A sepsis analysis system based on sublingual microcirculation imaging and PPG data

Abstract

This invention discloses a sepsis monitoring system based on microcirculation imaging and PPG data, belonging to the field of disease diagnosis and monitoring technology. Addressing the problem of data distortion caused by pressure differences between the base of the tongue and the measuring device during existing sublingual microcirculation imaging measurements, this invention compensates for data under different pressures through algorithmic processing and increases the data volume by measuring image data under different pressures. A pressure sensing and correction module is introduced, using pressure detection, uniformity assessment, and a quantification correction formula to correct pressure deviations between the acquired microcirculation data and PPG data. This invention fundamentally solves the problems of acquisition deviation and pressure interference, significantly improving the accuracy and repeatability of early sepsis identification and dynamic monitoring. It has advantages such as being non-invasive, bedside, rapid, artifact-resistant, and quantifiable, making it suitable for early warning and refined resuscitation management of high-risk sepsis patients in ICUs and emergency departments.

Description

Technical Field

[0001] This invention relates to the field of critical care medicine and bedside monitoring technology, specifically to the field of early identification of sepsis-related microvascular / microcirculatory damage based on microcirculation imaging and other auxiliary monitoring. Background Technology

[0002] Sepsis patients commonly experience microvascular endothelial dysfunction, increased heterogeneity of microcirculation perfusion, and limited tissue oxygen utilization. Even if macrocirculation indicators such as mean arterial pressure reach the target, there may still be "macrocirculation-microcirculation decoupling / inconsistency," leading to persistent organ hypoperfusion.

[0003] Sublingual microcirculation imaging (such as SDF / IDF and other sublingual in vivo microscopy) can provide a direct assessment of capillary perfusion.

[0004] The core of the sublingual microcirculation image acquisition device used for sepsis monitoring is to achieve non-invasive, real-time visualization of blood flow in sublingual microvessels through optical imaging technology, providing microscopic evidence for the diagnosis and treatment of sepsis. Its acquisition principle is based on the optical absorption characteristics of hemoglobin, and it mainly relies on three core technologies: OPS, SDF, and IDF. Although the optical path designs of the three are different, their core logic is the same.

[0005] As the third-generation microcirculation monitoring technology, IDF technology is an optimization based on SDF. It adopts a coaxial incident + digital dark field design and a 548nm ultra-short pulse light source to further improve resolution, reduce pressure interference, and make the equipment lighter and more suitable for bedside monitoring.

[0006] The core physical basis of the equipment's acquisition is the strong absorption of green light in the 530–548nm wavelength band by hemoglobin (oxyhemoglobin and deoxyhemoglobin), while the surrounding tissues such as the sublingual mucosa and plasma have high transmittance and weak absorption of green light in this wavelength band. This difference naturally forms an imaging contrast of "dark blood vessels (red blood cell-rich areas) + bright background (surrounding tissues)," providing a basis for microvascular identification.

[0007] However, in practical applications, the image quality acquired by IDF is greatly affected by factors such as the operator and probe pressure, which can easily lead to false detections. Summary of the Invention

[0008] To address the aforementioned problems, this invention provides a sepsis monitoring system that effectively considers (compensates) pressure and corroborates it with PPG data.

[0009] Specifically, the present invention provides a sepsis monitoring system based on sublingual microcirculation imaging, comprising: a microcirculation data acquisition unit, a pressure measurement device, a PPG signal acquisition unit, and a sepsis identification and decision output unit.

[0010] The microcirculation data acquisition unit is used to acquire or receive sublingual microcirculation imaging data of the target patient. At least one pressure measuring device is installed on the probe surface of the microcirculation data acquisition unit to measure the contact pressure between the tongue surface and the microcirculation data acquisition unit. During measurement, the contact pressure between the probe and the tongue of the target patient is finely adjusted to obtain microcirculation imaging data under multiple contact pressure conditions.

[0011] The PPG signal acquisition unit is used to acquire PPG signal data of the target patient; the sepsis identification and decision output unit contains a pre-trained sepsis identification model. The identification model receives contact pressure data measured by the pressure measuring device, microcirculation imaging data of the patient acquired by the microcirculation data acquisition unit under the same pressure, and PPG signal data measured by the PPG signal acquisition unit. It uses the microcirculation imaging data and PPG signal data to identify sepsis. The microcirculation imaging data includes microcirculation imaging data measured under different pressures. The microcirculation imaging data measured under different pressures are weighted according to corresponding weights. The identification model combines the pressure-weighted microcirculation imaging data and PPG signal data for model training.

[0012] Preferably, the sepsis identification model includes: an image feature extraction module, a pressure compensation module, a PPG time-series feature extraction module, a feature fusion module, and a decision output module. The image feature extraction module employs a CNN model with multiple convolutional and pooling layers, receives the measured microcirculation imaging data, extracts features from the microcirculation imaging data, and generates corresponding feature maps. The pressure compensation module receives pressure data measured by a pressure measuring device and constructs a pressure compensation weight feature matrix. k represents the weight parameter number, and P is the normalized current pressure measurement value.

[0013] Add attention weights to each channel in the CNN model. Calculate the weights for each channel using two fully connected layers, using the following formula: ,in The activation function is sigmoid, and FC1 and FC2 represent fully connected processing, respectively. Represents the pixel features of this channel; ReLU is a non-linear activation function.

[0014] For each pressure measurement, the attention weight W of the data under the current pressure measurement is... Z Pressure compensation weight W P Multiplying the CNN feature map by the attention-weighted image features yields the following formula: The feature maps after stress compensation and attention weighting are dimensionality-reduced and expanded into 1-dimensional feature vectors with dimensions of 16×16×256=65536, which are then used as image feature outputs. Principal component analysis is then used to further reduce the dimensionality of the expanded 1-dimensional features, resulting in 128-dimensional image features. .

[0015] Preferably, the PPG time series feature extraction module uses an LSTM network for feature extraction. This step includes: preprocessing the PPG data; using a two-layer LSTM structure to perform deep extraction of PPG time series features to obtain 128-dimensional PPG feature data; and calculating several dimensions of manual features, including one or more of the following: heart rate, heart rate variability, pulse amplitude, rise time, and fall time.

[0016] The feature fusion module integrates the 128-dimensional image features obtained by the principal component analysis method. The fused features were obtained by stitching together 128-dimensional PPG feature data and manually generated features from several dimensions.

[0017] The features output by the feature fusion module are fed into the decision output module for classification. The feature fusion module uses fully connected layers for classification.

[0018] The first dimensionality reduction is performed using a third fully connected layer (FC3): 261 → 64, with ReLU activation and dropout of 0.3. A second dimensionality reduction is then performed using a fourth fully connected layer (FC4): 64 → 1, with Sigmoid output probability. Model loss function:

[0019]

[0020] Sepsis probability output:

[0021] ;

[0022] FC3 and FC4 represent fully connected processing, with a threshold θ = 0.5. Determined to be at high risk of sepsis; Determined to be at low risk of sepsis. These represent microcirculation characteristics and ppg characteristics, respectively. The threshold can be adjusted based on advancements in medical research.

[0023] Beneficial effects

[0024] This invention addresses the problem of data distortion caused by pressure differences between the base of the tongue and the measuring device during existing sublingual microcirculation imaging measurements. Through algorithmic processing, it compensates for data under different pressures and increases the amount of image data by measuring images under varying pressures, thus providing more data support for subsequent model training. It also introduces pressure sensing and correction to correct pressure deviations in the acquired microcirculation data. This invention fundamentally solves the problems of acquisition bias and pressure interference, significantly improving the accuracy and repeatability of early sepsis identification and dynamic monitoring. It has advantages such as high efficiency, accuracy, and strong reference value, and is suitable for early warning and refined resuscitation management of high-risk sepsis patients in ICUs and emergency departments. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the architecture of the sepsis monitoring system of the present invention;

[0026] Figure 2 This is a schematic diagram of the architecture of the sepsis monitoring system of the present invention;

[0027] Figure 3 A schematic diagram of the identification process in the sepsis monitoring system of the present invention;

[0028] Figure 4-5 These are measured images of the microcirculation of the same patient under different pressures, collected by the sepsis monitoring system of the present invention. Detailed Implementation

[0029] like Figure 1 As shown, for any patient, two aspects of data collection are required: sublingual microcirculation data collection using a microcirculation data acquisition unit, and PPG signal acquisition simultaneously or subsequently.

[0030] In one implementation, such as Figure 2 As shown, the sepsis monitoring system architecture based on sublingual microcirculation imaging of the present invention includes: a microcirculation data acquisition unit, a pressure measurement device, a PPG signal acquisition unit, and a sepsis identification and decision output unit. The data acquisition unit is used to acquire or receive sublingual microcirculation imaging data. The sublingual microcirculation imaging device of the present invention is equipped with at least one pressure measurement device, or pressure sensor, which is used to measure the contact pressure between the tongue surface and the device. The sublingual microcirculation imaging data is measured using a third-generation dark-field imaging device (IDF). One to three pressure sensors are set on the edge or side area of ​​the device surface. The probe is gently placed against the patient's sublingual fold, keeping the tongue naturally relaxed. The device is irradiated with a 548nm pulsed light source.

[0031] The pressure measuring device employs one or more pressure sensors to measure pressure data simultaneously with the probe measurement. Preferably, the monitoring system further includes a pressure data processing module, which receives the contact pressure measured by the pressure measuring device at one or more locations on the tongue, eliminates outlier pressure values, and generates pressure correction data. In one implementation, the pressure measuring device uses an embedded thin-film pressure sensor to directly acquire the contact pressure P between the probe and the tongue mucosa.

[0032] The sepsis prediction model includes: an image feature extraction module, a stress compensation module, and a PPG time series feature extraction module.

[0033] like Figure 3 As shown, the working process of the system of the present invention will be described in detail below with reference to the above system architecture.

[0034] First, data measurement and acquisition are performed. During measurement, pressure is fine-tuned several times, adjusting the pressure value between the tongue and the acquisition probe to obtain different image data.

[0035] In this embodiment, the image feature extraction module adopts a CNN convolutional neural network model. The feature extraction process is as follows: a 4-layer structure of convolutional layers plus pooling layers is used to extract the spatial features of the image. The input image is a 256×256×3 RGB image, which is sequentially input into convolutional layer 1, pooling layer 1, convolutional layer 2, pooling layer 2, convolutional layer 3, pooling layer 3, convolutional layer 4, and pooling layer 4. The first convolutional layer has three input channels, corresponding to R, G, and B data respectively, and 32 output channels. The convolution kernel size is... The first convolutional layer has a 3x3 kernel size, stride = 1, padding = 1, and uses ReLU activation. The second convolutional layer has 32 input channels and 64 output channels, with a 3x3 kernel size and ReLU activation. Similarly, the third convolutional layer has 64 input channels and 128 output channels, and the fourth convolutional layer has 128 input channels and 256 output channels. All four pooling layers have the same structure, using 2x2 kernels with a stride of 2. After four convolutional and pooling layers, the output is a 16x16x256 feature map F. cnn .

[0036] Channel attention weighting and pressure weight compensation are applied to the extracted feature maps.

[0037] For the output feature map (16*16*256), the attention weights for each channel are calculated. In one implementation, the pixel features of each channel are first calculated: Where H=16 and W=16 are the feature map sizes, and c is the number of channels (256).

[0038] The weight of each channel is calculated using two fully connected (FC) layers, using the following formula: ,in FC1 and FC2 represent fully connected processing, respectively, and are Sigmoid activation functions.

[0039] The pressure compensation module is used to construct the pressure compensation weight feature matrix. , k represents the weight parameter number, and P is the normalized current pressure measurement value. The weights W in the matrix elements... k All initial values ​​are set to 1, and the learning process is continuously iterated and optimized.

[0040] For each set of data under stress, the attention weight W is... Z Pressure compensation weight W P Multiplying the CNN feature map by the double-weighted image features yields the double-weighted features, as shown in the formula: ;

[0041] The feature map (16×16×256) after stress compensation and attention weighting is reduced in dimension and expanded into a 1-dimensional feature vector with a dimension of 16×16×256=65536, which is used as the image feature output.

[0042] like Figure 4 and Figure 5 The image shows sublingual microcirculation images of a sepsis patient under different measurement pressures. The comparison of the images shows that there are significant differences in the clarity of features and blood flow concentration in the images measured under different pressures. Therefore, it is very necessary to compensate for different pressures.

[0043] To reduce dimensionality while retaining core information, this invention employs principal component analysis to perform linear dimensionality reduction on high-dimensional image features.

[0044] Zero-mean normalization was applied to the flattened 65536-dimensional image features:

[0045] Where μ is the mean vector of the features in the training set images, and σ' is the standard deviation vector of the features in the training set images.

[0046] Then, calculate the covariance matrix.

[0047]

[0048] Where N is the number of training samples, and Σ is the covariance matrix with dimensions 65536×65536; eigenvalue decomposition is performed on the covariance matrix:

[0049]

[0050] Where U is the eigenvector matrix and Λ is the eigenvalue diagonal matrix; eigenvalues ​​are sorted from largest to smallest, with larger eigenvalues ​​indicating richer micro-circulation information carried in the corresponding direction. The eigenvectors corresponding to the first 128 largest eigenvalues ​​are selected to form the dimension-reduced projection matrix:

[0051] ;

[0052] The standardized high-dimensional image features are projected into the low-dimensional principal component space:

[0053] ;

[0054] Output 128-dimensional image features ;

[0055] (ii) At the same time, or before or after, collect PPG data and extract features from the PPG data to obtain PPG-based features to be fused.

[0056] PPG feature extraction was performed using an LSTM network. The input to this network was PPG time-series data collected by the patient's wearable device (sampling frequency 100Hz, single input duration 10 seconds, total 1000 data points). The extraction process is as follows:

[0057] 1) Preprocess the PPG data, including at least noise reduction, such as using adaptive filtering (Kalman filtering) to eliminate motion artifacts and power frequency interference (50Hz) on the PPG signal; then perform normalization: normalize the PPG signal amplitude to between 0 and 1.

[0058] ,in, The amplitude of the PPG timing signal is the normalized value, where t represents the time point of the timing data. After normalization, the amplitude range is uniformly mapped to [0,1]. Let be the original PPG signal amplitude at time point t. The minimum amplitude in a single input PPG time series data (10 seconds, 1000 data points), This represents the maximum amplitude value in a single input PPG time series data.

[0059] 2) Next, temporal feature extraction (LSTM) is performed. In this embodiment, a two-layer LSTM (Long Short-Term Memory) network structure is used for deep extraction of PPG temporal features. The feature extraction layer receives the preprocessed PPG temporal data as a single-channel one-dimensional temporal signal.

[0060] The first LSTM layer has an input dimension of 1 (corresponding to a single-channel PPG time-series signal), a hidden layer dimension of 64, a dropout coefficient of 0.2, and is set to return sequences (return_sequences=True). The hidden layer dimension of 64 indicates that this LSTM network learns different features of the PPG time-series data in parallel through 64 hidden units (such as the rising segment, falling segment, and peak features of the pulse waveform). The dropout coefficient of 0.2 indicates that 20% of the hidden layer neurons are randomly dropped during training to avoid overfitting and improve the model's generalization ability. The return sequence is set to True because the output of LSTM1 needs to be used as the input of the second LSTM layer (LSTM2), requiring the preservation of complete time-series information rather than just outputting the features of the last time step, ensuring the continuous transmission of temporal dependencies.

[0061] The gating mechanism of LSTM1 works as follows: LSTM uses three gate control units—forget gate, input gate, and output gate—to selectively remember and forget time-series information, adapting to the dynamic changes in PPG time-series data. Specifically, the forget gate (controlled by the sigmoid activation function, with output value ∈...) The first gate is used to determine whether to discard the hidden state information from the previous time step. When the output value is close to 1, the information from the previous time step is retained; when it is close to 0, redundant information is discarded, which can effectively filter out the slight noise remaining in the PPG signal. The second gate is used to control which parts of the input signal at the current time step need to be updated to the cell state. Valid information is filtered through the sigmoid activation function, and then candidate cell states are generated through the tanh activation function. Combined with the cell state from the previous time step after processing by the forget gate, the cell state at the current time step is updated to obtain the cell state at the current time step. The third gate is used to control which parts of the cell state at the current time step need to be output to the hidden state. Output information is filtered through the sigmoid activation function, and then the cell state is normalized through the tanh activation function to obtain the hidden state at the current time step.

[0062] The first layer of LSTM1 processes the 1000-step PPG time-series data step-by-step using the aforementioned gating mechanism, with an output dimension of... The time-series feature sequence, that is, each time step corresponds to a 64-dimensional feature vector, which contains the local time-series features of the PPG signal at that moment.

[0063] The second LSTM layer has an input dimension of 64 (corresponding to the 64-dimensional features of each step of the LSTM1 output), a hidden layer dimension of 128, a dropout coefficient of 0.2, and is set to not return sequences (return_sequences=False). Increasing the hidden layer dimension to 128 is to further fuse and deeply mine the local temporal features output by LSTM1, capturing more complex long-distance temporal dependencies (such as the correlation features of pulse waveforms at different times, the dynamic trend of heart rate changes, etc.); the dropout coefficient is kept at 0.2 to continuously suppress overfitting; setting "not return sequences" to False is because after feature extraction by two LSTM layers, a fixed-dimensional deep temporal feature vector needs to be output for subsequent concatenation with manually generated features. Therefore, only the hidden state of the last time step is retained as the output of LSTM2, with an output dimension of... That is, a 128-dimensional deep temporal feature vector.

[0064] LSTM2 Feature Fusion and Output Process: LSTM2 receives the output of LSTM1. After processing the temporal feature sequence, the system uses its own gating mechanism to process the local features at each time step, fusing them to capture dynamic patterns related to sepsis in the PPG time-series data (such as abnormal heart rate fluctuations and unstable pulse amplitude commonly seen in sepsis patients). The final output is a 128-dimensional deep temporal feature vector. This feature vector integrates global dynamic information from the PPG time-series data, effectively reflecting the patient's microcirculation perfusion status and cardiovascular function, providing core temporal feature support for subsequent multimodal fusion and sepsis identification.

[0065] In addition, several additional manually calculated dimensions of features are used as supplementary features.

[0066] For example, heart rate: calculated from the time interval (pulse cycle T) between adjacent pulse wave peaks in the PPG time-series signal, measured in beats per minute, reflects the patient's cardiovascular function status. ;

[0067] Heart rate variability (HRV): Measures the degree of fluctuation in the pulse cycle. It has no unit and is an important indicator reflecting the autonomic function of the heart. Abnormal HRV is often observed in patients with sepsis.

[0068] Where T represents the time interval between two adjacent pulse peaks in the PPG time-series signal, in seconds, and i is the pulse count.

[0069] Pulse amplitude, rise time, and fall time are all parameters. Rise time represents the time it takes for the PPG pulse wave to rise from baseline to peak value, reflecting the pulse rise speed and being related to cardiovascular perfusion status; fall time represents the time it takes for the PPG pulse wave to fall from peak value to baseline, reflecting the pulse fall speed and helping to assess microcirculation perfusion.

[0070] 4) PPG Feature Concatenation: The 128-dimensional depth features are concatenated with the 5-dimensional handcrafted features to obtain 133-dimensional PPG time-series features. .

[0071] 4. Multimodal fusion and sepsis probability output

[0072] The image features (65,536 dimensions) were reduced to 128 dimensions using PCA and then concatenated with 133-dimensional PPG features to obtain 261-dimensional fused features. At this point, it's crucial to consider data volume matching. For example, if the microcirculation data volume is twice or more than the PPG data volume, then one set of PPG data should be fused separately with two sets of microcirculation data from the same patient. In other words, the features extracted from the smaller dataset are replicated and amplified proportionally to accommodate the larger dataset.

[0073] Fully connected layer classification: FC3: 2^61 → 6^4, ReLU, dropout 0.3; FC4: 6^4 → 1, Sigmoid output probability .

[0074] Channel attention weights:

[0075]

[0076]

[0077] Model loss function:

[0078]

[0079] Sepsis probability output:

[0080]

[0081] Set the threshold θ = 0.5: Determined to be at high risk of sepsis; The risk level was determined to be low for sepsis.

[0082] During model training, dual sets of data from both sepsis patients and non-sepsis patients were used to ensure the model's recognition accuracy.

[0083] Specifically, in the data collection phase, data from at least 50 confirmed sepsis patients and at least 100 non-sepsis patients were collected. For each patient, sublingual microcirculation images, probe-tongue mucosal contact pressure P, and PPG time-series data collected by wearable devices were collected simultaneously. Data from sepsis patients needed to be clinically confirmed and labeled, while data from non-sepsis patients needed to exclude sepsis and other severe diseases that might interfere with microcirculation indicators to ensure data validity.

[0084] Secondly, in the data preprocessing stage, outlier removal, standardization, feature extraction, and stress compensation operations were uniformly performed on both sets of data. The sepsis patient data was labeled as positive samples (label y=1), and the non-sepsis patient data was labeled as negative samples (label y=0). All processed samples were divided into training set, validation set, and test set in a 7:2:1 ratio. The training set was used for model parameter fitting, the validation set was used to adjust model hyperparameters (such as the number of CNN convolutional kernels, the dimension of LSTM hidden layers, the dropout coefficient of fully connected layers, etc.), and the test set was used to verify the final performance of the model.

[0085] Then, during the model training phase, the model parameters are iteratively updated using the gradient descent algorithm with binary cross-entropy as the loss function. Five-fold cross-validation is employed during training, randomly dividing the training set into five parts, alternating between four parts as the training subset and one part as the validation subset to avoid overfitting and ensure stable output results under different data distributions. Finally, in the model optimization phase, the model structure and parameters are adjusted based on performance feedback from the validation and test sets (combining metrics such as accuracy, sensitivity, specificity, and AUC) until the model performance reaches the preset evaluation criteria, ultimately resulting in a trained model suitable for clinical sepsis risk identification.

[0086] Model testing involved classifying the model using a test dataset. The test results showed that:

[0087] Prediction accuracy ≥ 97%, meaning that the overall classification accuracy is greater than 97% for both positive and negative samples;

[0088] Recall rate ≥ 97.5%, meaning that 97.5% of the true positive samples were successfully predicted.

[0089] This demonstrates that the method of the present invention can effectively predict cystic cystitis patients.

[0090] It should be noted that PPG data and sublingual microcirculation image data can be collected simultaneously or sequentially.

[0091] Although the principles of the present invention have been described in detail above with reference to preferred embodiments, those skilled in the art should understand that the above embodiments are merely illustrative explanations of the implementation of the present invention and are not intended to limit the scope of the present invention. The details in the embodiments do not constitute a limitation on the scope of the present invention. Any obvious changes, such as equivalent transformations or simple substitutions, based on the technical solutions of the present invention without departing from the spirit and scope of the present invention fall within the protection scope of the present invention.

Claims

1. A sepsis monitoring system based on sublingual microcirculation imaging and PPG data, characterized in that, include: Microcirculation data acquisition unit, pressure measurement device, PPG signal acquisition unit, sepsis identification and decision output unit, The microcirculation data acquisition unit is used to acquire or receive sublingual microcirculation imaging data of the target patient. At least one pressure measuring device is installed on the probe surface of the microcirculation data acquisition unit to measure the contact pressure between the tongue surface and the microcirculation data acquisition unit. During measurement, the contact pressure between the probe and the tongue of the target patient is finely adjusted to obtain microcirculation imaging data under multiple contact pressure conditions. The PPG signal acquisition unit is used to acquire PPG signal data of the target patient; the sepsis identification and decision output unit contains a pre-trained sepsis identification model. The identification model receives contact pressure data measured by the pressure measuring device, microcirculation imaging data of the patient acquired by the microcirculation data acquisition unit under the same pressure, and PPG signal data measured by the PPG signal acquisition unit. It uses the microcirculation imaging data and PPG signal data to identify sepsis. The microcirculation imaging data includes microcirculation imaging data measured under different pressures. The microcirculation imaging data measured under different pressures are weighted according to corresponding weights. The identification model combines the pressure-weighted microcirculation imaging data and PPG signal data for model training.

2. The sepsis monitoring system based on sublingual microcirculation imaging and PPG data according to claim 1, characterized in that, The sepsis identification model includes: an image feature extraction module, a pressure compensation module, a PPG time-series feature extraction module, a feature fusion module, and a decision output module. The image feature extraction module employs a CNN model with multiple convolutional and pooling layers, receives measured microcirculation imaging data, extracts features from the microcirculation imaging data, and generates corresponding feature maps. The pressure compensation module receives pressure data measured by a pressure measuring device and constructs a pressure compensation weight feature matrix. k represents the weight parameter number, and P is the normalized current pressure measurement value. Add attention weights to each channel in the CNN model. Calculate the weights for each channel using two fully connected layers, using the following formula: ,in The activation function is sigmoid, and FC1 and FC2 represent fully connected processing, respectively. Represents the pixel features of this channel; ReLU is a non-linear activation function. For each pressure measurement, the attention weight W of the data under the current pressure measurement is... Z Pressure compensation weight W P Multiplying the CNN feature map by the attention-weighted image features yields the following formula: The feature maps after stress compensation and attention weighting are dimensionality-reduced and expanded into 1-dimensional feature vectors with dimensions of 16×16×256=65536, which are then used as image feature outputs. Principal component analysis is then used to further reduce the dimensionality of the expanded 1-dimensional features, resulting in 128-dimensional image features. .

3. The sepsis monitoring system based on sublingual microcirculation imaging and PPG data according to claim 2, characterized in that, The PPG time series feature extraction module uses an LSTM network for feature extraction. This step includes: preprocessing the PPG data; using a two-layer LSTM structure to perform deep extraction of PPG time series features to obtain 128-dimensional PPG feature data; and calculating several dimensions of manual features, including one or more of the following: heart rate, heart rate variability, pulse amplitude, rise time, and fall time. The feature fusion module integrates the 128-dimensional image features obtained by the principal component analysis method. The fused features were obtained by stitching together 128-dimensional PPG feature data and manually generated features from several dimensions. The features output by the feature fusion module are fed into the decision output module for decision classification.

4. The sepsis monitoring system based on sublingual microcirculation imaging and PPG data according to claim 3, characterized in that, The feature fusion module uses a fully connected layer for classification. The first dimensionality reduction is performed using a third fully connected layer (FC3): 261 → 64, with ReLU activation and dropout of 0.

3. A second dimensionality reduction is then performed using a fourth fully connected layer (FC4): 64 → 1, with Sigmoid output probability. Model loss function: ； Sepsis probability output: ； FC3 and FC4 represent fully connected processing, with a threshold θ = 0.

5. Determined to be at high risk of sepsis; Determined to be at low risk of sepsis. These represent microcirculation characteristics and ppg characteristics, respectively.

Images