activation function applied to the layer, Y represents the convolution operation, W represents the set of one-dimensional convolution receptive fields implemented to extract features, and b represents the bias vector. n and X nRepresent the output of the current layer and the output of the previous layer respectively.
[0095] In this example, the upstream task (also called the pre-text task) is designed to adjust the parameters in the upstream network to identify important representations from the data. The tuned parameters can then be transferred to the downstream network to assist the downstream task.
[0096] Preferably, reference Figure 5 , six different transformations 502 can be applied to the unlabeled PCG samples 502 to produce a transformed dataset Xt and six sets of pseudo labels. In a preferred embodiment, a one-dimensional CNN as an upstream model in a multi-task learning manner is trained from the dataset and pseudo labels to learn the PCG representation. Similar methods are used to represent many other types of data, such as electrocardiogram and accelerometer data.
[0097] In this example, the transformation 502 applied to the unlabeled PCG samples includes:
[0098] Noise addition: Gaussian noise of the same length as the signal and with a scale of 0.2 is generated and added to the PCG data.
[0099] Scaling: The size of PCG data is scaled to 80% of the original.
[0100] Negate: Flips the PCG data in the amplitude dimension.
[0101] Flip: PCG data is flipped in the time dimension.
[0102] Permutation: PCG data is cut into 10 segments and reconnected in random order.
[0103] Stretch: Cut the PCG data into 10 segments and stretch or squeeze each segment in the time dimension.
[0104] refer to Figure 5 , shows a visualization of six types of transformations 502. During the transformation process, each unlabeled original PCG sample is calculated to have a 50% probability of being applied to each transformation. If a sample has a specific type of transformation applied, a positive pseudo label is generated under the corresponding task. Otherwise, a negative pseudo label is generated. At the end of the transformation process, each sample has six binary pseudo labels indicating whether the corresponding transformation occurred.
[0105] Refer to Figure 4, the one-dimensional CNN derived from the above laconic architecture is designed to perform multi-task learning from the transformed PCG dataset and pseudo-labeling. The upstream CNN 402 is equipped with the same structure of three convolutional maximum pooling layers as the downstream model 404. Then, the extracted feature vectors are flattened and fed into the six branches 406 of the fully connected dense layer respectively. Each branch is two layers deeper than the part of the downstream model to converge more easily and quickly. Since each task is a binary classification, the last layer is changed to output a pair of possibilities through the SoftMax function.
[0106] Preferably, the model can be trained with the Adam optimizer, and binary cross entropy can be used to calculate the loss on each task branch. Given that all six tasks are derived from pseudo-labeling and should be treated equally, a simple approach is adopted to calculate the loss of the upstream network as the sum of the losses of each individual task. By learning the six tasks in parallel, the convolutional layers of the upstream model can extract certain representations from the PCG data.
[0107] Reference Figure 4 , the upstream model 402 is compared with the downstream model 404. After the multi-task model converges in all tasks, the parameters of the first convolutional layer of the upstream model (θ1 in Equation 2) are transferred to the downstream model as the first convolutional layer that cannot be changed. As shown in the following Algorithm 1-pseudocode for SSL technology and how it is incorporated into the pipeline to train an efficient 1D-CNN to classify PCG data, the detailed process of pre-training the upstream model with unlabeled data and the generated pseudo-labels and transfer learning is abstracted into pseudocode.
[0108]
[0109]
[0110]
[0111] Preferably, the sound recording includes heartbeats, and thus heart disease can be screened based on the periodic sounds produced by the heart. Alternatively, breathing sounds (i.e., sounds produced in the chest), lung sounds, or sounds of bowel movements can be analyzed to diagnose health problems related to other parts of the body, where the neural network can be trained by different sets of sound training sets and diagnostic analysis provided by medical practitioners.
[0112] Those skilled in the art will appreciate that even though the classifier or lightweight network may be designed to analyze sounds produced by a specific organ and produce results based on the classification of sounds produced by that organ, the system is also capable of simultaneously receiving and processing sound recordings produced by multiple organs and selectively marking or classifying components related to sounds produced by a specific organ as needed. It is also possible that, in some cases, the absence of sounds produced by certain organs may also indicate that the user of the App may suffer from certain diseases.
[0113] The inventors conducted experimental evaluation and statistical analysis of the system. In the experiments, 10-fold cross validation was applied in most of the validation experiments for measuring model performance. In each of the ten iterations of cross validation, one fold was set aside as a test set to evaluate the model trained on the remaining nine folds. The average accuracy of 10 iterations was obtained from the data set as an aggregate measure, which was taken as 40 attempts. Whenever possible, the neural network model was trained 40 times to account for the randomness in model training due to the random initialization of weights, the shuffle order of feeding training samples, and random optimization. A two-sample unpaired t-test was used for two-category comparisons.
[0114] In some examples, the PCG data recorded by the real device may not be aligned with the phase of the cardiac cycle. The robust model designed by the inventors should be able to classify PCG recordings starting from random positions.
[0115] refer to Figure 6 , in order to evaluate the robustness of the models, the signals from the dataset were perturbed by cutting each recording into two parts at an arbitrary point and swapping the order of the two parts. This process was performed before the aforementioned preprocessing steps. The recordings would start from a random position in the cardiac cycle instead of from the beginning of S1 after processing. In addition, the PCG signals in the perturbed dataset were subsequently contaminated with Gaussian noise to further evaluate the robustness of the different models. The mean of the normal distribution was kept at zero, while the standard deviation of the distribution was gradually increased from 0.05 to 0.3.
[0116] As previously described, the health diagnostic analysis module may be implemented as a lightweight neural network that can be run on a computing device with a mobile or lightweight processor by inputting the recording 202 or specifically the identified audio track 214 after initial verification by noise reduction 206. Preferably, the system may be integrated with the functionality of generating a recording of the sound produced by the organs in the animal, which is collected by a built-in microphone or an external microphone connected to the smartphone through any suitable network or connection means.
[0117] In an example operation, a patient can launch an App running the health diagnostic system on their smartphone and place the smartphone's microphone close to their chest near their heart to start recording their heartbeat. The system begins by verifying the presence of a heartbeat. If the verification fails, the App can notify the patient to place the microphone in another location that can help better capture the sound of the heart, or until the App identifies a heartbeat.
[0118] Once the verification is successful and the patient's heartbeat is identified, the signal is further de-noised and the de-noised recording is then further analyzed by a lightweight network in the smartphone. The machine learning process can then return any potential health issues or risks to the patient performing the diagnosis, such as the risk of developing VHD. The analysis results and recordings can be further provided to a medical advisor for additional health advice to the patient.
[0119] Preferably, a lightweight deep learning model and a self-supervised learning (SSL) method are provided to utilize unlabeled data to improve the accuracy of valvular heart disease classification using phonocardiogram data. Advantageously, the detection and classification of heart murmurs using sounds collected by mobile phones can expand the screening of valvular heart disease at the population level. Highly accurate and lightweight AI models can be deployed in consumer-grade mobile devices, which can promote the adoption of artificial intelligence (AI) methods for mobile health (mHealth) applications.
[0120] In an example embodiment, referring to Figure 7 , by converting the KERAS model 602 to a TensorFlow.js web-friendly format 604 and developing a Progressive Web Application (PWA) 606 to keep the model behind a basic user interface, the model with pre-trained parameters is deployed to consumer-grade smartphones. PWA is an application software format that can use a standards-compliant browser as a medium and is designed to be reliable and installable, and can provide an experience that feels like a native application, especially on mobile devices. Advantageously, the relatively easy development, small installation size, and short startup time of PWA make it suitable for participating in academic tasks to collect or distribute information compared to hybrid or interpreted software.
[0121] refer to Figure 7, the user interface of the application and the logic flow thereafter are built on the Nuxt.js framework 608 with the BootstrapVue library, whose function is to extract the model and binary weights 610 stored in the JSON file to fine-tune or inference on the dataset placed in the application. 100 samples of the dataset from Yaseen are reserved to perform on-device fine-tuning and inference. The samples are input into the model with a batch size equal to 2 and trained for 1 epoch in the fine-tuning procedure.
[0122] Although the exemplary smartphone-based system is only used to demonstrate that lightweight models can be fine-tuned or quickly inferred in smartphones, and accuracy is not a major issue, the lightweight one-dimensional CNN is able to achieve an average 10-fold cross-validation accuracy of 98.65% (95% confidence interval (CI: [98.555%, 98.74%]), which is better than the accuracy of other example CNN models shown in the table below. The architecture of the model is also applied to the PASCAL dataset, which consists of records recorded by smartphones. The output layer of the model is modified to suit the binary classification task, but the other structures and hyperparameters of the proposed model are retained. The following table illustrates the comparison of the accuracy and number of parameters between the present disclosure and other example models that do not require any prior manual feature extraction or analysis. It can be observed that the system according to the preferred embodiment of the present disclosure includes only a limited number of parameters while maintaining high accuracy.
[0123]
[0124] To illustrate the impact of self-supervised learning (SSL), a new upstream model was retrained to generate a new downstream model to compare the performance difference between the models involving SSL and those not involving SSL. After 10-fold cross validation, the average accuracy of the one-dimensional CNN with pre-trained parameters was 99.435% (95% CI: [99.380%, 99.495%]). Fig. 8A As shown, the model applying SSL showed significantly higher accuracy when classifying the five categories of PCG data.
[0125] During upstream SSL, the upstream network can converge to all six tasks after a certain number of learning epochs to capture basic representations from PCG data. The training curves of the six tasks in the upstream network are shown in Fig. 9 shown.
[0126] To further validate the improvements brought by SSL, the same pipeline was applied to the train-test-split PASCAL dataset to see if the test accuracy can be significantly improved after incorporating the pre-trained parameters into the downstream model. The average test accuracy of the model with and without SSL was 79.78% (95% CI: [78.348%, 81.211%]) and 74.19% (95% CI: [72.014%, 76.369%]). These results indicate that SSL significantly improves the prediction accuracy of the model (p value = 4.27×10 -5 ). The relatively low test accuracy of the PASCAL dataset may be caused by the limited amount of training data and the lack of preprocessing (such as segmentation and denoising) available in the dataset. However, the results demonstrate that SSL can indeed be used to improve the accuracy of the proposed lightweight CNN.
[0127] The models were trained and tested on perturbed data, both with and without SSL. When the data was less well-formed, there was a larger gap between the performance of the two approaches. The average 10-fold cross validation accuracy of the model without SSL dropped to 92.538% (95% CI: [92.304%, 92.771%]). The SSL model was more robust and the number dropped to 96.123% (95% CI: [95.911%, 96.334%]). A summary of all attempts is plotted in Figure 8B The confusion matrix of the median attempt of the two methods is shown in the box plot. Fig.10 As shown in Figure 2. When the data has random starting points, the model without SSL performs significantly worse. Specifically, the accuracy of identifying MR and MVP categories is less than 90%. Alternatively, the model with SSL is more robust and less affected by data reshuffling.
[0128] Experiments comparing models with and without SSL are performed on a reformatted dataset with different levels of Gaussian noise. Fig.11 The model with the help of unlabeled data always outperforms the model trained purely on labeled data. Moreover, as the noise level increases, the accuracy growth gap between the two models can be seen. In addition, when the first convolutional layer is pre-trained with unlabeled data, the model converges much faster. Fig.11 In the b-plot, at 100 training epochs, even after only 15 training epochs, the model with SSL has achieved higher accuracy than the model without SSL.
[0129] In another experiment conducted in the evaluation study, PWAs on four different smartphones were installed with two mainstream operating systems, iOS and Android, to evaluate the time required to tune or execute the proposed lightweight model. Another model with more filters, larger kernels in the convolutional layers, and deeper parts of the fully connected dense layers was also built and deployed in the app for comparison. The large-size model contains 241,597 parameters. As shown in the table below, the operation of fine-tuning the small model with 100 samples on two iOS devices in collaboration with the Safari browser is completed in 5 seconds. The inference of 100 samples is completed instantly in tens of milliseconds. However, the larger model in the iPhone takes almost twice as long to perform fine-tuning, and the inference time of the two models differs by several orders of magnitude. There is a similar gap in the performance of the two models on Android devices with the Firefox browser, and the operation speed varies significantly based on hardware differences. The additional running time caused by the increased parameters affects the implementation of the model in resource-constrained devices. For example, the large model takes more than 5 seconds to complete inference in the Samsung Galaxy A30, while the small model only takes less than 0.4 seconds.
[0130] Referring to the table below, six exemplary smartphones with two mainstream operating systems installed are used to run the application.
[0131] The processing times reported in the table are the average of five attempts.
[0132]
[0133]
[0134] TensorFlow.js leverages the WebGL API as a backend to enable efficient GPU-accelerated training and inference of deep learning models. However, it has been observed that the maximum supported texture size varies across different browsers and operating systems, which may affect the ability of a model to successfully execute on a given device. It is worth noting that Chrome, the current most dominant browser for mobile devices, only supports a maximum texture size of 4096 on the tested Android devices, and because the texture size of the model exceeds the maximum value of WebGL, the execution of large-size models on three tested Android devices using the Chrome browser failed.
[0135] It is observed that deep learning models may not be successfully executed in the Chrome browser on Android devices due to limitations of WebGL and large-sized models have fewer parameters than other models. However, on iOS and other Android browsers, the maximum acceptable texture size is higher, allowing larger models to be successfully executed despite longer processing time.
[0136] Removing redundant parameters from deep neural networks reduces the time required to perform computations and significantly saves the cost of energy consumption. The power consumption of executing small and large models was measured and compared on a regular smartphone. Three PWAs were deployed, with 1) only the user interface, 2) an interface with ten inferences performed by a small-size model, and 3) an interface with ten inferences performed by a large-size model. Greenspector was used to analyze the battery consumption of the three PWAs on a Galaxy S7 with Android 8 installed with standardized access. The Greenspector website gives a detailed analysis method. Five measurement attempts were recorded. Compared to the small model, the energy consumption of the PWA running on the large model was nearly three times that of the small model, increasing from about 0.6mAh to 1.7mAh. This is a statistically significant increase (t-test, p=1.19×10 -7 ).
[0137] Advantageously, the effective and efficient execution of fine-tuning and inference of lightweight models on mobile devices demonstrates that it is feasible to provide classification of PCG data through deep learning models in a distributed computing paradigm without the need to transmit data and rely on central servers. In addition, the model can be fine-tuned locally in a reasonable cycle. Recent decentralized machine learning training techniques such as federated learning and swarm learning can be applied, so that more data can be included or more personalized classification can be provided by aggregating updated parameters from edge devices.
[0138] In another experiment conducted by the inventors, the class activation maps (CAMs) of the actual and predicted classes produced by GradCAM++ were generated by the convolutional layers of a representative CNN trained with and without SSL. The purpose of this experiment was to visualize the important features that would help in prediction. Fig. 12A and 12B ,Two tracks, No.17 and No.46, are selected as case studies because they are misclassified by the model without SSL and correctly classified by the model with SSL. The actual label of sample No.17 is AS, but the model without SSL misclassifies it as MR.
[0139] As shown in the CAM, the activations of the AS class from the non-SSL model are similar to those of the MR class. At the same time, most of the attention is focused on the diastolic period rather than the systolic period, and the murmur is located in the systolic period. In contrast, the activation area of the actual class from the SSL model covers the most systolic period of the sample, which ensures that the important features of the murmur of aortic stenosis are well captured (i.e., the murmur is mid-systolic, diamond-shaped, and high-pitched).
[0140] A similar situation was observed in sample No.46. As shown in the CAM, the convolutional layers in the non-SSL model did not notice the diastolic murmur. This neglect of the murmur gave the model an incorrect judgment. In contrast, the model with SSL involved a larger range of activation areas for the MS category, including S1, S2, and diastolic murmur, which drove the model to identify the murmur and make a correct judgment about the location of the murmur.
[0141] Leveraging SSL and a large amount of unlabeled data, a lightweight neural network provides fast, accurate, and robust classification of the five-class PCG dataset, even on consumer-grade smartphones with relatively limited computing resources.
[0142] The inventors have designed that, outside laboratory conditions, heart sounds recorded by mobile devices may be weak, noisy, and chaotic at various stages. Although SSL was tested to improve the robustness of PCG classification, it can provide future realistic scenarios for mobile phone-based VHD automatic screening from two aspects.
[0143] Advantageously, an effective heart sound denoising and segmentation algorithm has been developed to process real-world raw PCG signals. Optionally, a signal amplification or enhancement process may also be applied to exploit the differences between mobile phones and digital stethoscopes.
[0144] Preferably, in alternative embodiments, a customized SSL task for PCG data can also be developed, while the upstream contrast task used in this study is generally applicable to other biomedical signals. It is designed that the design pattern for balancing upstream and downstream models can be further tuned / optimized. For example, transferring and freezing the parameters of more than one convolutional layer to the downstream model leads to performance degradation. In the small model used in this study, setting a large number of unavailable parameters may cause the convolutional layers from the upstream task to be compressed and may cause the focus from the downstream task to be lost.
[0145] Furthermore, a PCG database of phone records can be built to handle real-world phone-based VHD screening scenarios.
[0146] The advantages of these embodiments are that a lightweight one-dimensional CNN is provided. The system achieves an average accuracy of 98.6% in 10-fold cross validation in a five-class PCG dataset. Compared with other example deep learning models, the parameters of the neural network are reduced by 92% to 97%.
[0147] Advantageously, the present disclosure significantly improves the classification average accuracy to 99.4% and enhances the robustness of the model by involving approximately 25,000 unlabeled data from PhysioNet and performing comparative self-supervised learning. In addition, the model can be converted into a browser executable version and deployed to an application for instant PCG classification on a smartphone. As a lightweight model, it allows deployment in more smartphones. In addition, power consumption can also be reduced.
[0148] In addition, class activation maps were generated and compared to show that the model with pre-trained layers can better capture the identification features with biological relevance. Early screening of VHD is an important medical topic, and the present disclosure can improve the efficiency of deep learning-based VHD screening with the help of unlabeled data and blend reliable PCG classification into mobile health.
[0149] As mentioned above, lung sounds can be analyzed to diagnose health problems related to other parts of the body, namely lung diseases. The inventors have designed an alternative embodiment of the health diagnosis system, which is called "Auscultation Lungs" and can be used to detect abnormalities in lung sounds to screen for airway related diseases.
[0150] Lung disease is a significant cause of morbidity and mortality worldwide. Early detection and diagnosis of respiratory abnormalities are essential for the effective management and treatment of lung disease. Auscultation is a non-invasive technique that involves listening to the sounds produced by the lungs during breathing. According to the inventor's design, by using AI technology, the lung sound analysis process can be automated, and the accuracy and reliability of diagnosis can be improved.
[0151] In an exemplary embodiment, auscultating the lungs may consist of three main parts: identification, noise reduction, and classification.
[0152] Preferably, the recognition phase involves the identification and localization of lung sounds to ensure the quality of the captured data. The recordings are downsampled to 1.6KHz and windowed into 2 second segments, unlike the heartbeat detection which processes 1 second segments of the recordings. Preferably, each segment may contain a complete respiratory cycle, or a portion of the cycle, which may also be sufficient for some screening procedures. Urban noise and generated Gaussian noise are considered positive data, and a convolutional neural network (CNN) with a similar structure to the heart sound recognition model may be used, where the input layer is different.
[0153] Similar to the process in VHD screening, the noise reduction stage can include removing background noise and other artifacts that interfere with the quality of the captured lung sounds. The process is similar to the noise reduction process in heart sound analysis. However, the boundary can be set to 800 Hz instead of 300 Hz based on the characteristics of the lung sound data.
[0154] The classification stage can include classifying lung sounds as normal, rales, or stridor. The recording can be downsampled to 1.6KHz and chopped into respiratory cycles. For example, each cycle can be trimmed or padded to 3 seconds in length, with the average length of the cycle in the training set being approximately 2.8 seconds. A CNN with a structure similar to the heart sound classification model can be used, where the input layer is different to classify the cycle as normal / contains rales / contains stridor. Preferably, hyperparameter tuning can be used to improve accuracy.
[0155] The inventors also tested an embodiment of auscultating the lungs using a respiratory sound database containing 920 recordings from 126 patients with 6898 respiratory cycles. Of these, 1864 contained rales, 886 contained stridor, and 506 contained rales and stridor. The test set was formed by recordings from 26 patients, and the training set was formed by recordings from the other 100 patients. Without observation, the test accuracy of the recognition phase was 91.10%.
[0156] Advantageously, auscultating the lungs can be used to detect breathing abnormalities and screen for airway-related diseases. It can also be used to supplement traditional auscultation techniques and improve the accuracy and reliability of diagnosis. In addition, auscultating the lungs can be used to monitor the progression of lung diseases and adjust treatment plans accordingly.
[0157] Although not required, the embodiments described with reference to the accompanying drawings may be implemented as an application programming interface (API) or a series of libraries used by developers, or may be included in another software application, such as a terminal or personal computer operating system or a portable computing device operating system. Generally, since program modules include routines, programs, objects, components, and data files that help perform specific functions, those skilled in the art will understand that the functionality of a software application may be distributed among multiple routines, objects, or components to achieve the same functionality required herein.
[0158] It should also be understood that in the case where the methods and systems of the present disclosure are implemented in whole or in part by a computing system, any appropriate computing system architecture may be used. This may include tablet computers, wearable devices, smart phones, IoT devices, edge computing devices, stand-alone computers, network computers, cloud-based computing devices, and dedicated hardware devices. Where the terms "computing system" and "computing device" are used, these terms are intended to cover any appropriate configuration of computer hardware that can implement the described functions.
[0159] It will be appreciated by those skilled in the art that many changes and / or modifications may be made to the present disclosure as shown in the specific embodiments without departing from the spirit or scope of the present disclosure as broadly described. Therefore, the present embodiments are to be considered in all aspects as illustrative and not restrictive.
[0160] Unless otherwise indicated, any reference to prior art contained herein should not be taken as an admission that the information is common general knowledge.
Claims
1. A health diagnosis system, characterized in that: include: a signal receiving module for receiving a recording of a sound produced by an organ in an animal when the organ performs a predetermined function for a predetermined period of time; A signal noise reduction module, used to reduce noise signals in the recording; as well as A health diagnosis analysis module is used to analyze the health problems of the animal based on the normal sounds and occasional sounds produced by the organ in the noise-reduced recording.
2. The health diagnosis system according to claim 1, characterized in that: in, The incidental sounds include murmurs and arrhythmia sound patterns produced by the animal's heart.
3. The health diagnosis system according to claim 2, characterized in that: in, The health diagnosis analysis module includes a neural network for classifying the records so as to mark at least one health indicator associated with the health problem.
4. The health diagnosis system according to claim 3, characterized in that: in, The noise reduction module is used to reduce interference caused by the noise signal in the record, thereby improving the prediction accuracy of the health problem provided by the health diagnosis analysis module.
5. The health diagnosis system according to claim 4, characterized in that: in, The neural network includes a lightweight neural network for running on a computing device having a mobile or lightweight processor.
6. The health diagnosis system according to claim 1, characterized in that: in, The signal receiving module is also used to verify normal sounds produced by one or more organs in the animal's body for further processing.
7. The health diagnosis system according to claim 6, characterized in that: in, The signal receiving module includes a signal identification module, which is used to verify the normal sound by identifying the presence of the sound in the recording received by the signal receiving module; and the signal noise reduction module and the health diagnosis analysis module are used to process the sound recording after successful verification of the normal sound in the recording.
8. The health diagnosis system according to claim 7, characterized in that: in, The record includes one or more segments extracted from the record received by the signal receiving module.
9. The health diagnosis system according to claim 7, characterized in that: in, The signal receiving module is also used to extend the sound record for further processing after successfully confirming the normal sound.
10. The health diagnosis system according to claim 9, characterized in that: in, The successful verification is indicated by a positive classification of two consecutive segments, each of the two consecutive segments including the normal sound.
