A health condition diagnosis apparatus based on human voice

Abstract

This invention provides a health status diagnosis device based on human voice, including a human voice acquisition module, a signal processing module, a health diagnosis module, a display and interaction module, and an online learning module. The human voice acquisition module acquires the user's voice signal through a microphone array; the signal processing module filters and amplifies the signal and removes noise outside a preset frequency range; the health diagnosis module receives acoustic features and outputs diagnostic results through a preset model and fusion algorithm; the display and interaction module visualizes the results and provides a user feedback interface; and the online learning module dynamically adjusts model parameters based on user feedback using an adaptive learning algorithm. This invention enables non-invasive and convenient health diagnosis, with high accuracy and environmental adaptability, and can be widely applied to real-time monitoring of various health conditions such as respiratory diseases and cardiovascular diseases.

Description

Technical Field

[0001] This invention relates to the field of medical and health testing technology, and more specifically, to a health status diagnostic device based on human voice. Background Technology

[0002] Currently, there are various methods for diagnosing human health conditions. Traditional diagnostic methods, such as blood tests and imaging examinations, can detect many diseases relatively accurately, but they often suffer from problems such as complex testing processes, the need for specialized equipment and personnel, and high costs. Moreover, these testing methods usually require patients to go to specialized medical institutions such as hospitals, making it impossible to achieve health monitoring anytime, anywhere. With the development of health monitoring technology, people's demand for non-invasive and convenient health diagnostic methods is increasing.

[0003] In recent years, with the development of artificial intelligence and signal processing technologies, assessing health status through the analysis of human physiological signals has become a research hotspot. Among these, the human voice, as an easily accessible physiological signal, contains rich information about human health. Studies have shown a certain correlation between vocal characteristics and health status. However, existing technologies for analyzing health status based on the human voice are not yet mature enough. Currently, there is no convenient and efficient device capable of rapid diagnosis of health status through vocal analysis. On the one hand, most studies focus only on single characteristics of the human voice, such as pitch and timbre, failing to comprehensively assess health status; on the other hand, existing devices often suffer from low accuracy and poor stability, easily affected by environmental noise and individual speaker differences, making them difficult to widely apply in real-life scenarios. For example, in the early stages of common illnesses such as colds and respiratory infections, patients' voices may exhibit subtle changes, but existing detection methods struggle to capture and analyze these changes promptly and accurately to aid in diagnosis.

[0004] CN202510064919.5 discloses a speech recognition-based emotion recognition method, which improves the accuracy of emotion recognition through multi-dimensional feature extraction and enhancement. However, this method is based on speech recognition and cannot recognize sounds such as breathing, ignoring important health-related phonemes such as breathing rhythm, thus limiting its application scenarios. CN202411731250.2 provides a multimodal mental health prediction method, electronic device, and medium, which detects mental health by fusing user micro-expression features, user micro-movement features, user facial features, user emotion categories, and user micro-tone categories. However, this method requires an image processing device, resulting in high costs. CN202410363734.X discloses a psychological assessment method and system based on the BPRS Brief Mental Illness Scale, which converts speech into text using speech recognition technology and then performs semantic analysis on the text to obtain emotion prediction results. However, this method only considers the meaning conveyed by the user's language, ignoring the information carried by the speech itself, including features such as intonation, pauses, and volume.

[0005] In implementing the embodiments of the present invention, the prior art has at least the following problems or defects: On the one hand, existing health diagnosis technologies based on human voice are mostly limited to single feature analysis and cannot comprehensively assess health status; on the other hand, existing devices have poor adaptability in complex environments and are easily affected by environmental noise and individual differences, resulting in low diagnostic accuracy and insufficient stability, making it difficult to meet practical application needs. In addition, the prior art lacks an effective dynamic learning mechanism for model optimization and cannot adjust model parameters in a timely manner based on user feedback to improve diagnostic performance. Summary of the Invention

[0006] This invention provides a health status diagnosis device based on human voice, comprising: The voice acquisition module is configured to acquire the user's voice signal through a microphone array; The signal processing module is connected to the human voice acquisition module and is configured to perform filtering and amplification operations on the human voice signal to remove noise signals outside the preset frequency range; The health diagnosis module is connected to the signal processing module and is configured to receive extracted acoustic features and output health status diagnosis results through a preset health diagnosis model and model fusion algorithm. The display interaction module is connected to the health diagnosis module and is configured to present the health status diagnosis results in a visual form and provide a user feedback interface. The online learning module connects the health diagnosis module and the display interaction module, and is configured to dynamically adjust the parameters of the health diagnosis model based on user feedback information through an adaptive learning algorithm.

[0007] Furthermore, the microphone array consists of four evenly distributed microelectromechanical system (MEMS) microphones, with a sampling range covering a 360-degree spatial area and a sensitivity error controlled within ±2 dB.

[0008] Furthermore, the signal processing module includes: The feature extraction unit is configured to construct a graph structure of the human voice signal using a graph convolutional network, where different frequency components are nodes and frequency correlations are edge relationships; The preprocessing unit is configured to simultaneously process the breath sound signal of the non-vocal segment and the speech vocal segment signal, and adjust the signal amplitude to a preset range through an amplifier.

[0009] Furthermore, the health diagnosis module includes: Base model units include support vector machine classifiers, random forest classifiers, multilayer perceptron neural networks, and recurrent neural networks and their variants; The fusion decision unit is configured to fuse the prediction results of each base model through a decision tree algorithm, thereby reducing the prediction bias of a single model and generating health status diagnosis results.

[0010] Furthermore, the online learning module includes: The data collection unit is configured to continuously collect users' voice data, diagnostic results, and user feedback data. The parameter adjustment unit is configured to adjust the relevant weight parameters using a stochastic gradient descent algorithm when a specific acoustic feature dimension continuously deviates from the range of model parameters.

[0011] Furthermore, the display interaction module includes: The visualization unit is configured to present diagnostic results through text descriptions and color icons, where green icons indicate a healthy state, yellow icons indicate a mildly abnormal state, and red icons indicate a severely abnormal state. The feedback entry unit, integrated into the diagnostic results display interface, is configured to receive written evaluations or optional feedback from users regarding the accuracy of the diagnostic results.

[0012] Furthermore, the health status diagnosis results include a category identifier and severity rating of at least one of the four disease types: respiratory disease, cardiovascular disease, asthma, and depression.

[0013] Furthermore, the feature extraction unit is further configured as follows: By learning the relationship features between nodes through the convolution operation of graph convolutional networks, the changing features of frequency component coupling relationships under different health conditions can be captured. The graph convolutional network processes non-Euclidean structured data and mines signal relationship features that are difficult to capture by traditional algorithms.

[0014] Furthermore, the training process of the fusion decision unit includes: The prediction results of support vector machine, random forest, multilayer perceptron and recurrent neural network on the validation set are combined into a multidimensional feature vector; Using the multidimensional feature vector as input and the real health label as output, a decision tree fusion model is trained, and the decision tree hyperparameters are adjusted through a validation set.

[0015] Furthermore, the preset model of the health diagnosis module is trained through the following steps: Collect human voice sample data labeled with disease type and severity; The sample data is divided into a training set, a validation set, and a test set; The base model is trained using the training set, and its hyperparameters are adjusted using the validation set to prevent overfitting. The performance of the final model is evaluated using a test set.

[0016] The embodiments of the present invention have at least the following beneficial effects: 1. By using the microphone array acquisition and signal processing module's filtering and noise reduction technology, the problem of environmental noise interference is effectively solved. It can accurately extract human voice characteristics in complex acoustic environments, thereby improving the reliability and stability of health diagnosis.

[0017] 2. By constructing graph structure features of human voice signals using graph convolutional networks and combining them with multi-model fusion algorithms, the limitations of traditional single feature analysis are overcome. This approach can comprehensively capture changes in acoustic features under different health conditions, significantly improving the accuracy of diagnosis for respiratory diseases, cardiovascular diseases, and other conditions.

[0018] 3. Through the adaptive parameter adjustment and user feedback mechanism of the online learning module, the diagnostic model can be continuously optimized to adapt to the individual pronunciation differences of different users, which solves the problem of insufficient generalization ability of traditional detection methods and realizes the personalization and intelligence of long-term health monitoring. Attached Figure Description

[0019] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein: Figure 1 This is a schematic diagram of the structure of a health status diagnosis device based on human voice provided in an embodiment of the present invention. Detailed Implementation

[0020] The principles and spirit of the invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided to make the invention more thorough and complete, and to fully convey the scope of the invention to those skilled in the art.

[0021] Those skilled in the art will recognize that embodiments of the present invention can be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0022] It should be noted that the number of any elements in the accompanying drawings is for illustrative purposes only and not as a limitation, and any naming is for distinction only and has no limiting meaning.

[0023] The following is for reference. Figure 1 , Figure 1 This is a schematic diagram of a health status diagnosis device based on human voice provided in an embodiment of the present invention. Figure 1 As shown, a health status diagnosis device based on human voice includes: The voice acquisition module 101 is configured to acquire the user's voice signal through a microphone array; The signal processing module 102 is connected to the human voice acquisition module and is configured to perform filtering and amplification operations on the human voice signal to remove noise signals outside the preset frequency range; The health diagnosis module 103 is connected to the signal processing module and is configured to receive extracted acoustic features and output health status diagnosis results through a preset health diagnosis model and model fusion algorithm. The display interaction module 104 is connected to the health diagnosis module and is configured to present the health status diagnosis results in a visual form and provide a user feedback interface. The online learning module 105 connects the health diagnosis module and the display interaction module, and is configured to dynamically adjust the parameters of the health diagnosis model based on user feedback information through an adaptive learning algorithm.

[0024] It's important to note that the voice acquisition module is the foundation of the entire device. Its function is to acquire the user's voice signals for subsequent processing and analysis. A microphone array is a device composed of multiple microphones that can simultaneously receive sound from multiple directions, thereby improving the accuracy and completeness of voice acquisition. Voice signals refer to the sounds emitted by the user, including speaking, coughing, breathing, and other health-related vocalizations. Acquiring voice signals through a microphone array provides raw data for subsequent health diagnosis.

[0025] Specifically, the microphone array consists of four evenly distributed MEMS microphones, covering a 360-degree spatial area with a sensitivity error controlled within ±2 dB. MEMS microphones are miniaturized microphones with advantages such as small size, low power consumption, and high sensitivity. The four MEMS microphones, evenly distributed around the device, form a 360-degree acquisition range, capable of capturing the user's voice from all directions, accurately recording it regardless of the user's direction of voice. Sensitivity error refers to the deviation between the microphone's response intensity to the sound signal and the actual sound intensity. Controlling it within ±2 dB means that the microphone can accurately reflect changes in sound intensity, thus ensuring the quality of the acquired voice signal.

[0026] Preferably, in practical applications, the microphone array parameters of the voice acquisition module can be adjusted according to different usage scenarios and needs. For example, in noisy environments, the microphone sensitivity can be increased to improve the ability to acquire voice signals; in quiet environments, the sensitivity can be appropriately reduced to minimize noise interference. Furthermore, the acquisition range and direction of the microphone array can be optimized based on the user's vocal characteristics, such as volume and pitch, to further improve the acquisition effect.

[0027] In some embodiments, the microphone array consists of four uniformly distributed microelectromechanical system (MEMS) microphones, with a sampling range covering a 360-degree spatial area and a sensitivity error controlled within ±2 dB.

[0028] It should be noted that the microphone array of this device consists of four evenly distributed MEMS microphones. This design aims to achieve omnidirectional sound acquisition, ensuring the device can capture the user's voice signal from all directions. MEMS microphones are miniature microphones based on MEMS technology, offering advantages such as small size, high sensitivity, and low power consumption, making them suitable for portable devices and applications requiring high-sensitivity sound acquisition. A 360-degree acquisition range means the microphone array can receive sound signals from all directions without blind spots, while sensitivity error controlled within ±2 dB ensures the microphone's accuracy in acquiring sound signals, avoiding sound signal distortion caused by differences in microphone sensitivity.

[0029] Specifically, a MEMS microphone is a miniature microphone manufactured using MEMS technology. Its working principle involves using miniature mechanical structures to convert sound signals into electrical signals. The four microphones are evenly distributed, meaning they are arranged at equal intervals in space, such as in a square or circular layout, thus achieving uniform coverage of 360 degrees. This arrangement ensures that the microphone array receives the signal with the same sensitivity regardless of the user's direction of voice. Sensitivity error refers to the deviation between the microphone's response intensity to sound signals at different frequencies and the theoretical value. A tolerance of ±2 dB is a relatively strict specification, meaning that the microphone's response differences at different frequencies are strictly controlled, thereby ensuring the consistency of the collected human voice signal across different directions and frequencies.

[0030] Preferably, to further optimize the performance of the microphone array, the layout and parameters of the microphones can be finely adjusted during the design phase. For example, the sensitivity range of the microphones can be adjusted according to the actual usage scenario, such as indoor or outdoor, to adapt to different environmental noise levels. In an indoor environment, the sensitivity can be appropriately reduced to decrease background noise interference; while in an outdoor environment, the sensitivity can be increased to ensure clear capture of the user's voice. Furthermore, the acquired signals can be calibrated using software algorithms to further compensate for sensitivity differences between microphones, ensuring the consistency and accuracy of the acquired voice signals in all directions.

[0031] In some embodiments, the signal processing module includes: The feature extraction unit is configured to construct a graph structure of the human voice signal using a graph convolutional network, where different frequency components are nodes and frequency correlations are edge relationships; The preprocessing unit is configured to simultaneously process the breath sound signal of the non-vocal segment and the speech vocal segment signal, and adjust the signal amplitude to a preset range through an amplifier.

[0032] It should be noted that the signal processing module is a crucial part of the device used for preprocessing and feature extraction of the acquired human voice signals. Its main functions are filtering and amplifying the human voice signals, removing noise signals outside the preset frequency range, and extracting acoustic features. The feature extraction unit uses a Graph Convolutional Network (GCN) to construct the graph structure of the human voice signal. This is an advanced signal processing technique capable of uncovering the complex relationships between different frequency components in the human voice signal. The preprocessing unit is responsible for simultaneously processing the non-vocal segment respiratory sound signals and the vocal segment signals, and adjusting the signal amplitude to the preset range through amplifiers to ensure signal stability and consistency, providing high-quality input data for subsequent health diagnosis.

[0033] Specifically, the feature extraction unit in the signal processing module processes the human voice signal using a Graph Convolutional Network (GCN). A GCN is a neural network used to process graph-structured data. It treats different frequency components of the human voice signal as nodes in a graph, and the correlations between frequency components as edge relationships between nodes. This graph structure can more intuitively represent the interactions between various frequency components in the human voice signal. The preprocessing unit is responsible for filtering and amplifying the raw human voice signal. The filtering operation removes noise from the signal that is not within the human voice frequency range, such as low-frequency current noise and high-frequency electromagnetic interference. The amplifier adjusts the signal amplitude to a range suitable for subsequent processing, for example, adjusting the peak amplitude of the signal to about 1 volt to ensure signal stability and consistency.

[0034] Preferably, when using a graph convolutional network, the feature extraction unit first decomposes the human voice signal into multiple frequency components and calculates the correlation between these components to construct a graph structure. Then, it learns the relationship features between nodes through graph convolution operations to capture changes in the coupling relationships of frequency components under different health conditions. For example, when suffering from respiratory diseases, the coupling relationships of certain frequency components in the human voice signal may change, and the graph convolutional network can effectively extract these changing features. When performing filtering operations, the preprocessing unit can set the filter parameters according to the frequency range of the human voice signal, for example, using a bandpass filter with a passband range of 85 Hz to 255 Hz to remove noise signals outside this range. The amplifier gain can be dynamically adjusted according to the initial amplitude of the signal to ensure that the amplitude of the amplified signal conforms to a preset range.

[0035] In some embodiments, the health diagnostic module includes: Base model units include support vector machine classifiers, random forest classifiers, multilayer perceptron neural networks, and recurrent neural networks and their variants; The fusion decision unit is configured to fuse the prediction results of each base model through a decision tree algorithm, thereby reducing the prediction bias of a single model and generating health status diagnosis results.

[0036] It's important to note that the health diagnosis module is the core component for diagnosing health conditions based on human voice signal characteristics. It analyzes the extracted acoustic features using multiple base models and then integrates the predictions from each base model using a fusion decision unit to generate the final health condition diagnosis. The base model unit includes various machine learning and deep learning models, such as Support Vector Machines (SVM), Random Forest, Multilayer Perceptron Neural Networks (MLP), and Recurrent Neural Networks (RNN) and their variants. Each of these models has its advantages, enabling analysis of human voice signal characteristics from different perspectives. The fusion decision unit then uses a decision tree algorithm to fuse the predictions from each base model, reducing the prediction bias of a single model and improving the accuracy and stability of the diagnosis.

[0037] Specifically, the base model units in the health diagnosis module include the following models: Support Vector Machine (SVM), a classification model based on statistical learning theory, separates data of different categories by finding the optimal hyperplane, suitable for classification problems with small samples and high-dimensional data; Random Forest, an ensemble learning model composed of multiple decision trees, constructs multiple decision trees through random sampling and feature selection, and combines their prediction results, exhibiting strong anti-overfitting ability and the ability to handle complex data relationships; Multilayer Perceptron Neural Network (MLP), a feedforward neural network that achieves complex feature mapping through the nonlinear combination of multiple layers of neurons, suitable for handling complex nonlinear relationships; Recurrent Neural Networks (RNNs) and their variants, such as LSTM and GRU, can process sequence data, particularly suitable for analyzing time-series features in human voice signals, and can capture the temporal variation patterns of sound signals. The fusion decision unit uses a decision tree algorithm to fuse the prediction results of each base model. By constructing a new decision tree model, using the prediction results of each base model as input features and the real health label as output, a fusion model that integrates the advantages of each base model is trained.

[0038] Preferably, the construction process of the health diagnosis module can be divided into the following steps: First, collect a large amount of voice sample data labeled with health conditions, including various conditions such as health, colds, respiratory diseases, and cardiovascular diseases. Then, divide this data into training sets, validation sets, and test sets. Next, train and optimize each base model. For example, for support vector machines, it is necessary to select an appropriate kernel function, such as radial basis functions, and adjust kernel parameters; for random forests, it is necessary to determine the number and depth of decision trees; for multilayer perceptrons, it is necessary to design the network structure, such as the number of layers and the number of neurons in each layer, and adjust the learning rate; for recurrent neural networks, it is necessary to select an appropriate network structure, such as LSTM or GRU, and adjust parameters such as sequence length. After the training of each base model is completed, the prediction results on the validation set are used to construct the input features of the fusion model, and the fusion model is trained using a decision tree algorithm. Finally, the performance of the fusion model is evaluated using a test set to ensure its diagnostic accuracy and stability under different health conditions.

[0039] In some embodiments, the online learning module includes: The data collection unit is configured to continuously collect users' voice data, diagnostic results, and user feedback data. The parameter adjustment unit is configured to adjust the relevant weight parameters using a stochastic gradient descent algorithm when a specific acoustic feature dimension continuously deviates from the range of model parameters.

[0040] It should be noted that the online learning module of this device is a key component for achieving dynamic model optimization. It continuously collects user voice data, diagnostic results, and user feedback data, and uses adaptive learning algorithms to adjust the parameters of the health diagnostic model. This process enables the model to dynamically adjust based on actual user feedback, thereby improving the accuracy and adaptability of the diagnosis. The data collection unit is responsible for collecting various types of data, while the parameter adjustment unit adjusts the weight parameters using a stochastic gradient descent algorithm based on specific conditions, such as when the acoustic feature dimension deviates from the model parameter range, to optimize model performance.

[0041] Specifically, the online learning module includes a data collection unit and a parameter tuning unit. The data collection unit continuously collects users' voice data, diagnostic results, and user feedback on those results. This data serves as the basis for model optimization. The parameter tuning unit is responsible for adjusting the model parameters based on the collected data. When the model detects that a specific acoustic feature dimension continuously deviates from the range of the model parameters, it indicates that the model's prediction on that feature may be inaccurate. At this point, the parameter tuning unit will initiate the stochastic gradient descent algorithm to adjust the relevant weight parameters. Stochastic gradient descent is a commonly used optimization algorithm that gradually adjusts model parameters by calculating the gradient of the loss function to minimize prediction error.

[0042] Preferably, the data collection unit can be implemented in various ways in practical applications. For example, it can be integrated with the user interface to receive user feedback in real time; or it can upload the user's voice data and diagnostic results to a server for centralized processing via a network connection. When using the stochastic gradient descent algorithm, the parameter adjustment unit needs to set some key parameters, such as the learning rate. The learning rate determines the step size of each parameter update; too large a rate will lead to model instability, while too small a rate will result in slow model convergence. For example, the learning rate can be set to 0.01 and dynamically adjusted according to the model's convergence.

[0043] Furthermore, a momentum term can be introduced to accelerate convergence and avoid getting trapped in local minima. When adjusting parameters, weight adjustment strategies can be set separately for different acoustic feature dimensions. For example, different learning rates or adjustment frequencies can be used for high-frequency features and low-frequency features to better adapt to the changing patterns of different features.

[0044] In some embodiments, the display interaction module includes: The visualization unit is configured to present diagnostic results through text descriptions and color icons, where green icons indicate a healthy state, yellow icons indicate a mildly abnormal state, and red icons indicate a severely abnormal state. The feedback entry unit, integrated into the diagnostic results display interface, is configured to receive written evaluations or optional feedback from users regarding the accuracy of the diagnostic results.

[0045] It's important to note that the display and interaction module is the core component of the device for user interaction. It presents the health status diagnosis results output by the health diagnosis module in a visual format and provides a feedback interface for users to evaluate the results. The visualization unit intuitively displays health status through text descriptions and color icons, with green, yellow, and red representing healthy, mild, and severe abnormalities, respectively. This intuitive display method facilitates quick understanding of the diagnosis results. The feedback entry unit allows users to provide feedback on the accuracy of the diagnosis results; this feedback information will be used to optimize the health diagnosis model.

[0046] Specifically, the display and interaction module includes a visualization unit and a feedback entry unit. The visualization unit transforms complex health diagnostic results into a user-friendly format, such as using text descriptions and color icons to represent health status. A green icon represents a healthy state, indicating the user's current health is good; a yellow icon represents a mild abnormality, suggesting potential health issues but not requiring excessive concern; and a red icon represents a severe abnormality, advising the user to seek medical attention as soon as possible. The feedback entry unit is integrated into the diagnostic results display interface, allowing users to express their satisfaction or assessment of the diagnostic results through written evaluations or by selecting feedback options. This feedback mechanism not only enhances the interactivity between the user and the device but also provides crucial information for model optimization.

[0047] Preferably, to further enhance the user experience and functionality of the display interaction module, more details can be added to the visualization unit. For example, in addition to text descriptions and color icons, charts such as bar charts or line charts can be used to display the changing trends of the user's health status, helping users better understand the dynamic changes in their own health status. In the design of the feedback entry unit, more specific feedback options can be provided, such as allowing users to select options like "the diagnosis result is very accurate," "the diagnosis result is basically accurate," or "the diagnosis result is inaccurate," and allowing users to input specific feedback.

[0048] In addition, a help button can be added to provide users with guidance on how to interpret diagnostic results and how to provide feedback. Through these optimizations, the display interaction module can not only present diagnostic results more intuitively but also better collect user feedback, thereby providing richer data support for the continuous optimization of the health diagnostic model.

[0049] In some embodiments, the health status diagnosis results include a category identifier and severity rating of at least one of four disease types: respiratory disease, cardiovascular disease, asthma, and depression.

[0050] It should be noted that the health status diagnosis results include a category identifier and severity rating for at least one of four disease types: respiratory diseases, cardiovascular diseases, asthma, and depression. This indicates that the device can classify and assess a variety of common health problems, not only determining the presence of a disease but also quantifying its severity. This comprehensive diagnostic capability allows the device to provide users with more specific health information, helping them better understand their own health status and take appropriate measures.

[0051] Specifically, the category labels in the health status diagnosis results refer to the clear classification of disease types. For example, respiratory diseases include colds, chronic obstructive pulmonary disease (COPD), cardiovascular diseases include hypertension, coronary heart disease, asthma is a chronic inflammatory airway disease, and depression is a common mental illness. These category labels allow users to quickly understand the type of disease they may have. Severity grading, on the other hand, is a quantitative assessment of the severity of the disease, typically divided into three levels: mild, moderate, and severe. For example, a mild respiratory disease might manifest as a slight cough, while a severe one might be accompanied by symptoms such as difficulty breathing. This grading method helps users more intuitively understand the severity of their illness and provides a basis for subsequent medical decisions.

[0052] Preferably, to achieve accurate health status diagnosis results, the device needs to train on a large amount of voice sample data labeled with disease types and severity when constructing the health diagnosis model. For example, during training, features of the voice signal, such as frequency, amplitude, and tone, can be correlated with the corresponding disease type and severity. For respiratory diseases, the model can determine the presence of the disease by analyzing changes in the frequency of breathing sounds; for cardiovascular diseases, it can be assessed by analyzing the amplitude and rhythm of the sound. Regarding severity grading, the model can perform quantitative assessment based on the degree of abnormality of features, for example, by setting thresholds to distinguish between mild, moderate, and severe conditions.

[0053] Furthermore, to improve diagnostic accuracy, expert knowledge can be incorporated to calibrate and optimize the model's output. For example, doctors can be invited to review the model's diagnostic results to ensure they meet clinical diagnostic criteria.

[0054] In some embodiments, the feature extraction unit is further configured to: By learning the relationship features between nodes through the convolution operation of graph convolutional networks, the changing features of frequency component coupling relationships under different health conditions can be captured. The graph convolutional network processes non-Euclidean structured data and mines signal relationship features that are difficult to capture by traditional algorithms.

[0055] It should be noted that the feature extraction unit processes the human voice signal using a Graph Convolutional Network (GCN) to construct a graph structure of the voice signal, where different frequency components are nodes and frequency correlations are edges. This graph structure can capture the changing characteristics of frequency component coupling relationships under different health conditions, thereby uncovering signal relationship features that are difficult for traditional algorithms to capture. A Graph Convolutional Network is a neural network specifically designed for processing graph-structured data. It can handle non-Euclidean data and is suitable for feature extraction from human voice signals with complex relationships.

[0056] Specifically, Graph Convolutional Networks (GCNs) are a type of neural network based on a graph structure. They treat different frequency components of a human voice signal as nodes in a graph, and the correlations between these frequency components as edge relationships between nodes. This graph structure can more intuitively represent the interactions between various frequency components in a human voice signal. For example, if two frequency components frequently appear simultaneously or influence each other during vocalization, their edge relationship is strong. Through graph convolution operations, GCNs can learn the relationship features between nodes and capture changes in the coupling relationships of frequency components under different health conditions. For example, in patients with cardiovascular disease, the coupling relationship between certain low-frequency and high-frequency components of the human voice may change, and GCNs can effectively extract this graph structure feature. This approach can uncover signal relationship-based features that are difficult for traditional algorithms to capture, providing a new feature dimension for health diagnosis.

[0057] Preferably, when using a graph convolutional network, the feature extraction unit first needs to preprocess the human voice signal to extract its frequency components. For example, the human voice signal is transformed from the time domain to the frequency domain using Fourier transform to obtain the amplitude and phase information of different frequency components. Then, a graph structure is constructed based on the correlation between the frequency components. The correlation can be determined by calculating the cross-correlation coefficient or covariance between the frequency components. Next, the graph convolutional network learns the relationship features between nodes through convolution operations. During training, the network input is the constructed graph structure, and the output is the extracted feature vector. These feature vectors can reflect the changing patterns of the human voice signal under different health conditions. For example, for the human voice signal of someone suffering from respiratory diseases, its feature vector may show obvious anomalies in the coupling relationships of certain frequency components. In this way, the feature extraction unit can provide high-quality feature input for the subsequent health diagnosis module.

[0058] In some embodiments, the training process of the fusion decision unit includes: The prediction results of support vector machine, random forest, multilayer perceptron and recurrent neural network on the validation set are combined into a multidimensional feature vector; Using the multidimensional feature vector as input and the real health label as output, a decision tree fusion model is trained, and the decision tree hyperparameters are adjusted through a validation set.

[0059] It's important to note that the training process of the fusion decision unit is a crucial step in the health diagnosis module. Its purpose is to generate more accurate health status diagnoses by fusing the predictions from multiple base models. This process involves combining the predictions from base models such as Support Vector Machines (SVM), Random Forests (Random Forest), Multilayer Perceptrons (MLP), and Recurrent Neural Networks (RNN) on the validation set into a multidimensional feature vector, which is then used as input to train the decision tree fusion model. In this way, the fusion decision unit can reduce the prediction bias of individual models, improving the overall accuracy and stability of the diagnosis.

[0060] Specifically, the training process of the fusion decision unit includes the following key steps. First, the prediction results of various base models, such as SVM, Random Forest, MLP, and RNN, on the validation set are combined into a multi-dimensional feature vector. These prediction results reflect the judgments of different models on the same input data, and each model's prediction result can be regarded as a feature dimension. For example, if the SVM prediction result is 0.8 (representing the probability of disease), the Random Forest prediction result is 0.7, the MLP prediction result is 0.6, and the RNN prediction result is 0.9, then this multi-dimensional feature vector can be represented as [0.8, 0.7, 0.6, 0.9]. Next, using this multi-dimensional feature vector as input and the true health label as output, the decision tree fusion model is trained. A decision tree is a tree-based classification model that classifies samples through a series of conditional judgments. Each internal node is a test on an attribute, the branch is the test output, and the leaf node is the category. During training, the hyperparameters of the decision tree, such as the tree depth and the minimum number of samples for splitting, are adjusted using the validation set to optimize model performance.

[0061] Preferably, the training process can be further refined when constructing the fusion decision unit. First, the prediction results of each base model are standardized to ensure they are on the same scale, for example, by normalizing all prediction results to between 0 and 1. This avoids the possibility that the prediction results of some models, due to their large numerical range, will have an excessive impact on the fusion model. Second, when training the decision tree fusion model, cross-validation can be used to select the optimal hyperparameters. For example, the validation set can be divided into multiple subsets, and the model performance under different hyperparameter combinations can be evaluated through multiple training and validation cycles, ultimately selecting the hyperparameters with the best performance.

[0062] Furthermore, regularization techniques can be introduced to prevent overfitting in decision tree models, such as using pruning algorithms to reduce tree complexity. Through these optimization steps, the fusion decision unit can more effectively integrate the prediction results of multiple base models, thereby improving the accuracy and robustness of health diagnosis.

[0063] In some embodiments, the preset model of the health diagnosis module is trained through the following steps: Collect human voice sample data labeled with disease type and severity; The sample data is divided into a training set, a validation set, and a test set; The base model is trained using the training set, and its hyperparameters are adjusted using the validation set to prevent overfitting. The performance of the final model is evaluated using a test set.

[0064] It's important to note that the pre-training process of the health diagnosis module is a crucial step in realizing the health status diagnosis function based on human voice signals. This process involves collecting human voice sample data labeled with disease types and severity, and dividing it into training, validation, and test sets. These sets are used to train the base model, adjust hyperparameters, and evaluate the final model's performance, respectively. This process ensures that the model can learn the acoustic characteristic patterns under different health conditions from a large amount of data, thereby providing users with accurate health diagnosis suggestions.

[0065] Specifically, the training process of the health diagnosis module involves the following key steps. First, a large amount of voice sample data labeled with disease types and severity is collected. This sample data includes voice signals under various health conditions, such as health, cold, respiratory diseases, and cardiovascular diseases. The labeling information refers to the health condition label corresponding to each sample, such as healthy, mild cold, severe cardiovascular disease, etc. Next, this sample data is divided into training, validation, and test sets. The training set is used to train base models, such as Support Vector Machines (SVM), Random Forests, Multilayer Perceptrons (MLP), and Recurrent Neural Networks (RNNs). These models build predictive models by learning the relationship between acoustic features in the training set and health conditions. The validation set is used to adjust the hyperparameters of the base models, such as the kernel function parameters of SVM, the number of trees in a random forest, and the number of hidden layer neurons in an MLP, to prevent overfitting. The test set is used to evaluate the performance of the final model, measuring its diagnostic ability by calculating metrics such as accuracy, recall, and F1 score.

[0066] Preferably, to further refine the training process of the health diagnosis module, the following measures can be taken: During the data collection phase, ensure the diversity and representativeness of the sample data, such as covering users of different age groups, genders, and regional accents, to improve the model's generalization ability. During the model training phase, different optimization algorithms and loss functions can be used for each base model. For example, for SVM, the Sequence Minimum Optimization (SMO) algorithm can be used for training; for MLP, the Stochastic Gradient Descent (SGD) algorithm can be used, with appropriate learning rate and momentum parameters set. During the hyperparameter tuning phase, methods such as grid search or Bayesian optimization can be used to find the optimal combination of hyperparameters. For example, for random forests, different numbers of trees, such as 100, 200, and 300, and maximum depths, such as 5, 10, and 15, can be tried to find the best performance. During the model evaluation phase, in addition to commonly used accuracy, recall, and F1 scores, a confusion matrix can be introduced to more comprehensively analyze the model's classification performance, especially for multi-class problems, such as the diagnosis of multiple disease types. Through these optimization measures, the accuracy and reliability of the health diagnosis module can be significantly improved.

[0067] The above embodiments of the present invention have the following beneficial effects: 1. By using the microphone array acquisition and signal processing module's filtering and noise reduction technology, the problem of environmental noise interference is effectively solved. It can accurately extract human voice characteristics in complex acoustic environments, thereby improving the reliability and stability of health diagnosis.

[0068] 2. By constructing graph structure features of human voice signals using graph convolutional networks and combining them with multi-model fusion algorithms, the limitations of traditional single feature analysis are overcome. This approach can comprehensively capture changes in acoustic features under different health conditions, significantly improving the accuracy of diagnosis for respiratory diseases, cardiovascular diseases, and other conditions.

[0069] 3. Through the adaptive parameter adjustment and user feedback mechanism of the online learning module, the diagnostic model can be continuously optimized to adapt to the individual pronunciation differences of different users, which solves the problem of insufficient generalization ability of traditional detection methods and realizes the personalization and intelligence of long-term health monitoring.

[0070] Furthermore, the storage medium in the embodiments of this application stores program instructions capable of implementing all the above methods. These program instructions can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0071] The above description is merely an explanation of some preferred embodiments of the present invention and the technical principles employed. Those skilled in the art should understand that the scope of the invention as described in the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A health status diagnosis device based on human voice, characterized in that, include: The voice acquisition module is configured to acquire the user's voice signal through a microphone array; The signal processing module is connected to the human voice acquisition module and is configured to perform filtering and amplification operations on the human voice signal, remove noise signals outside the preset frequency range, and extract acoustic features. The health diagnosis module is connected to the signal processing module and is configured to receive extracted acoustic features and output health status diagnosis results through a preset health diagnosis model and model fusion algorithm. The display interaction module is connected to the health diagnosis module and is configured to present the health status diagnosis results in a visual form and provide a user feedback interface. The online learning module connects the health diagnosis module and the display interaction module, and is configured to dynamically adjust the parameters of the health diagnosis model based on user feedback information through an adaptive learning algorithm.

2. The apparatus according to claim 1, characterized in that, The microphone array consists of four evenly distributed microelectromechanical system (MEMS) microphones, covering a 360-degree spatial area, with a sensitivity error controlled within ±2 dB.

3. The apparatus according to claim 1, characterized in that, The signal processing module includes: The feature extraction unit is configured to construct a graph structure of the human voice signal using a graph convolutional network, where different frequency components are nodes and frequency correlations are edge relationships; The preprocessing unit is configured to simultaneously process the breath sound signal of the non-vocal segment and the speech vocal segment signal, and adjust the signal amplitude to a preset range through an amplifier.

4. The apparatus according to claim 1, characterized in that, The health diagnosis module includes: Base model units include support vector machine classifiers, random forest classifiers, multilayer perceptron neural networks, and recurrent neural networks and their variants; The fusion decision unit is configured to fuse the prediction results of each base model through a decision tree algorithm, thereby reducing the prediction bias of a single model and generating health status diagnosis results.

5. The apparatus according to claim 1, characterized in that, The online learning module includes: The data collection unit is configured to continuously collect users' voice data, diagnostic results, and user feedback data. The parameter adjustment unit is configured to adjust the relevant weight parameters using a stochastic gradient descent algorithm when a specific acoustic feature dimension continuously deviates from the range of model parameters.

6. The apparatus according to claim 1, characterized in that, The display interaction module includes: The visualization unit is configured to present diagnostic results through text descriptions and color icons, where green icons indicate a healthy state, yellow icons indicate a mildly abnormal state, and red icons indicate a severely abnormal state. The feedback entry unit, integrated into the diagnostic results display interface, is configured to receive written evaluations or optional feedback from users regarding the accuracy of the diagnostic results.

7. The apparatus according to claim 1, characterized in that, The health status diagnosis results include a category identifier and severity rating for at least one of the four disease types: respiratory disease, cardiovascular disease, asthma, and depression.

8. The apparatus according to claim 3, characterized in that, The feature extraction unit is further configured as follows: By learning the relationship features between nodes through the convolution operation of graph convolutional networks, the changing features of frequency component coupling relationships under different health conditions can be captured. The graph convolutional network processes non-Euclidean structured data and mines signal relationship features that are difficult to capture by traditional algorithms.

9. The apparatus according to claim 4, characterized in that, The training process of the fusion decision unit includes: The prediction results of support vector machine, random forest, multilayer perceptron and recurrent neural network on the validation set are combined into a multidimensional feature vector; Using the multidimensional feature vector as input and the real health label as output, a decision tree fusion model is trained, and the decision tree hyperparameters are adjusted through a validation set.

10. The apparatus according to claim 1, characterized in that, The preset model of the health diagnosis module is trained through the following steps: Collect human voice sample data labeled with disease type and severity; The sample data is divided into a training set, a validation set, and a test set; The base model is trained using the training set, and its hyperparameters are adjusted using the validation set to prevent overfitting. The performance of the final model is evaluated using a test set.

Images