Intraoperative hypotension prediction equipment based on adaptive filtering and multi-channel deep learning

Abstract

The invention provides intraoperative hypotension prediction equipment based on adaptive filtering and multi-channel deep learning, the equipment is provided with a main machine body, the front side of the main machine body is provided with a mobile handheld device, a touch screen and a switch button, and the upper part of the mobile handheld device is provided with a handheld device screen, function buttons and a data input port; a network interface, a universal auxiliary equipment interface, a USB interface, a data output port, a power interface and a high-decibel buzzer are arranged on the rear side of the host; the equipment can continuously monitor and accurately record key physiological index data such as heart rate, blood pressure, oxyhemoglobin saturation and end-expiratory CO2 of a patient in an operation; carrying out pre-data processing by adopting a self-adaptive filtering algorithm; a complex algorithm formed through deep learning model training is used for analyzing the preprocessed data, the physiological change condition of the patient after five minutes is predicted for clinical judgment and treatment of a doctor, and if it is predicted that the patient is in danger, an alarm device is activated immediately to remind the clinician to conduct treatment in time; and the whole operation process is protected.

Description

Technical Field

[0001] The present invention relates to the technical field of medical devices, and specifically to an intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning. Background Art

[0002] Intraoperative hypotension (IOH) is an important issue in perioperative care, and is closely related to postoperative complications such as postoperative delirium and increased mortality. Research shows that the longer the duration of IOH, the higher the risk of serious complications; therefore, being able to predict the occurrence of IOH in advance, so as to assist doctors to take preventive measures in a timely manner, is of great significance for reducing the postoperative risk and mortality of patients; currently, traditional hypotension prediction models mainly rely on arterial waveform data monitored during surgery, extract features from the original waveform data through manually designed feature extraction algorithms, and then use machine learning methods such as logistic regression, naive Bayes, support vector machines, etc. for classification; however, these methods rely on manually designed features, are difficult to capture complex physiological signal patterns, and have limited performance when dealing with multivariate time series; with the development of deep learning technology, more and more researchers have begun to try to use deep learning methods to improve the intraoperative hypotension prediction model; deep learning methods can automatically learn and identify features, avoiding the limitations of manually designed features.

[0003] At present, the existing hypotension prediction models still have the following deficiencies:

[0004] 1. Lack of preoperative baseline: Most of the existing models use only intraoperative waveform data for modeling, lacking the correction of preoperative baseline data, and the accuracy of the model is limited.

[0005] 2. Insufficient utilization of frequency domain information: The existing models mainly rely on time domain features and ignore frequency domain information, while frequency domain analysis has shown the potential to reveal patterns that are not obvious in the time domain in other fields.

[0006] 3. Limited ability to handle noise and redundant data: The physiological signal data monitored during surgery usually contains noise and redundant information, and the existing methods lack effective mechanisms to suppress noise and retain key signal features.

[0007] Therefore, according to the requirements, the applicant proposes an intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning. By combining preoperative static data such as patient age, gender, weight, height, etc. and intraoperative multi-channel biosignal data, this device can capture the patient's physiological state more comprehensively. The device monitoring system uses an adaptive filtering module, which can convert time-domain data to frequency-domain, extract frequency-domain features, and apply adaptive threshold technology to remove noise and retain key signal information. Furthermore, to improve the effect of feature extraction, a combination of a convolutional neural network (CNN) and a Transformer model is adopted to provide decision support for real-time monitoring and early warning in clinical surgery. The combination of the two can significantly improve the accuracy and reliability of prediction. Summary of the Invention

[0008] To solve the above technical problems, the present invention proposes an intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning. By setting a main body, a mobile handheld device, a touch screen, and a switch button are arranged on the front side of the main body. A handheld screen, function buttons, and a data input port are arranged on the upper part of the mobile handheld device; a network interface, a general auxiliary device interface, a USB interface, a data output port, a power supply interface, and a high-decibel buzzer are arranged on the rear side of the main body; the device can continuously monitor and accurately record key physiological index data such as the patient's heart rate, blood pressure, blood oxygen saturation, end-tidal CO2, etc. during the operation; an adaptive filtering algorithm is used for pre-data processing; and a complex algorithm formed by training a deep learning model is used to analyze the pre-processed data to predict the patient's physiological changes 5 minutes later for doctors' clinical judgment and treatment. For example, when it is predicted that the patient is in danger, the alarm device will be immediately activated to remind the clinician to deal with it in time, escorting the entire operation.

[0009] To achieve the above object, the technical solution adopted by the present invention is:

[0010] An intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning includes a main body, a touch screen, a switch button, a mobile handheld device, function buttons, a handheld screen, a data input port, a USB interface, a high-decibel buzzer, a power supply interface, a network cable interface, a general auxiliary device interface, and a data output port. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning is provided with a main body. The mobile handheld device is arranged on the left side of the front side of the main body, the touch screen and the switch button are arranged on the right side of the front side of the main body, and the switch button is located below the touch screen; the handheld screen is arranged on the upper part of the mobile handheld device, the function buttons are arranged in the middle, and the data input port is arranged on the lower part; three groups of network cable interfaces are arranged in the middle of the rear side of the main body, the general auxiliary device interface is arranged below the network cable interface, the data output port and two groups of USB interfaces are arranged below the general auxiliary device interface; the high-decibel buzzer is arranged in the upper right corner of the rear side of the main body.

[0011] Furthermore, a power interface is provided at the lower left corner of the rear side of the main body. The system of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning uses an adaptive filtering algorithm for pre-data processing; and analyzes the pre-processed data using a complex algorithm formed by training a deep learning model to predict the physiological changes of the patient after 5 minutes for clinical judgment and treatment by doctors. For example, when it is predicted that the patient is in danger, the alarm device will be immediately activated to trigger a high-decibel buzzer to remind the clinician to handle it in time.

[0012] Furthermore, the specific working steps of the system of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning are as follows:

[0013] Step 1: Data preprocessing; preprocess the data of two parts, namely the preoperative static baseline data and the intraoperative dynamic waveform data of the patient to obtain the input matrix that the model can receive.

[0014] Step 2: Adaptive filtering module processing; after data preprocessing, the input matrix enters the adaptive filtering module for further processing; the main function of the adaptive filtering module is to extract frequency domain features and suppress noise.

[0015] Step 3: Multi-channel deep learning modeling; use multi-channel deep learning to build models using the CNN model and the Transformer model.

[0016] Step 4: Perform data feature fusion; splice the output of the Transformer and the preoperative static features in one dimension to form the feature vector required for final prediction.

[0017] Step 5: Enter the classification module layer for prediction; input the fused feature vector into the fully connected layer and output the prediction probability of the hypotension event through the Sigmoid activation function; according to the prediction probability, set a threshold to judge whether a hypotension event occurs; if the prediction probability is greater than the threshold, it is determined as a hypotension event; otherwise, it is determined as a non-hypotension event; finally, the model can indicate whether a hypotension event will occur within the next 5 minutes, 10 minutes or 15 minutes.

[0018] Furthermore, the specific processing of the static and dynamic waveform data in the data preprocessing in Step 1 is as follows:

[0019] Static baseline data processing:

[0020] 1) Standardization: Use the Z-score standardization method to convert the numerical features into a form with a mean of 0 and a standard deviation of 1.

[0021] 2) Missing value filling: Use the mean or median to fill in the missing baseline data.

[0022] Dynamic waveform data processing:

[0023] 1) Outlier filtering: Exclude data segments where MAP < 20 mmHg or MAP > 160 mmHg;

[0024] 2) Sliding window filling: Use interpolation method to fill in the missing time points to ensure that each data segment is continuous and complete;

[0025] 3) Data segmentation: Segment the original data according to time windows of fixed length;

[0026] 4) Label definition: Label the data of each time window based on whether a hypotensive event occurs within the next 5 minutes:

[0027] If a hypotensive event occurs within the next 5 minutes, the label is 1, positive sample;

[0028] If no hypotensive event occurs within the next 5 minutes, the label is 0, negative sample.

[0029] Furthermore, the specific processing steps of the adaptive filtering module in the second step are as follows:

[0030] 1) Select input sequence: Use a 30 - second waveform data sequence as input; at the time point 5 minutes before the occurrence of the hypotensive event T, extract the 30 - second waveform data before this time point, that is, the data from T - 5:00 to T - 4:30; similarly, for the time points of 10 minutes and 15 minutes, the corresponding 30 - second waveform data will also be extracted;

[0031] 2) Frequency - domain feature extraction and noise suppression: The input 30 - second waveform data will pass through the adaptive filtering block, which can extract frequency - domain features from the time - domain signal and suppress noise through the adaptive filtering mechanism.

[0032] 6. The intraoperative hypotension prediction device based on adaptive filtering and multi - channel deep learning according to claim 5, characterized in that: the frequency - domain feature extraction and noise suppression; the specific steps are as follows:

[0033] First, Fourier transform: Convert the input multi - channel time - domain signal into a frequency - domain signal

[0034] X(f) = FFT(x(t))

[0035] where: is the original waveform data;

[0036] C is the number of channels;

[0037] T is the number of time samples;

[0038] F is the number of frequency components;

[0039] Secondly, adaptive filtering: Introduce a trainable parameter matrix and constrain it within the range of [0, 1] through the sigmoid activation function to generate an importance mask:

[0040]

[0041] The mask emphasizes relevant frequency components and suppresses irrelevant noise, which is used to dynamically adjust the weights of each frequency component; then calculate the filtered frequency-domain signal through element-wise multiplication:

[0042] X'(f) = X(f) ⊙ σ(W(f))

[0043] where: ⊙ represents element-wise multiplication;

[0044] Finally, inverse Fourier transform: Convert the filtered frequency-domain signal back to the time domain through the inverse Fourier transform to obtain the enhanced time-domain signal.

[0045] Furthermore, the CNN model and Transformer model for multi-channel deep learning modeling in the second step are specifically as follows:

[0046] CNN model: After being processed by the adaptive filtering module, the data enters the CNN model for feature extraction; the CNN module consists of five convolutional layers, and the kernel size of each convolutional layer is 10, with a stride of 1, which is used to learn and extract the high-frequency details of the signal; the convolutional layer extracts local features through a sliding window, which can effectively capture short-term dependencies and perform preliminary feature extraction on the input time series data; after each convolutional layer, batch normalization and ReLU activation functions are connected to accelerate the training process and improve the stability of the model; after the convolutional layer, a max-pooling layer with a stride of 2 is used to reduce the data dimension and computational amount; at the same time, a Dropout layer is added after each convolutional layer and pooling layer to prevent the model from overfitting.

[0047] Transformer model: The features extracted by the CNN model are then input into the Transformer model for global dependency modeling; the Transformer model consists of multiple encoder and decoder layers, and each encoder layer contains a multi-head self-attention mechanism and a feed-forward neural network; the specific process is as follows:

[0048] 1) Positional encoding: Add temporal information to the feature sequence output by the CNN through positional encoding to ensure that the model can capture the temporal relationship in the sequence; the positional encoding formula is:

[0049]

[0050] Wherein:

[0051] k ∈ {1, …, [d model / 2]};

[0052] d model is the dimension of the model;

[0053] 2) Multi-Head Self-Attention Mechanism: The core of Transformer is the multi-head self-attention mechanism, which is used for global modeling of the input sequence; each attention head independently calculates the attention weights and concatenates the results, and then obtains the final attention output through a linear transformation; the calculation formula for multi-head attention is:

[0054] MultiHead(Q, K, V) = Concat(head1, head2…, head h )W O

[0055] For each head h, calculate the query (Query), key (Key), and value (Value) matrices:

[0056]

[0057] At the same time, calculate the attention weights and apply them to the value matrix:

[0058]

[0059] d k is the dimension of the key;

[0060] 3) Encoder Layer: The Transformer model contains four identical encoder layers, and each encoder layer consists of a multi-head self-attention mechanism and a feed-forward neural network, which are used to further extract and fuse features;

[0061] 4) Decoder Layer: The output of the encoder generates the final prediction result through the decoder layer; the decoder layer also contains a multi-head self-attention mechanism and a feed-forward neural network, which are used to combine the target sequence and shallow high-dimensional features.

[0062] The benefits brought by this application are:

[0063] 1. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning can integrate the preoperative static data of patients. By combining the preoperative static data, such as patient age, gender, weight, height, etc. and intraoperative multi-channel biosignal data, it can capture the physiological state of patients more comprehensively; the introduction of preoperative baseline data helps to improve the accuracy and robustness of the model;

[0064] 2. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning adopts an adaptive filtering module. By introducing the adaptive filtering module, the time-domain data is converted to the frequency domain through Fourier transform, frequency-domain features are extracted, and the adaptive threshold technology is applied to remove noise and retain key signal information. This module enhances the model's ability to process data in a noisy clinical environment.

[0065] 3. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning uses a multi-channel deep learning architecture, combining a convolutional neural network (CNN) and a Transformer model, which are respectively used to extract local temporal and global temporal dependencies. The CNN can capture short-range temporal patterns, while the Transformer is good at modeling long-range dependencies. Through this dual-path architecture, the model can better understand complex physiological signal patterns.

[0066] 4. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning can predict the physiological changes of patients 5 minutes later, providing decision support for real-time monitoring and early warning in clinical surgery. Brief Description of the Drawings

[0067] Figure 1 Schematic diagram of the overall structure of the present invention Figure 1 ;

[0068] Figure 2 Schematic diagram of the overall structure of the present invention Figure 2 ;

[0069] Figure 3 Schematic diagram of the operation flow of the device system of the present invention;

[0070] Figure 4 Schematic diagram of the model construction of the present invention;

[0071] Figure 5 Schematic diagram of the method steps of the model training of the present invention;

[0072] Figure 6 Schematic diagram of the frequency-domain feature extraction and noise suppression steps of the present invention;

[0073] Figure 7 Schematic diagram of the multi-channel deep learning modeling of the present invention.

[0074] In the figure, the markings are: 1. Main body; 2. Touch screen; 3. Switch button; 4. Mobile handheld device; 5. Function button; 6. Handheld device screen; 7. Data input port; 8. USB interface; 9. High-decibel buzzer; 10. Power interface; 11. Network cable interface; 12. General auxiliary device interface; 13. Data output port. Detailed Description of the Invention

[0075] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0076] As Figures 1-3 shown, the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning includes a main body 1, a touch screen 2, a switch button 3, a mobile handheld device 4, function buttons 5, a handheld screen 6, a data input port 7, a USB interface 8, a high-decibel buzzer 9, a power supply interface 10, a network cable interface 11, a general auxiliary device interface 12, and a data output port 13. As Figure 1 shown, the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning is provided with a main body 1. The mobile handheld device 4 is arranged on the left side of the front side of the main body 1. The mobile handheld device 4 communicates and controls with the main body 1 through a wireless network. The touch screen 2 and the switch button 3 are arranged on the right side of the front side of the main body 1. The touch screen 2 can perform operations such as data interaction and switching of monitoring data. The switch button 3 controls the on / off of the entire device. The switch button 3 is located below the touch screen 2. The handheld screen 6 is arranged on the upper part of the mobile handheld device 4. The handheld screen 6 is synchronously displayed with the touch screen arranged on the main body 1 and can be operatively controlled with each other. The function buttons 5 are arranged in the middle of the mobile handheld device 4. The function buttons 5 can quickly switch to the corresponding monitoring data items. The data input port 7 is arranged on the lower part of the mobile handheld device 4. The data input port 7 can be connected to an external device to quickly input the required data. As Figure 2 shown, three groups of network cable interfaces 11 are arranged in the middle of the rear side of the main body 1. The network cable interfaces 11 can connect to the internal network, the external network, or transmit information to the device through a network interface. The general auxiliary device interface 12 is arranged below the network cable interfaces 11. The general auxiliary device interface 12 can be connected to a general interface of some monitoring devices commonly used in the surgical process to access monitoring data. The data output port 13 and two groups of USB interfaces 8 are arranged below the general auxiliary device interface 12. The monitoring data can be output to an external data for display through the data output port 13. The USB interfaces 8 are used for debugging the device or connecting to monitoring devices using the USB interface. The high-decibel buzzer 9 is arranged in the upper right corner of the rear side of the main body 1. When the monitoring data is abnormal, an alarm is issued through the high-decibel buzzer 9 to remind relevant personnel to handle it in time. The power supply interface 10 is arranged in the lower left corner of the rear side of the main body 1. The power supply is connected in through the power supply interface 10. As Figure 3As shown in the figure, the system of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning can continuously monitor and accurately record key physiological index data of patients during surgery, such as heart rate, blood pressure, blood oxygen saturation, end-tidal CO2, etc.; these real-time data will be automatically and quickly transmitted to the hypotension prediction module in the device; the system uses an adaptive filtering algorithm to perform preprocessing on multi-channel physiological data, such as denoising and feature extraction; and uses a complex algorithm formed by training a deep learning model based on a large amount of clinical data to conduct in-depth analysis on the preprocessed data to predict whether the patient will develop hypotension in 5 minutes: if the prediction result is "yes", it indicates that the patient is very likely to have hypotension in 5 minutes, and the device will immediately activate the alarm device, and use a dual reminder method of a high-decibel beeping sound and a prominent flashing light to remind the clinician to closely monitor the deteriorating blood pressure of the patient; if the prediction result is "no", the system will continue to monitor the data and loop through the above process to escort the entire surgical process.

[0077] As Figure 4 shown in the figure, the system model of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning includes a data preprocessing layer, an adaptive filtering module layer, a multi-channel deep learning modeling layer, a feature fusion layer, and a classification module layer.

[0078] As Figure 5 shown in the figure, the specific working steps of the system of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning are as follows:

[0079] Step 1: Data preprocessing; preprocess two parts of data, namely the preoperative static baseline data and the intraoperative dynamic waveform data of the patient, to obtain the input matrix that the model can receive.

[0080] Step 2: Adaptive filtering module processing; after data preprocessing, the input matrix enters the adaptive filtering module for further processing; the main function of the adaptive filtering module is to extract frequency domain features and suppress noise.

[0081] Step 3: Multi-channel deep learning modeling; use multi-channel deep learning to build models using the CNN model and the Transformer model.

[0082] Step 4: Perform data feature fusion; splice the output of the Transformer and the preoperative static features in one dimension to form the feature vector required for final prediction.

[0083] Step 5: Enter the classification module layer for prediction; input the fused feature vector into the fully connected layer and output the prediction probability of hypotensive events through the Sigmoid activation function; set a threshold according to the prediction probability to determine whether a hypotensive event occurs; if the prediction probability is greater than the threshold, it is determined as a hypotensive event; otherwise, it is determined as a non-hypotensive event; the final model can indicate whether a hypotensive event will occur within the next 5 minutes, 10 minutes, or 15 minutes.

[0084] The specific processing of static and dynamic waveform data in data preprocessing in Step 1 of the system working steps of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning is as follows:

[0085] Static baseline data processing:

[0086] 1) Standardization: Use the Z-score standardization method to convert numerical features into a form with a mean of 0 and a standard deviation of 1;

[0087] 2) Missing value filling: Fill in the missing baseline data using the mean or median;

[0088] Dynamic waveform data processing:

[0089] 1) Outlier filtering: Exclude data segments with MAP < 20 mmHg or MAP > 160 mmHg;

[0090] 2) Sliding window filling: Use interpolation to fill in the missing time points to ensure that each segment of data is continuous and complete;

[0091] 3) Data segmentation: Segment the original data according to a fixed-length time window;

[0092] 4) Label definition: Label the data of each time window based on whether a hypotensive event occurs within the next 5 minutes:

[0093] If a hypotensive event occurs within the next 5 minutes, the label is 1, a positive sample;

[0094] If a hypotensive event does not occur within the next 5 minutes, the label is 0, a negative sample.

[0095] The specific processing steps of the adaptive filtering module in Step 2 of the system working steps of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning are as follows:

[0096] Step 1: Select the input sequence: Use a 30-second waveform data sequence as the input; at the time point 5 minutes before the occurrence of the hypotensive event T, extract the waveform data of the 30 seconds before this time point, that is, the data from T - 5:00 to T - 4:30; similarly, for the time points of 10 minutes and 15 minutes, the corresponding 30-second waveform data will also be extracted;

[0097] Step 2: Frequency-domain feature extraction and noise suppression: The input 30-second waveform data passes through an adaptive filtering block, which can extract frequency-domain features from the time-domain signal and suppress noise through an adaptive filtering mechanism; as Figure 6 shown, the specific steps are as follows:

[0098] 1) Fourier transform: Convert the input multi-channel time-domain signal into a frequency-domain signal

[0099] X(f) = FFT(x(t))

[0100] where: is the original waveform data;

[0101] C is the number of channels;

[0102] T is the number of time samples;

[0103] F is the number of frequency components;

[0104] 2) Adaptive filtering: Introduce a trainable parameter matrix and constrain it within the range of [0,1] through a sigmoid activation function to generate an importance mask:

[0105]

[0106] The mask emphasizes relevant frequency components and suppresses irrelevant noise, and is used to dynamically adjust the weights of each frequency component; then calculate the filtered frequency-domain signal through element-wise multiplication:

[0107] X'(f) = X(f) ⊙ σ(W(f))

[0108] where: ⊙ represents element-wise multiplication;

[0109] 3) Inverse Fourier transform: Convert the filtered frequency-domain signal back to the time domain through the inverse Fourier transform to obtain the enhanced time-domain signal.

[0110] As Figure 7 shown, the specific CNN model and Transformer model for multi-channel deep learning modeling in Step 2 of the system working steps of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning are as follows:

[0111] CNN Model: After being processed by the adaptive filtering module, the data enters the CNN model for feature extraction. The CNN module consists of five convolutional layers, each with a kernel size of 10 and a stride of 1, which are used to learn and extract the high-frequency details of the signal. The convolutional layer extracts local features through a sliding window, which can effectively capture short-term dependencies and perform preliminary feature extraction on the input time series data. After each convolutional layer, batch normalization and the ReLU activation function are connected to accelerate the training process and improve the stability of the model. After the convolutional layer, a max pooling layer with a stride of 2 is used to reduce the data dimension and computational amount. At the same time, a Dropout layer is added after each convolutional layer and pooling layer to prevent the model from overfitting.

[0112] Transformer Model: The features extracted by the CNN model are then input into the Transformer model for global dependency modeling. The Transformer model consists of multiple encoder and decoder layers, and each encoder layer contains a multi-head self-attention mechanism and a feed-forward neural network. The specific process is as follows:

[0113] 1) Positional Encoding: Add temporal information to the feature sequence output by the CNN through positional encoding to ensure that the model can capture the temporal relationship in the sequence. The positional encoding formula is:

[0114]

[0115] Where:

[0116] k ∈ {1, …, [d model / 2]};

[0117] d model is the dimension of the model;

[0118] 2) Multi-Head Self-Attention Mechanism: The core of the Transformer is the multi-head self-attention mechanism, which is used to perform global modeling on the input sequence. Each attention head independently calculates the attention weights, and the results are concatenated and then passed through a linear transformation to obtain the final attention output. The calculation formula for multi-head attention is:

[0119] MultiHead(Q, K, V) = Concat(head1, head2…, head h )W O

[0120] For each head h, calculate the query (Query), key (Key), and value (Value) matrices:

[0121]

[0122] Simultaneously calculate the attention weights and apply them to the value matrix:

[0123]

[0124] d k is the dimension of the key;

[0125] 3) Encoder layer: The Transformer model contains four identical encoder layers, and each encoder layer consists of a multi-head self-attention mechanism and a feed-forward neural network, which are used to further extract and fuse features;

[0126] 4) Decoder layer: The output of the encoder generates the final prediction result through the decoder layer; the decoder layer also contains a multi-head self-attention mechanism and a feed-forward neural network, which are used to combine the target sequence and shallow high-dimensional features.

[0127] To further illustrate the performance of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning, 4 of the most common intraoperative waveform data are used to predict intraoperative hypotension, namely arterial blood pressure (ABP), electrocardiogram (ECG), photoplethysmogram (PPG), and end-tidal carbon dioxide (ETCO2) waveforms; the preoperative static baseline data included are: age, gender, BMI, ASA grade, whether it is an emergency, surgical method (open / closed), patient position (supine / non-supine), anesthesia method (general anesthesia / non-general anesthesia), and comorbidities (hypertension, diabetes, arrhythmia), as well as preoperative laboratory test results, hemoglobin content, albumin content, platelet count, activated partial thromboplastin time, Na + 、K + 、glucose, alanine aminotransferase, aspartate aminotransferase, urea nitrogen, creatinine, etc., and experiments were carried out 5 minutes, 10 minutes, and 15 minutes before the hypotensive event; the deep learning methods for comparison include single CNN, LSTM, and TCN; the overall test results are shown in Table 1 below; it is shown that it is significantly better than the other three deep learning methods in all indicators;

[0128] Table 1 Overall test results of the device

[0129]

[0130] The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning introduces an adaptive filtering module and a multi-channel deep learning model, which can effectively utilize multi-channel physiological signals and perform well in processing noise and redundant data; this device has high prediction accuracy and robustness, is suitable for real-time hypotension prediction in clinical surgery, helps to take intervention measures in advance, and improves the postoperative prognosis of patients.

[0131] The above are only the preferred embodiments of the present invention, and are not any other form of limitation to the present invention. Any modification or equivalent change made according to the technical essence of the present invention still belongs to the scope claimed by the present invention.

Claims

1. An intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning comprises a main body (1), a touch screen (2), a switch button (3), a mobile handheld device (4), a function button (5), a handheld device screen (6), a data input port (7), a USB port (8), a high-decibel buzzer (9), a power port (10), a network cable port (11), a general auxiliary device port (12) and a data output port (13), wherein: The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning is provided with a main body (1), a mobile handheld device (4) is provided on the left side of the front side of the main body (1), a touch screen (2) and a switch button (3) are provided on the right side of the front side of the main body (1), and the switch button (3) is located below the touch screen (2); a handheld screen (6) is provided on the upper part of the mobile handheld device (4), a function button (5) is provided in the middle part, and a data input port (7) is provided in the lower part; three groups of network cable interfaces (11) are provided in the middle part of the rear side of the main body (1), a universal auxiliary device interface (12) is provided below the network cable interface (11), and a data output port (13) and two groups of USB interfaces (8) are provided below the universal auxiliary device interface (12); a high-decibel buzzer (9) is provided at the upper right corner of the rear side of the main body (1), and a power interface (10) is provided at the lower left corner of the rear side of the main body (1).

2. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning according to claim 1, characterized in that; The system of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning adopts an adaptive filtering algorithm for pre-data processing; and uses a complex algorithm formed by deep learning model training to analyze the pre-processed data, predicting the patient's physiological changes after 5 minutes for the doctor to make clinical judgment and treatment. If it is predicted that the patient is in danger, the alarm device will be activated immediately, triggering a high-decibel buzzer (9) to remind the clinician to deal with it in time.

3. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning according to claim 1, characterized in that: The specific steps of the system operation of the intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning are as follows: Step 1: Data preprocessing: Preprocess the patient's preoperative static baseline data and intraoperative dynamic waveform data to obtain an input matrix that the model can accept. Step 2: Adaptive filter module processing: After data preprocessing, input matrix Enter the adaptive filtering module for further processing; the main function of the adaptive filtering module is to extract frequency domain features and suppress noise; Step 3: Multi-channel deep learning modeling: Use multi-channel deep learning to model using CNN model and Transformer model; Step 4: Perform data feature fusion: concatenate the output of Transformer with the preoperative static features in one dimension to form the feature vector required for the final prediction; Step 5: Enter the classification module layer for prediction; input the fused feature vector into the fully connected layer and output the predicted probability of a hypotensive event through the Sigmoid activation function; according to the predicted probability, set a threshold to determine whether a hypotensive event has occurred; if the predicted probability is greater than the threshold, it is determined to be a hypotensive event; otherwise, it is determined to be a non-hypotensive event; the final model can indicate whether a hypotensive event will occur in the next 5 minutes, 10 minutes or 15 minutes.

4. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning according to claim 3, characterized in that: The static and dynamic waveform data processing of the data preprocessing in step 1 is specifically as follows: Static baseline data processing: 1) Standardization: Use the Z-score standardization method to convert numerical features into a form with a mean of 0 and a standard deviation of 1; 2) Missing value imputation: use the mean or median to fill in missing baseline data; Dynamic waveform data processing: 1) Outlier filtering: exclude data segments with MAP < 20 mmHg or MAP > 160 mmHg; 2) Sliding window filling: Use interpolation to fill in missing time points to ensure that each segment of data is continuous and complete; 3) Data segmentation: split the original data into time windows of fixed length; 4) Label definition: Label the data of each time window based on whether a hypotensive event occurs within the next 5 minutes: If a hypotensive event occurs within the next 5 minutes, the label is 1, a positive sample; If no hypotensive event occurs in the next 5 minutes, the label is 0, a negative sample.

5. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning according to claim 3, characterized in that: The adaptive filtering module processing steps in step 2 are specifically as follows: 1) Select input sequence: use 30-second waveform data sequence as input; at the time point 5 minutes before the occurrence of hypotension event T, extract the waveform data of 30 seconds before this time point, that is, the data from T-5:00 to T-4:30; similarly, for the time points of 10 minutes and 15 minutes, the corresponding 30-second waveform data will also be extracted; 2) Frequency domain feature extraction and noise suppression: The input 30-second waveform data will pass through the adaptive filtering block, which can extract frequency domain features from the time domain signal and suppress noise through an adaptive filtering mechanism.

6. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning according to claim 5, characterized in that: The frequency domain feature extraction and noise suppression; the specific steps are as follows: First, Fourier transform: transform the input multi-channel time domain signal Convert to frequency domain signal X(f)=FFT(x(t)) in: is the original waveform data; C is the number of channels; T is the number of time samples; F is the number of frequency components; Secondly, adaptive filtering: introducing a trainable parameter matrix And constrain it to the range of [0,1] through the sigmoid activation function to generate an importance mask: The mask emphasizes relevant frequency components and suppresses irrelevant noise, and is used to dynamically adjust the weight of each frequency component; the filtered frequency domain signal is then calculated by element-wise multiplication: X'(f)=X(f)⊙σ(W(f)) Among them: ⊙ represents element-by-element multiplication; Finally, inverse Fourier transform: the filtered frequency domain signal is transformed through inverse Fourier transform Convert back to time domain to get the enhanced time domain signal.

7. The intraoperative hypotension prediction device based on adaptive filtering and multi-channel deep learning according to claim 3, characterized in that: The CNN model and Transformer model of the multi-channel deep learning modeling in step 2 are specifically: CNN model: After being processed by the adaptive filtering module, the data enters the CNN model for feature extraction; the CNN module consists of five convolutional layers, each with a kernel size of 10 and a step size of 1, which is used to learn and extract high-frequency details of the signal; the convolutional layer extracts local features through a sliding window, which can effectively capture short-term dependencies and perform preliminary feature extraction on the input time series data; each convolutional layer is followed by batch normalization and ReLU activation functions to accelerate the training process and improve the stability of the model; after the convolutional layer, a maximum pooling layer with a step size of 2 is used to reduce the data dimension and reduce the amount of calculation; at the same time, a Dropout layer is added after each convolutional layer and pooling layer to prevent the model from overfitting. Transformer model: The features extracted by the CNN model are then input into the Transformer model for modeling global dependencies; the Transformer model consists of multiple encoder and decoder layers, each encoder layer contains a multi-head self-attention mechanism and a feedforward neural network; the specific process is: 1) Position encoding: The feature sequence output by CNN is encoded with time information to ensure that the model can capture the temporal relationship in the sequence; the position encoding formula is: in: k∈{1,…,[d model / 2]}; d model is the dimension of the model; 2) Multi-head self-attention mechanism: The core of Transformer is the multi-head self-attention mechanism, which is used to globally model the input sequence; each attention head independently calculates the attention weight, and the results are concatenated and linearly transformed to obtain the final attention output; the calculation formula of multi-head attention is: MultiHead(Q,K,V)=Concat(head1,head2…,head h )W O For each head h, compute the query, key, and value matrices: At the same time, the attention weights are calculated and applied to the value matrix: d k is the dimension of the key; 3) Encoder layer: The Transformer model contains four identical encoder layers, each of which consists of a multi-head self-attention mechanism and a feed-forward neural network for further feature extraction and fusion; 4) Decoder layer: The output of the encoder is passed through the decoder layer to generate the final prediction result; the decoder layer also contains a multi-head self-attention mechanism and a feedforward neural network to combine the target sequence and shallow high-dimensional features.

Images