Intelligent wearable device for breast health monitoring and data analysis system and analysis method

Abstract

The invention discloses intelligent wearable equipment for breast health monitoring and a data analysis system and method. The equipment comprises a plurality of wearable ultrasonic patches, a wireless communication module, a signal processing and control unit and a power management module. The data analysis system comprises a data acquisition system, a data transmission system, a data processing and analysis system and a user interaction system, and the data acquisition system is used for acquiring ultrasonic images of mammary tissue and physiological parameters of a mammary area; the data processing and analysis system comprises a data preprocessing module, a feature extraction module and a multi-modal deep learning algorithm module, data preprocessing, feature extraction and data fusion are achieved, and the user interaction system comprises a data visualization system and an early warning notification module. According to the invention, ultrasonic images and various physiological parameters of the mammary gland can be collected in real time, long-term, repeatable and non-invasive monitoring can be carried out on the mammary gland tissue, and early warning and risk assessment of mammary gland diseases can be realized.

Description

Technical Field

[0001] The present invention relates to intelligent devices and data analysis methods, and particularly to an intelligent wearable device for breast health monitoring, a data analysis system, and an analysis method. Background Art

[0002] Breast cancer is one of the most common malignant tumors among women globally, and early detection is crucial for improving the survival rate and quality of life of patients. Current breast health monitoring methods mainly include:

[0003] · Breast self-examination and clinical palpation: Dependent on the experience of individuals and doctors, with low sensitivity and specificity, and prone to missing early tiny lesions.

[0004] · Mammography: It is a standard screening method, but there is a radiation risk, the sensitivity is reduced for women with dense breasts, and the examination frequency is limited.

[0005] · Ultrasound examination: Suitable for dense breasts, but requires professional operation and cannot achieve continuous and real-time monitoring.

[0006] · Magnetic resonance imaging (MRI): Provides high-resolution images, but is costly and time-consuming, and not suitable for large-scale screening.

[0007] Existing wearable devices are mostly used for general health monitoring. There are few devices for breast health, and they mainly monitor parameters such as surface temperature, lacking direct monitoring of the internal breast tissue.

[0008] Therefore, there is an urgent need in the market for a wearable, non-invasive, and sustainable breast health monitoring system. Summary of the Invention

[0009] Object of the Invention: The object of the present invention is to provide an intelligent wearable device for breast health monitoring, a data analysis system, and an analysis method. Through a breast health monitoring system based on wearable ultrasound technology, long-term, repeatable, and non-invasive monitoring of breast tissue is carried out, so as to achieve early warning and risk assessment of breast diseases (such as breast cancer, breast hyperplasia, fibroadenoma, etc.). The system is designed as a flexible ultrasound patch that can be attached to the inner side of a bra, capable of real-time collecting ultrasound images of the breast and various physiological parameters such as temperature and pressure. By integrating multi-modal deep learning algorithms, the system can comprehensively analyze the morphological characteristics and physiological parameters of breast tissue, and combine temporal information to achieve early warning and risk assessment of breast diseases (such as breast cancer, breast hyperplasia, fibroadenoma, etc.).

[0010] Technical solution: On the one hand, the intelligent wearable device for breast health monitoring provided by the present invention includes a plurality of wearable ultrasonic patches, a wireless communication module, a signal processing and control unit, and a power management module. An outer shell is provided outside. The plurality of wearable ultrasonic patches are connected by flexible connection bands and are respectively communicatively or electrically connected to the wireless communication module, the signal processing and control unit, and the power management module.

[0011] Further, the wearable ultrasonic patch includes an upper non-woven fabric, an intermediate flexible circuit board, and a lower conductive gel. The flexible circuit board integrates ultrasonic and physiological parameter sensors, and the flexible circuit board serves as the substrate of the sensor array and forms an ultrasonic transducer with the ultrasonic sensor.

[0012] Further, N×M elements are designed to form a two-dimensional matrix to cover the high-incidence areas of the breast. The size and spacing d of each element satisfy the Nyquist sampling theorem to avoid spatial aliasing, that is where λ is the ultrasonic wavelength, and the operating frequency of the ultrasonic transducer is set to f 0 = 7 - 15 MHz.

[0013] Further, the physiological parameter sensors are temperature sensors and pressure sensors.

[0014] On the other hand, the present invention provides a data analysis system for the intelligent wearable device for breast health monitoring, including a data acquisition system, a data transmission system, a data processing and analysis system, and a user interaction system. The data acquisition system is used to acquire ultrasonic images of breast tissue and physiological parameters of the breast region. The data transmission system realizes real-time data transmission. The data processing and analysis system includes a data preprocessing module, a feature extraction module, and a multimodal deep learning algorithm module to realize data preprocessing, feature extraction, and data fusion. The user interaction system includes a data visualization system and a warning notification module.

[0015] On the third aspect, the present invention provides an analysis method for the data analysis system of the intelligent wearable device for breast health monitoring, including the following steps:

[0016] (1) Data processing;

[0017] (2) Feature extraction;

[0018] (3) Multimodal data fusion;

[0019] (4) Multimodal deep learning algorithm.

[0020] Further, the processing method of step (1):

[0021] 1.1 Ultrasonic image preprocessing

[0022] d) Denoising processing: Median filtering is used to remove noise;

[0023] e) Image enhancement: Histogram equalization is used to improve the image contrast;

[0024] f) Size normalization: The image is adjusted to a unified size of M×N pixels, and bilinear interpolation is used for image scaling. The scaling factor is calculated as follows: where H and W are the height and width of the original image, and the value of each pixel g(i, j) of the new image is calculated using the following formula

[0025] 1.2 Physiological parameter preprocessing

[0026] a) Filtering processing: A low-pass filtering method is used to remove high-frequency noise in the signal. For discrete signals, the output of the low-pass filter can be obtained through convolution operation. A discrete smoothing filter is used:

[0027] where h(k) is the impulse response of the filter, K is the window size of the filter, that is, the number of sampling points covered by the filter impulse response h[n] determines the window size of the filter and affects the degree of smoothing. k is the index variable in the convolution operation, which is used to traverse the impulse response of the filter;

[0028] b) Scale the physiological parameter values to the range [0, 1] according to a ratio. The formula is: x is the original physiological parameter value, x max and x min are the maximum and minimum values of this physiological parameter, respectively.

[0029] Furthermore, the processing method in step (2):

[0030] 2.1 Ultrasonic image morphological extraction

[0031] By using a convolutional neural network, high-level features in the image, such as the shape, size, edge sharpness, and texture features of the mass, are automatically extracted. Input: The preprocessed ultrasonic image with a size of m×n. The specific steps are as follows: First, convolution operation is used to extract local features, such as edges and corner points, and further convolution operations are performed: Among them, O(i, j) is the pixel value in the convolutional output feature map, W(k, l) is the convolutional kernel, I(i + k, j + l) is the pixel value of the input image, b is the bias. The output of the convolutional layer is multiple feature maps, and each feature map extracts specific types of features through different convolutional kernels. The pooling layer is used to downsample the output of the convolutional layer, retain the key information of the features, reduce the data dimension and computational amount, and use the ReLU activation function to perform a non-linear transformation on the convolutional result, A(i, j) = max(0, O(i, j)). Further, the local features extracted by the convolutional layer are mapped to the global feature space to generate a global feature vector F img = Flatten(A), where Flatten is to flatten the multi-dimensional feature map output by the convolutional layer into a one-dimensional feature vector

[0032] 2.2 Physiological Parameter Feature Extraction

[0033] c) High-level features of physiological parameters are extracted through a fully connected neural network (FCNN). First, the normalized physiological parameter data is input into the network: X = [T norm , P norm , HR norm , where T norm , P norm , HR norm are the normalized temperature, pressure, and heart rate data respectively. Through the fully connected layer, feature extraction of physiological parameters is performed. Each node is connected to all nodes in the previous layer. The formula: A l = σ(W l ·A l-1 + b l ), where A l is the activation value of the i-th layer, W l is the weight matrix, b l is the bias term, σ is the activation function, and finally the physiological parameter feature vector F phy , the output of the last layer of the fully connected network, is the extracted high-level feature vector of physiological parameters

[0034] d) Statistical type features, extract the statistical features and frequency domain features of physiological parameters. Fourier transform: Convert the time-domain signal to the frequency-domain signal and extract the frequency components. The formula is: where x(t), F(ω) is the frequency-domain signal, ω is the frequency, and the frequency components (the main frequency components of the heart rate) are extracted for identifying periodic changes

[0035] Further, the processing method in step (3):

[0036] Calculation formula: F fusion = [F img , F phy , where Ffusion is the fused feature vector, F img is the feature vector extracted from the ultrasound image, F phy is the feature vector extracted from physiological parameters. By splicing, information from different modalities is integrated together, enabling subsequent models to comprehensively consider different features.

[0037] Furthermore, the processing method in step (4):

[0038] 4.1 Model architecture. The input layer receives the comprehensive feature vector sequence {F fusion (t)} from the ultrasound image and the physiological parameter sensor. Here, F fusion (t) is the feature vector at time t, which contains information on multi-modal data and can provide a comprehensive view of the breast health status. The model uses a Gated Graph Sequence Neural Network, combining a Graph Convolutional Network (GCN) and LSTM, which can effectively capture the temporal dynamics and relationships between features.

[0039] j) Graph convolutional layer. Graph structure construction: The comprehensive features are transformed into a graph structure, with nodes representing feature vectors and edges representing the relationships between features. In the graph convolutional layer, feature update is performed through the following formula: H (l+1 ) = σ(AH (l) W (l) ), where H (l) is the node feature matrix of the th layer, representing the features of the current layer, A is the normalized adjacency matrix used to represent the connections between nodes, W (l) is the weight matrix, and σ is the activation function (ReLU) used to introduce non-linearity;

[0040] k) LSTM layer. The feature sequence output from the graph convolutional layer enters the LSTM layer to capture the dynamic information in the time series. The state update formula of the LSTM cell is as follows:

[0041] i t = σ(W i ·[H t-1 , F fusion (t)] + b i )

[0042] f t = σ(W f ·[H t-1 , F fusion (t)] + b f )

[0043] o t = σ(W o ·[H t-1 , Ffusion (t)] + b o )

[0044]

[0045] H t = o t ·σ(C t )

[0046] i t , f t , o t are the input gate, forget gate, and output gate respectively, which determine the inflow, retention, and outflow of information. C t is the cell state of the current unit, and is the hidden state, storing the information at the current moment. C t is the cell state of the current unit, and H t is the hidden state, storing the information at the current moment;

[0047] 1) Fully connected layer. The output of the LSTM layer passes through the fully connected layer for feature extraction to obtain the final feature vector, F output = σ(W fc ·H T + b fc ), where H T is the last hidden state of the LSTM, and W fc and b fc are the learnable parameters of the fully connected layer;

[0048] m) Output layer. The output layer uses the Softmax activation function to calculate the prediction probabilities of various breast diseases: where P(k) represents the probability that the model predicts as class k, and the sum of the probabilities of all classes is 1;

[0049] 4.2 Model Training

[0050] n) To optimize the model performance, cross-entropy is used as the loss function: where N is the number of samples, C is the number of disease classes, y i，k is the true label of the sample on class k, is the probability predicted by the model;

[0051] o) The Adam optimizer is used, and the parameters are updated in combination with the adaptive learning rate,

[0052] 4.3 Disease Risk Assessment

[0053] p) Calculate the total risk score according to the prediction probabilities of each disease class: w k is the weight of each disease, which is set considering its severity;

[0054] q) Risk level judgment: According to the risk score, the risk level is divided into: low risk: R ≤ θ low , medium risk: θ low < R < θ high , high risk: R ≥ θ high

[0055] Warning mechanism: When the risk level reaches medium and high risks, the system will send a warning notice.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] 1. Flexible ultrasonic patch: It is a flexible ultrasonic patch that can be attached to the inner side of the underwear, integrating a high-density flexible ultrasonic sensor array, which can closely fit the breast curve and provide high-quality breast ultrasound images.

[0058] 2. Integration of multi-parameter sensors: Integrate various physiological parameter sensors such as temperature, pressure, and heart rate in the patch to realize real-time monitoring of multi-dimensional, physiological, and morphological characteristics of breast tissue.

[0059] 3. It can provide high-quality internal breast image data and perform intelligent analysis by combining multi-modal information.

[0060] 4. Early detection of breast abnormalities: Through continuous monitoring of the wearable ultrasonic device, it can timely capture the tiny changes in breast tissue, improve the early diagnosis rate, and increase the possibility of treatment success.

[0061] 5. Non-invasiveness and safety: Using ultrasonic technology, there is no ionizing radiation, which is harmless to the human body, suitable for long-term and frequent use, especially suitable for close monitoring of high-risk populations.

[0062] 6. Improve patient compliance: The device is designed to be comfortable and portable, and can be used by attaching it to the underwear, reducing the frequency of going to the hospital for examinations, and enhancing the patient's willingness and compliance to use.

[0063] 7. Real-time monitoring and warning: Integrate intelligent deep learning algorithms to analyze multi-modal data in real time, provide instant health status feedback and abnormal warnings, and facilitate timely intervention measures.

[0064] 8. Personalized medical support: Through the accumulation and analysis of long-term data, it helps doctors understand the breast health status of patients more comprehensively and formulate personalized diagnosis, treatment, and nursing plans. Description of the Drawings

[0065] Figure 1 It shows the overall appearance of the device and the layout of each module, including a flexible ultrasonic sensor array, physiological parameter sensors, a signal processing and control unit, a wireless communication module, and a power management module;

[0066] Figure 2 It is a front schematic diagram of the device;

[0067] Figure 3 It is a back schematic diagram of the device. Specific Embodiments

[0068] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0069] The intelligent wearable device for breast health monitoring in this embodiment includes several wearable ultrasonic patches, a wireless communication module, a signal processing and control unit, and a power management module. An outer shell is provided outside. The several wearable ultrasonic patches are connected by flexible connection bands and are respectively communicatively or electrically connected to the wireless communication module, the signal processing and control unit, and the power management module. The wearable ultrasonic patch includes an upper non-woven fabric, a middle flexible circuit board, and a lower conductive gel. The flexible circuit board is integrated with ultrasonic and physiological parameter sensors, and the flexible circuit board serves as the substrate of the sensor array and forms an ultrasonic transducer with the ultrasonic sensor.

[0070] Further, N×M elements are designed to form a two-dimensional matrix to cover the high-incidence areas of the breast. The size and spacing d of each element satisfy the Nyquist sampling theorem to avoid spatial aliasing, that is where λ is the ultrasonic wavelength, and the operating frequency of the ultrasonic transducer is set to f 0 = 7 - 15 MHz.

[0071] Further, the physiological parameter sensor is a temperature sensor and a pressure sensor.

[0072] The data analysis system of the intelligent wearable device for breast health monitoring in this embodiment includes a data acquisition system, a data transmission system, a data processing and analysis system, and a user interaction system. The data acquisition system is used to acquire ultrasonic images of breast tissue and physiological parameters of the breast area. The data transmission system realizes real-time data transmission. The data processing and analysis system includes a data preprocessing module, a feature extraction module, and a multi-modal deep learning algorithm module to realize data preprocessing, feature extraction, and data fusion. The user interaction system includes a data visualization system and an early warning notification module.

[0073] The system architecture design of this embodiment mainly includes four parts: a data acquisition system, a data transmission system, a data processing and analysis system, and a user interaction system.

[0074] 1. Data Acquisition System

[0075] · Wearable ultrasonic patch device:

[0076] ο Function: Collect ultrasonic images of breast tissue and obtain morphological features.

[0077] ο Design: Adopt flexible materials, with a miniaturized ultrasonic sensor array built-in, conforming to the breast curve and providing comfortable wearing.

[0078] · Physiological parameter sensor:

[0079] ο Function: Collect physiological parameters in the breast area, including temperature, pressure (elasticity), heart rate, etc.

[0080] ο Design: Integrated in a patch device for real-time monitoring of physiological indicators.

[0081] 2. Data transmission system

[0082] · Wireless communication module:

[0083] ο Function: Use low-power Bluetooth (BLE) or Wi-Fi to achieve real-time data transmission.

[0084] 3. Data processing and analysis system

[0085] · Data preprocessing module:

[0086] ο Function: Perform preprocessing such as noise reduction and normalization on the collected ultrasonic images and physiological parameters.

[0087] ο Method: Use technologies such as filters and normalization.

[0088] · Feature extraction module:

[0089] ο Function: Extract morphological features of ultrasonic images and physiological parameter features.

[0090] ο Method: Adopt a deep learning model to automatically extract high-level features.

[0091] · Multimodal deep learning algorithm module:

[0092] ο Function: Integrate multimodal data, combine temporal information, and perform disease prediction and risk assessment.

[0093] 4. User interaction system

[0094] · Data visualization system:

[0095] ο Function: Display monitoring results, risk assessment, and medical advice.

[0096] ο Design: Provide intuitive charts and reports for easy user understanding.

[0097] · Early warning notification module:

[0098] ο Function: When an abnormality is detected, send a warning message to the user.

[0099] ο Method: Notify users via mobile applications or text messages and provide follow-up suggestions.

[0100] (1) Detailed description of the data acquisition system:

[0101] Use medical-grade flexible materials as the substrate of the sensor array to ensure that the device can fit the breast curve. Use flexible piezoelectric materials to fabricate miniaturized ultrasonic transducers to achieve ultrasonic transmission and reception. According to the anatomical characteristics of the breast and monitoring requirements, design N×M array elements to form a two-dimensional matrix covering the high-incidence areas of the breast (such as the upper outer quadrant of the breast).: The size and spacing d of each array element satisfy the Nyquist sampling theorem to avoid spatial aliasing, that is where λ is the ultrasonic wavelength. The operating frequency of the ultrasonic transducer is set to f 0 = 7 - 15 MHz to achieve high-resolution imaging of shallow tissues. Use flexible conductive materials to connect the array elements to the signal processing circuit to ensure the reliability of signal transmission and the flexibility of the device.

[0102] Ultrasonic beamforming circuit:

[0103] · Analog front end (AFE), low-noise amplifier (LNA): The received signal of each array element is amplified by the LNA to increase the signal strength and reduce the influence of noise. Filter: Use a band-pass filter to filter out unnecessary frequency components and retain the target signal.

[0104] · Analog-to-digital converter (ADC), high-precision ADC: Convert the analog signal into a digital signal with a bit number of 12 to 16 bits, and the sampling rate satisfies the Nyquist sampling theorem.

[0105] · Digital signal processor (DSP), beamforming algorithm: Implement the Delay and Sum method to perform delay compensation and superposition on the received signal. The formula is:

[0106]

[0107] where w n is the array element weight coefficient, τ n is the received signal of the nth array element, τ n is the delay time, and the calculation formula is:

[0108]

[0109] d n is the distance between the nth array element and the reference point, θ is the beam pointing angle, and c is the propagation speed of ultrasonic waves in tissues (about 1540 m / s).

[0110] The outer shell is made of medical-grade silicone or polyurethane materials to ensure safety during long-term contact with the skin. The device has a certain degree of waterproof performance to prevent the influence of sweat or other liquids on the internal electronic components. It exists in the form of a patch and uses medical-grade adhesives to be directly attached to the skin or the inner side of the underwear. According to the breast sizes and shapes of different users, the device has a certain degree of adjustability to ensure full contact between the sensor array and the skin. Schematic diagram of the structure of the wearable ultrasonic patch device Figure 1 。

[0111] The physiological parameter sensor module is used to collect physiological parameters in the breast area and combine with ultrasonic images to provide multimodal data.

[0112] (1) Temperature sensor

[0113] · Type: High-precision digital temperature sensor,

[0114] · Layout: Multiple temperature sensors are distributed at key positions of the device to form a temperature monitoring network. The specific layout is as shown in Figure 2 shown.

[0115] · Function: Real-time monitor the temperature changes on the breast surface, detect local temperature abnormalities, and assist in disease diagnosis.

[0116] (2) Pressure sensor

[0117] · Type: Thin-film or flexible piezoresistive pressure sensor, with a thin thickness and high sensitivity.

[0118] · Layout: Integrated on a flexible substrate and jointly cover the breast area with the ultrasonic transducer array.

[0119] · Function: Measure the pressure distribution on the breast surface, evaluate the elasticity and hardness of tissues, and assist in judging the nature of lesions.

[0120] · Tissue elasticity evaluation: By measuring the deformation amount of tissues under pressure, calculate the elastic modulus. The formula is:

[0121]

[0122] where F is the applied force, A is the force-bearing area, and L 0 is the initial length.

[0123] (3) Heart rate sensor

[0124] · Type: Photoplethysmogram (PPG) sensor, using an LED light source and a photodiode.

[0125] · Location: Placed in the area where the pulse wave can be sensed, near the lower edge of the breast

[0126] · Function: Detect blood volume changes using optical methods, measure the user's heart rate, and provide a reference for overall health assessment.

[0127] Figure 2 : The device is fixed on the breast surface through adhesives or elastic straps to ensure good contact between the sensor and the skin.

[0128] (II) Detailed description of the data processing and analysis system:

[0129] 1 Data preprocessing module

[0130] 1.1 Ultrasonic image preprocessing

[0131] g) Denoising processing: Median filtering is a non-linear filtering method. It effectively removes noise, especially impulse noise, by replacing the value of each pixel with the median of the pixel values in its neighborhood. First, for each pixel f(i, j), a k×k window is selected. Further, all the pixel values within the window are collected, and the median of the pixel values within the window is calculated. Finally, the value of each pixel is replaced with the median of the pixel values in its neighborhood.

[0132] h) Image enhancement: Histogram equalization is adopted to improve the image contrast. Histogram equalization enhances the contrast by adjusting the histogram of the image to make it evenly distributed.

[0133] i) Size normalization: The image is adjusted to a unified size of M×N pixels. The specific steps are to use bilinear interpolation for image scaling. Calculate the scaling factor: where H and W are the height and width of the original image. Further, calculate the value of each pixel g(i, j) of the new image, and the calculation formula is After denoising, enhancing, and normalizing the size of the ultrasonic image using median filtering, the data quality can be significantly improved.

[0134] 1.2 Physiological parameter preprocessing

[0135] j) Filtering processing, using a low-pass filtering method to remove high-frequency noise in the signal. For a discrete signal, the output of the low-pass filter can be obtained through convolution operation, using a simple discrete smoothing filter: where h(k) is the impulse response of the filter, and K determines the window size of the filter, which affects the degree of smoothing.

[0136] k) Scale the physiological parameter values to the interval [0, 1] proportionally, and the formula is: x is the original physiological parameter value, x max and x min are the maximum and minimum values of this physiological parameter respectively.

[0137] 2 Feature extraction module

[0138] 2.1 Ultrasonic Image Morphological Extraction

[0139] Ultrasonic images contain a large amount of noise and detailed information. By using a Convolutional Neural Network (CNN), high-level features in the images, such as the shape, size, edge sharpness, and texture features of the mass, are automatically extracted. Input: The preprocessed ultrasonic image, with a size of m×n. The specific steps are as follows: First, convolution operations are performed to extract local features, such as edges and corners, and further convolution operations are carried out: where O(i, j) is the pixel value in the convolution output feature map. W(k, l) is the convolution kernel, I(i + k, j + l) is the pixel value of the input image, and b is the bias. The output of the convolutional layer is multiple feature maps, and each feature map extracts specific types of features through different convolution kernels. The pooling layer is used to downsample the output of the convolutional layer, retain the key information of the features, reduce the data dimension and computational amount. The ReLU activation function is used to perform a non-linear transformation on the convolution result, A(i, j) = max(0, O(i, j)). Further, the local features extracted by the convolutional layer are mapped to the global feature space to generate a global feature vector F img = Flatten(A), where Flatten is to flatten the multi-dimensional feature map output by the convolutional layer into a one-dimensional feature vector.

[0140] 2.2 Physiological Parameter Feature Extraction

[0141] e) Physiological parameters such as temperature, pressure, and heart rate are used to extract high-level features through a Fully Connected Neural Network (FCNN). First, the normalized physiological parameter data is input into the network: X = [T norm , P norm , HR norm , where T norm , P norm , HR norm are the normalized temperature, pressure, and heart rate data respectively. Through the fully connected layer, feature extraction is performed on the physiological parameters. Each node is connected to all nodes in the previous layer. Formula: A l = σ(W l .A l-1 + b l ), where A l is the activation value of the l-th layer, W l is the weight matrix, b l is the bias term. σ is the activation function. Finally, the physiological parameter feature vector F phy , the output of the last layer of the fully connected network, that is, the extracted high-level feature vector of the physiological parameters, is output.

[0142] f) Statistical type features, extract the statistical features (such as mean, standard deviation, trend of change) and frequency domain features (Fourier transform) of physiological parameters. Fourier transform: Convert the time-domain signal into a frequency-domain signal and extract the frequency components. The formula is: where \(x(t)\) and \(F(\omega)\) are frequency-domain signals, and \(\omega\) is the frequency. Extract the frequency components (the main frequency component of the heart rate) for identifying periodic changes.

[0143] 3 Multimodal data fusion

[0144] The feature vectors of different modalities are integrated into a comprehensive feature vector for subsequent model training and inference. The morphological feature vector and the physiological parameter feature vector are concatenated to form a comprehensive feature vector.

[0145] Formula: \(F\) fusion \( = [F\) img , \(F\) phy , where \(F\) fusion is the fused feature vector, \(F\) img is the feature vector extracted from the ultrasound image, and \(F\) phy is the feature vector extracted from physiological parameters. By concatenation, the information of different modalities is integrated together, enabling the subsequent model to comprehensively consider different features.

[0146] 4 Multimodal deep learning algorithms

[0147] 4.1 Model architecture, the input layer receives the sequence of comprehensive feature vectors \(\{F\) fusion (t)\} from the ultrasound image and physiological parameter sensors. Here, \(F\) fusion (t) is the feature vector at time \(t\), which contains the information of multimodal data and can provide a comprehensive view of the breast health status. The model uses the Gated Graph Sequence Neural Network (GGSNN), which combines the Graph Convolutional Network (GCN) and LSTM, and can effectively capture the temporal dynamics and the relationships between features.

[0148] r) Graph convolutional layer, graph structure construction: Convert the comprehensive features into a graph structure, with nodes representing feature vectors and edges representing the relationships between features. In the graph convolutional layer, the feature update is performed through the following formula: \(H\) (l+1) \( = \sigma(AH\) (l) W\) (l) ), where \(H\) (l) is the node feature matrix of the \(i\)-th layer, representing the features of the current layer. \(A\) is the normalized adjacency matrix, used to represent the connections between nodes. \(W\) (l) is the weight matrix. \(\sigma\) is the activation function (ReLU), used to introduce non-linearity.

[0149] s) LSTM layer, the feature sequence output from the graph convolution layer enters the LSTM layer to capture the dynamic information in the time series. The state update formula of the LSTM unit is as follows:

[0150] i t =σ(W i ·[H t-1 , F fusion (t)]+b i )

[0151] f t =σ(W f ·[H t-1 , F fusion (t)]+b f )

[0152] o t =σ(W o ·[H t-1 , F fusion (t)]+b o )

[0153]

[0154] H t =o t ·σ(C t )

[0155] Among them, i t , f t , o t They are the input gate, forget gate and output gate, which determine the inflow, retention and outflow of information. t It is the cell state of the current unit, which is a hidden state and stores the information at the current moment. t is the cell state of the current unit, H t It is a hidden state that stores the information at the current moment.

[0156] t) Fully connected layer: The output of the LSTM layer is passed through the fully connected layer for feature extraction to obtain the final feature vector. output =σ(W fc .H T +b fc ), where HT 是 The last hidden state of the LSTM, W fc and b fc are the learnable parameters of the fully connected layer.

[0157] u) Output layer: The output layer uses the Softmax activation function to calculate the predicted probability of various breast diseases: Where P(k) represents the probability that the model predicts class k, and the sum of the probabilities of all categories is 1.

[0158] 4.2 Model Training

[0159] v) To optimize the model performance, cross-entropy is used as the loss function: where N is the number of samples, C is the number of disease categories, and y i，k is the true label of the sample for class k, and is the probability predicted by the model.

[0160] w) The Adam optimizer is adopted, and the parameters are updated in combination with an adaptive learning rate

[0161] 4.3 Disease Risk Assessment

[0162] x) Calculate the total risk score based on the predicted probabilities for each disease category: w k is the weight for each disease, which is set considering its severity.

[0163] y) Risk level judgment: The risk level is divided according to the risk score as follows: low risk: R ≤ θ low , medium risk: θ low <R<θ high , high risk: R ≥ θ high

[0164] Early warning mechanism: When the risk level reaches medium or high risk, the system will send an early warning notice so that measures can be taken in a timely manner.

[0165] The above is only the preferred embodiment of the present invention and does not impose any limitation on the present invention. Any person skilled in the art, without departing from the technical solution of the present invention, makes any form of equivalent replacement or modification and other changes to the technical solution and technical content disclosed by the present invention, all of which fall within the content of the technical solution of the present invention and still belong to the protection scope of the present invention.

Claims

1. A smart wearable device for breast health monitoring, characterized in that: The invention comprises a plurality of wearable ultrasonic patches, a wireless communication module, a signal processing and control unit and a power management module, and an external shell is arranged. The plurality of wearable ultrasonic patches are connected by flexible connecting belts and are respectively connected in communication or electrical connection with the wireless communication module, the signal processing and control unit and the power management module.

2. The breast health monitoring smart wearable device according to claim 1, characterized in that: The wearable ultrasonic patch includes an upper layer of non-woven fabric, a middle layer of flexible circuit board and a lower layer of conductive gel. Ultrasonic and physiological parameter sensors are integrated on the flexible circuit board. The flexible circuit board serves as the substrate of the sensor array and forms an ultrasonic transducer with the ultrasonic sensor.

3. The breast health monitoring smart wearable device according to claim 2, characterized in that: Design N×M array elements to form a two-dimensional matrix to cover the high incidence area of ​​breast. The size and spacing d of each array element meet the Nyquist sampling theorem to avoid spatial aliasing, that is, Where λ is the ultrasonic wavelength, and the operating frequency of the ultrasonic transducer is set to f0=7~15MHz.

4. The breast health monitoring smart wearable device according to claim 2, characterized in that: The physiological parameter sensors are temperature sensors and pressure sensors.

5. A data analysis system using the smart wearable device for breast health monitoring according to any one of claims 1 to 4, characterized in that: It includes a data acquisition system, a data transmission system, a data processing and analysis system and a user interaction system. The data acquisition system is used to collect ultrasonic images of breast tissue and physiological parameters of the breast area. The data transmission system realizes real-time transmission of data. The data processing and analysis system includes a data preprocessing module, a feature extraction module and a multimodal deep learning algorithm module to realize data preprocessing, feature extraction and data fusion. The user interaction system has a data visualization system and an early warning notification module.

6. An analysis method for the data analysis system of the breast health monitoring smart wearable device according to claim 5, characterized in that: The steps include: (1) Data processing; (2) Feature extraction; (3) Multimodal data fusion; (4) Multimodal deep learning algorithm.

7. The analysis method of the data analysis system of the breast health monitoring intelligent wearable device according to claim 6, characterized in that: The processing method of step (1) is: 1.1 Ultrasound image preprocessing a) Denoising: Use median filtering to remove noise; b) Image enhancement: Use histogram equalization to improve image contrast; c) Size normalization: resize the image to a uniform size of M×N pixels, use bilinear interpolation to scale the image, and calculate the scaling factor: Where H and W are the height and width of the original image, and the calculation formula for each pixel value g(i, j) of the new image is: 1.2 Physiological parameter preprocessing a) Filtering: Use low-pass filtering to remove high-frequency noise from the signal. For discrete signals, The output of the low-pass filter can be obtained by convolution, using a discrete smoothing filter: Where h(k) is the impulse response of the filter, K is the window size of the filter, that is, the number of sampling points covered by the filter impulse response h[n] determines the window size of the filter and affects the degree of smoothing. k is the index variable in the convolution operation, which is used to traverse the impulse response of the filter. b) Scale the physiological parameter value to the interval [0, 1] according to the formula: x is the original physiological parameter value, x max and x min are the maximum and minimum values ​​of the physiological parameter respectively.

8. The analysis method of the data analysis system of the breast health monitoring intelligent wearable device according to claim 6, characterized in that the processing method of step (2) is: 2.1 Ultrasound Image Morphology Extraction By using convolutional neural networks, high-level features in the image are automatically extracted, including the shape, size, edge clarity, and texture features of the mass. Input: preprocessed ultrasound image with a size of m×n. The specific steps are as follows: First, convolution operation is performed to extract local features, edges, corners, and further convolution operation is performed: Where O(i, j) is the pixel value in the convolution output feature map, W(k, l) is the convolution kernel, I(i+k, j+l) is the pixel value of the input image, b is the bias, and the output of the convolution layer is multiple feature maps. Each feature map extracts a specific type of feature through a different convolution kernel. The pooling layer is used to downsample the convolution layer output, retain the key information of the feature, reduce the data dimension and the amount of calculation, and use the ReLU activation function to perform a nonlinear transformation on the convolution result, A(i, j) = max(0, O(i, j)). Furthermore, the local features extracted by the convolution layer are mapped to the global feature space to generate a global feature vector F. img = Flatten(A), where is the multi-dimensional feature map output by the convolutional layer flattened into a one-dimensional feature vector, 2.2 Physiological parameter feature extraction a) Physiological parameters are extracted through a fully connected neural network (FCNN). First, the normalized physiological parameter data is input into the network: X = [T norm , P norm , HR norm ], where T norm , P norm , HR norm They are normalized temperature, pressure, and heart rate data. The physiological parameters are extracted through the fully connected layer. Each node is connected to all nodes in the previous layer. The formula is: l =σ(W l .A l-1 +b l ), where A l is the activation value of the ,th layer, W l is the weight matrix, b l is the bias term, σ is the activation function, and finally outputs the physiological parameter feature vector F phy , the output of the last layer of the fully connected network, i.e., the extracted high-level feature vector of physiological parameters; b) Statistical type features: extract the statistical features and frequency domain features of physiological parameters. Fourier transform: convert the time domain signal into the frequency domain signal and extract the frequency component. The formula is: Among them, x(t), F(ω) are frequency domain signals, ω is the frequency, and the frequency component (the main frequency component of the heart rate) is extracted to identify periodic changes.

9. The analysis method of the data analysis system of the breast health monitoring intelligent wearable device according to claim 6, characterized in that: The processing method of step (3): Calculation formula: F fusion =[F img , F phy ], where F fusion is the fused feature vector, F img is the feature vector extracted from the ultrasound image, F phy It is the feature vector extracted from physiological parameters. By splicing, the information of different modalities is integrated together so that the subsequent model can comprehensively consider different features.

10. The analysis method of the data analysis system of the breast health monitoring intelligent wearable device according to claim 6, characterized in that: The processing method of step (4) is: 4.1 Model Architecture,The input layer receives the comprehensive feature vector sequence {F fusion (t)}, where F fusion (t) is the feature vector at time t, which contains information of multimodal data and can provide a comprehensive perspective of breast health status. The model uses Gated Graph Sequence Neural Network, combined with graph convolutional network (GCN) and LSTM, which can effectively capture the relationship between time series dynamics and features. a) Graph convolution layer, graph structure construction: The comprehensive features are converted into a graph structure, with nodes representing feature vectors and edges representing the relationship between features. In the graph convolution layer, feature updates are performed using the following formula: H (l+1) =σ(AH (l) W (l) ), where H (l) is the node feature matrix of the lth layer, representing the features of the current layer, A is the normalized adjacency matrix used to represent the connections between nodes, and W (l) is the weight matrix, σ is the activation function (ReLU), which is used to introduce nonlinearity; b) LSTM layer, the feature sequence output from the graph convolution layer enters the LSTM layer to capture the dynamic information in the time series. The state update formula of the LSTM unit is as follows: i t =σ(W i .[H t-1 ,F fusion (t)]+b i ) f t =σ(W f .[H t-1 ,F fusion (t)]+b f ) o t =σ(W o .[H t-1 F fusion (t)]+b o ) H t =o t .σ(C t ) i t , f t , o t They are input gate, forget gate and output gate, which determine the inflow, retention and outflow of information. t is the cell state of the current unit, is a hidden state, stores the information of the current moment, C t is the cell state of the current unit, H t It is a hidden state, storing the information at the current moment; c) Fully connected layer: The output of the LSTM layer is subjected to feature extraction through the fully connected layer to obtain the final feature vector, F output =σ(W fc .H T +b fc ), where H T is the last hidden state of LSTM, W fc and b fc are the learnable parameters of the fully connected layer; d) Output layer: The output layer uses the Softmax activation function to calculate the predicted probability of various breast diseases: Where P(k) represents the probability that the model predicts class k, and the sum of the probabilities of all categories is 1; 4.2 Model Training e) In order to optimize the model performance, cross entropy is used as the loss function: Where N is the number of samples, C is the number of disease categories, and y i，k is the true label of the sample in category k, is the probability predicted by the model; f) Adopt Adam optimizer and adaptive learning rate to update parameters. 4.3 Disease risk assessment g) Calculate the total risk score based on the predicted probability of each disease category: w k is the weight of each disease, which is set considering its severity; h) Risk level judgment: Risk levels are divided into low risk: R≤θ according to risk score. low , medium risk: θ low <R<θ high , high risk: R ≥ θ high i) Early warning mechanism: When the risk level reaches medium or high risk, the system will send an early warning notification.

Images