Blood multimodal infection monitoring methods, systems and dialysis catheters
By collecting pressure and temperature data in real time during hemodialysis, and combining time-spectrum analysis and domain adversarial neural networks to dynamically adjust thresholds, the problem of misjudgment caused by individual differences in existing technologies has been solved, achieving highly sensitive and specific infection risk early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN TIANKE MEDICAL TECH CO LTD
- Filing Date
- 2025-08-11
- Publication Date
- 2026-05-26
Smart Images
Figure CN121034669B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent sensor technology, and more specifically, to a method, system, and dialysis catheter for monitoring multimodal blood infections. Background Technology
[0002] Chinese patent CN118797245B discloses an online monitoring system and method for patients undergoing hemodialysis: acquiring and preprocessing multimodal monitoring data; using a deep learning model to fuse the preprocessed multimodal monitoring data into a comprehensive feature vector; generating a blood status evaluation matrix and behavioral status evaluation coefficients based on the fused comprehensive feature vector; constructing an infection risk prediction model based on the status values of blood components and behavioral status evaluation coefficients; optimizing the infection risk prediction model through adaptive learning; and generating an online monitoring report for hemodialysis based on the infection risk prediction results. This invention, by acquiring and preprocessing multimodal monitoring data, fusing data using a deep learning model, constructing a blood status evaluation matrix and behavioral status evaluation coefficients, establishing and continuously optimizing an infection risk prediction model, ultimately generates an online monitoring report for hemodialysis. This effectively monitors the infection risk of patients during dialysis and improves the accuracy of risk prediction by comprehensively assessing blood and behavioral status.
[0003] While the above methods can meet the needs of most scenarios, research and practical application of these methods and existing technologies have revealed at least the following shortcomings:
[0004] Existing blood infection risk prediction models employ a single-parameter, fixed-threshold design. However, hemodialysis patients exhibit significant individual differences due to underlying diseases (such as diabetes and hypertension), age, and dialysis duration. Elderly patients generally have lower baseline blood pressure, and using a uniform pressure fluctuation threshold can easily misinterpret these physiological fluctuations as infection warnings. Furthermore, long-term dialysis patients often experience autonomic nervous system dysfunction, resulting in a baseline body temperature 0.3-0.5°C lower than healthy individuals. A fixed temperature threshold can also lead to missed detection of early-stage low-grade fever (37.5-38°C) due to these individual baseline differences. This high misjudgment rate ultimately limits the accuracy of infection risk prediction.
[0005] In view of this, the present invention proposes a blood multimodal infection monitoring method, system and dialysis catheter to solve the above problems. Summary of the Invention
[0006] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a blood multimodal infection monitoring method, system, and dialysis catheter, comprising the following steps:
[0007] The dialysis catheter, which integrates intelligent sensing elements into the interventional blood vessel, intelligently collects and analyzes pressure and temperature data in real time to obtain the pressure time spectrum and temperature correction matrix.
[0008] Feature extraction is performed on the pressure-time spectrum and temperature correction matrix to obtain analytical features, including pressure features, temperature features, and cross features.
[0009] By acquiring synthetic and clinical data and combining them with transfer learning, a domain adversarial neural network is built to intelligently analyze real-time features and obtain predicted values of patient infection risk probability and inflammatory factor concentration ranges.
[0010] The system acquires the patient's baseline data and, based on the patient's infection risk probability and the predicted values of inflammatory factor concentration ranges, dynamically adjusts the pressure fluctuation tolerance bandwidth and temperature rise slope threshold using reinforcement learning to generate personalized alarm commands.
[0011] Furthermore, methods for obtaining the spectrum under pressure include:
[0012] The original pressure signal is subjected to discrete Fourier transform to obtain the pressure signal, and the pressure signal is divided into physiological fluctuation frequency band and pathological low frequency band by spectrum segmentation.
[0013] Scaling functions and wavelet functions are constructed, and pressure signals are combined to decompose the physiological fluctuation frequency band and the pathological low frequency band to obtain physiological and pathological components.
[0014] The baseline value is defined as the state vector, the pathological component is defined as the observation value, and the observation noise is defined in combination with the sensor accuracy. The current pathological component is predicted based on the state vector of the previous moment by the Kalman filter algorithm. The prediction is then updated by combining the observation value and the observation noise to obtain the corrected pathological component. The physiological component and the corrected pathological component are superimposed in the time domain to obtain the reconstructed signal.
[0015] The reconstructed signal is windowed and framed, and then a discrete Fourier transform is performed on each frame of the reconstructed signal to obtain the pressure time spectrum.
[0016] Furthermore, methods for obtaining the temperature correction matrix include:
[0017] Based on the original sensitivity coefficient, temperature drift deviation, and random noise of the temperature sensor, the sensor measurement value at each spatial location is modeled against the actual physical quantity to obtain the spatial temperature deviation.
[0018] Under the premise of preset constant physical quantities, measurements at different temperatures are collected at each spatial location, and a dataset is established for each spatial location;
[0019] For each spatial location, the temperature drift deviation is set to an L-order polynomial. The spatial temperature deviation is replaced, and the replaced spatial temperature deviation is solved by the least squares method to obtain the calibration spatial temperature deviation value and the calibration sensitivity coefficient.
[0020] The temperature drift correction matrix is obtained by subtracting the temperature drift estimate from the calibration space temperature deviation value.
[0021] The sensitivity matrix is obtained by calculating the calibration sensitivity coefficient and the normalized sensitivity, and the temperature correction matrix is obtained by calculating the sensitivity matrix and the temperature drift correction matrix.
[0022] Furthermore, methods for obtaining stress characteristics include:
[0023] Pressure characteristics include dynamic characteristics and frequency domain characteristics;
[0024] The time-frequency spectrum matrix is obtained from the pressure time-frequency spectrum. The time-frequency spectrum matrix is integrated over the respiratory frequency range to obtain the respiratory modulation intensity signal A(t). The respiratory cycle is segmented according to the peak value of A(t). The mean and standard deviation of the pulse pressure sequence in each respiratory cycle are statistically analyzed, and the statistical coefficient of variation is calculated.
[0025] After differentiating the pressure signal, SIFT is performed to obtain the time spectrum of the derivative. Energy abrupt change points are found in the preset high-frequency region to obtain the rising and falling edges of the contraction period. The corresponding intervals are truncated in the time domain to calculate the slope of the contraction period.
[0026] The kinetic characteristics are obtained by concatenating the statistical coefficient of variation and the slope during contraction.
[0027] Extract the power H1 at the fundamental frequency and the power H0 of the DC component of each time spectrum in the pressure time spectrum, and calculate the proportion of the first harmonic energy;
[0028] For each time point in the pressure spectrum, the power spectral density is integrated within a preset frequency range to calculate the low-frequency power and obtain the low-frequency power integral.
[0029] The frequency domain characteristics are obtained by splicing together the proportion of first harmonic energy and the low-frequency band power integral.
[0030] Furthermore, the cross features include mutual information entropy and two-dimensional spatiotemporal correlation images. Methods for obtaining cross features include:
[0031] The pressure and temperature signals are denoised and normalized to obtain standard pressure and temperature signals.
[0032] The standard pressure signal value range is divided into N1 intervals, and the standard temperature signal value range is divided into N2 intervals. Each interval contains an equal number of data points. The standard pressure signal and standard temperature signal are mapped to the corresponding intervals to obtain discrete pressure sequence and discrete temperature sequence.
[0033] The frequency of standard pressure signals and standard temperature signals falling into each interval combination (i,j) at the same time is statistically analyzed, where i∈[1,N1] and j∈[1,N2]. After normalization, the joint probability distribution is obtained, the marginal probability is calculated, and the mutual information entropy is obtained based on the joint probability distribution and the marginal probability.
[0034] The standard pressure signal and standard temperature signal are converted into an angle sequence using the inverse cosine function;
[0035] The angle sequence is converted into a two-dimensional matrix of length M using trigonometric identities, thus obtaining the GAF matrix of the pressure signal and the GAF matrix of the temperature signal.
[0036] The GAF matrix of the pressure signal and the GAF matrix of the temperature signal are used as the first two channels of the RGB channel to form an M×M×2 input image, thus obtaining a two-dimensional spatiotemporal correlated image.
[0037] Furthermore, methods for obtaining the patient's infection risk probability and predicted values of inflammatory factor concentration ranges include:
[0038] Generative analysis features of the source domain are generated by pre-setting an external pulsating flow system;
[0039] Collect monitoring data and extract features to obtain analytical features of the target domain;
[0040] The source and target domain data are preprocessed, and cross-domain sample pairs are generated through mixed-precision training.
[0041] A shared feature extractor is built using 3D ResNet-50. The analyzed features are used as input to the shared feature extractor to obtain high-dimensional feature vectors.
[0042] A domain adversarial neural network was constructed, and source domain data, target domain data, and high-dimensional feature vectors were used as inputs to obtain predicted values of patient infection risk probability and inflammatory factor concentration ranges.
[0043] Methods for obtaining high-dimensional feature vectors of the source domain include:
[0044] Different enhancement results of the same analytical feature in the source domain data are used as positive samples, and the projected features of any enhancement results of other analytical features are used as negative samples, forming a mixed batch according to a preset batch size and preset ratio.
[0045] A normalized temperature-scale cross-entropy loss is constructed based on the cosine similarity of positive sample pairs and the temperature hyperparameter.
[0046] Samples are extracted from the mixed batch as input to the encoder to obtain high-dimensional feature vectors of the source domain.
[0047] Furthermore, training methods for domain adversarial neural networks include:
[0048] Construct a discriminator with dual task headers, including a domain classification header and an auxiliary task header;
[0049] The domain classification head outputs the domain discrimination probability, which is used to determine whether a high-dimensional feature vector comes from the source domain or the target domain;
[0050] The auxiliary task header is used to predict infection labels from source domain data;
[0051] A gradient inversion layer is inserted between the shared feature extractor and the discriminator. During training, gradient inversion is applied to the domain classification loss so that the shared feature extractor learns domain-invariant features.
[0052] The overall task loss function is designed to update the discriminator and shared feature extractor until a preset optimal performance condition is reached. The domain adversarial neural network corresponding to the optimal performance condition is then used as the final output domain adversarial neural network. The overall task loss function is calculated from the task loss, domain adversarial loss, and auxiliary loss. The task loss is calculated as the total task loss of the target domain and the source domain. The task loss is calculated by quantifying the deviation between the predicted probability and the true label using cross-entropy loss, and then adding the mean squared error loss that quantifies the deviation between the predicted interval and the true value. The domain adversarial loss is calculated by quantifying the error of domain classification using binary cross-entropy loss. The auxiliary loss is calculated by quantifying the error of infection prediction within the source domain using cross-entropy loss.
[0053] Furthermore, methods for generating personalized alarm commands include:
[0054] Define a state space, which includes a state s. The state s consists of real-time features, historical data, and patient baseline. Real-time features include pressure features, temperature features, and cross features. Historical data includes the number of alarms, true positives, false positives, and false negatives within a historical time period.
[0055] Define the action space, which includes action vectors. The action vectors consist of the pressure fluctuation tolerance bandwidth and the temperature rise slope threshold. The pressure fluctuation tolerance bandwidth is set to E1 level based on the current patient baseline. The temperature rise slope threshold is set to E2 level based on the current patient baseline.
[0056] The threshold adjustment is triggered once every preset time interval, and the action is selected according to the current state s to update the alarm threshold;
[0057] A reward function is defined, which is obtained by weighting accuracy reward, stability reward and timeliness reward. Among them, accuracy reward is obtained by weighting the number of true positives, false positives and false negatives with preset true positive reward, false positive penalty and false negative penalty; stability reward is obtained by cosine similarity between current feature and feature of previous time step; timeliness reward is obtained by time difference from the occurrence of real infection to system alarm.
[0058] An Actor-Critic architecture is adopted, where state s serves as the input to the Actor network to obtain the action probability distribution; state s also serves as the input to the Critic network to obtain the state value; the Actor network uses log probability to measure the policy probability of choosing action a in state s, and uses an advantage function to evaluate the advantage of choosing action a in state s compared to the average case; the Critic network uses the sum of discounted estimates of the current reward and the future cumulative reward as the target of the state value, calculates the squared difference between the current state value estimate and the target, and measures the prediction error of the value network; the action with the highest state value is selected as the final action to be executed, and the stress fluctuation tolerance bandwidth and temperature rise slope threshold are obtained.
[0059] A blood-based multimodal infection surveillance system, implementing the aforementioned blood-based multimodal infection surveillance method, including:
[0060] Acquisition and Analysis Module: The dialysis catheter with integrated intelligent sensing elements in the interventional blood vessel intelligently acquires and analyzes pressure and temperature data in real time to obtain pressure time spectrum and temperature correction matrix.
[0061] Feature extraction module: Extracts features from the pressure-time spectrum and temperature correction matrix to obtain analytical features, including pressure features, temperature features, and cross features;
[0062] Monitoring and Analysis Module: Acquires synthetic and clinical data, combines transfer learning to build a domain adversarial neural network, performs intelligent analysis on real-time features, and obtains predicted values of patient infection risk probability and inflammatory factor concentration range.
[0063] Dynamic alarm module: Obtain patient baseline, predict patient infection risk probability and inflammatory factor concentration range based on reinforcement learning, dynamically adjust pressure fluctuation tolerance bandwidth and temperature rise slope threshold, and generate personalized alarm commands.
[0064] Dialysis catheters are used in the aforementioned blood multimodal infection monitoring method.
[0065] The technical effects and advantages of the blood multimodal infection monitoring method, system, and dialysis catheter of this invention are as follows:
[0066] This invention simultaneously collects pressure and temperature data through a dialysis catheter inserted into the blood vessel. After time-spectrum analysis and calibration matrix processing, multi-dimensional analytical features are extracted. Combining transfer learning and domain adversarial neural networks, and utilizing synthetic data pre-training and clinical data fine-tuning, it overcomes the limitations of small samples and improves cross-domain generalization ability, enabling real-time prediction of patient infection risk probability and inflammatory factor concentration. Based on patient baseline data, it dynamically adjusts the pressure fluctuation tolerance bandwidth and temperature rise slope threshold through reinforcement learning to generate personalized alarm commands. Compared with traditional single-parameter monitoring, this invention significantly improves the sensitivity and specificity of infection detection, can provide early warning of infection risk, and reduces false alarm and false negative rates. It provides an intelligent and personalized solution for early intervention of blood infections, effectively shortening the diagnosis time. Attached Figure Description
[0067] Figure 1 This is a schematic diagram of the blood multimodal infection monitoring method of the present invention;
[0068] Figure 2 This is a schematic diagram of the data flow in this invention;
[0069] Figure 3 This is a schematic diagram of the method for locating suspected foci of infection according to the present invention;
[0070] Figure 4 This is a schematic diagram of the blood multimodal infection monitoring system of the present invention. Detailed Implementation
[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0072] Example 1:
[0073] Please see Figure 1 , Figure 2 As shown, this embodiment provides a blood multimodal infection monitoring method, including the following steps:
[0074] The dialysis catheter, which integrates intelligent sensing elements into the interventional blood vessel, intelligently collects and analyzes pressure and temperature data in real time to obtain the pressure time spectrum and temperature correction matrix.
[0075] The pressure-time spectrum extracts dynamic features such as pulse pressure variability and systolic slope through frequency domain analysis, capturing vasodynamic abnormalities caused by infection. The temperature correction matrix eliminates sensor noise and environmental interference through spatiotemporal calibration, accurately extracting features such as the temperature difference between the puncture point and the tip, and the rate of temperature rise, ensuring the spatiotemporal consistency of temperature data. This enables the system to identify signs of infection from both hemodynamic and thermometrological abnormalities, providing cross-modal feature support for infection risk prediction, dynamic threshold adjustment, and infection focus localization. This improves the sensitivity and specificity of the monitoring system for early infection from the data source.
[0076] Methods for obtaining the spectrum of pressure include:
[0077] A 16-channel MEMS piezoresistive sensor is used, which is circumferentially distributed on the surface of the dialysis catheter. The radial pressure distribution data of the blood vessel wall is acquired through the dialysis catheter of the interventional blood vessel to obtain the raw pressure signal.
[0078] The original pressure signal is subjected to discrete Fourier transform to obtain the pressure signal. The pressure signal is then divided into physiological fluctuation frequency bands (e.g., 0.1Hz-2Hz) and pathological low frequency bands (e.g., 0-0.05Hz) according to a preset frequency range.
[0079] A scaling function is constructed to cover the pathological low-frequency band, and a wavelet function is constructed to cover the physiological fluctuation band. The scaling function and the wavelet function meet the following preset conditions: their frequency bands are non-overlapping and complementary; their energies satisfy the Parseval identity and orthogonality in the frequency domain; the scaling function exhibits high-order continuous change; the wavelet function has an L-order vanishing moment; the scaling function and the wavelet function must satisfy the recursive relationship of multi-resolution analysis; and the frequency domain can be completely reconstructed. Based on the scaling function, the wavelet function, and the pressure signal, the physiological fluctuation band and the pathological low-frequency band are decomposed to obtain the physiological and pathological components.
[0080] The baseline value is defined as the state vector, the pathological component is defined as the observation value, and the observation noise is defined in combination with the sensor accuracy. The current pathological component is predicted based on the state vector of the previous moment by the Kalman filter algorithm. The prediction is then updated by combining the observation value and the observation noise to obtain the corrected pathological component. The physiological component and the corrected pathological component are superimposed in the time domain to obtain the reconstructed signal.
[0081] The reconstructed signal is windowed and framed, and then a discrete Fourier transform is performed on each frame of the reconstructed signal to obtain the pressure time spectrum.
[0082] The above process accurately captures infection characteristics through multi-dimensional signal processing. A circumferential sensor layout supports spatial localization of vascular wall pressure distribution, providing a basis for mapping the anatomical location of infection foci. Spectrum segmentation and wavelet decomposition effectively remove physiological noise, highlighting infection-related low-frequency pathological fluctuations. Kalman filtering further eliminates sensor accuracy errors and environmental interference, ensuring the reliability of pathological components. The time-frequency spectrum dynamically presents the frequency domain feature evolution of the pressure signal through two-dimensional time-frequency analysis. This processing provides pressure feature inputs with both spatiotemporal resolution and pathological specificity for multimodal blood infection monitoring. It can effectively help distinguish between infectious and non-infectious pressure abnormalities, improve the detection sensitivity of hemodynamic changes caused by early infection, and provide a high-quality frequency domain feature foundation for subsequent domain adversarial neural network modeling and infection foci localization.
[0083] Methods for obtaining the temperature correction matrix include:
[0084] Using temperature sensors circumferentially distributed on the surface of the dialysis catheter, the sensor measurements at each spatial location are modeled against the actual physical quantities based on the original sensitivity coefficient, temperature drift deviation, and random noise of the temperature sensors to obtain the spatial temperature deviation.
[0085] Under the premise of preset constant physical quantities, measurements at different temperatures are collected at each spatial location, and a dataset is established for each spatial location;
[0086] For each spatial location, the temperature drift deviation is set to an L-order polynomial. The spatial temperature deviation is replaced, and the replaced spatial temperature deviation is solved by the least squares method to obtain the calibration spatial temperature deviation value and the calibration sensitivity coefficient.
[0087] The temperature drift correction matrix is obtained by subtracting the temperature drift estimate from the calibration space temperature deviation value.
[0088] The sensitivity matrix is obtained by calculating the calibration sensitivity coefficient and the normalized sensitivity, and the temperature correction matrix is obtained by calculating the sensitivity matrix and the temperature drift correction matrix.
[0089] The core role of the above process in blood multimodal infection monitoring is reflected in the following aspects: Fine calibration eliminates sensor errors, ensuring consistency between temperature measurements and actual physical quantities, providing a reliable data foundation for extracting infection-related temperature features; the circumferential sensor layout supports improved spatial resolution of temperature distribution, enabling the system to capture abnormal temperature gradients at different sites on the dialysis catheter; and the combination of corrected temperature data with pressure-time spectrum allows for accurate calculation of temperature-pressure cross-features, avoiding misjudgments of cross-modal feature associations caused by sensor errors, significantly improving the sensitivity and positioning accuracy of the infection monitoring system for early inflammatory responses, and providing reliable temperature dimension support for multimodal data fusion analysis and dynamic threshold adjustment.
[0090] Feature extraction is performed on the pressure-time spectrum and temperature correction matrix to obtain analytical features, including pressure features, temperature features, and cross features.
[0091] By extracting features from the pressure-time spectrum and temperature correction matrix to generate pressure features, temperature features, and cross features, multi-dimensional pathological information fusion can be achieved. Pressure features capture vasodynamic abnormalities caused by infection, while temperature features reflect local thermo-metabolic disorders. The two provide infection early warning signals from the perspectives of hemodynamics and thermo-metabolic metabolism, respectively. Cross features reveal pathological mechanisms that are difficult to detect by a single modality by mining the spatiotemporal correlation patterns between pressure and temperature. For example, infection is often accompanied by synergistic abnormalities of pressure and temperature, and cross features are indicators that quantify this synergy. In the early stages of septic shock, microcirculatory disturbances lead to a strong correlation between low-frequency pressure fluctuations and local inflammatory heat production, while a single feature may be caused by non-infectious factors, easily leading to false alarms. Temperature response during infection may lag behind pressure fluctuations, while texture abnormalities in GAF images can be identified by reinforcement learning models. This correlation pattern cannot be captured by a single pressure or temperature feature. Without cross features, the system will be unable to distinguish between infectious synergistic abnormalities and non-infectious independent fluctuations, resulting in threshold adjustment losing its pathological mechanism basis. Cross-features, by constructing "pressure-temperature correlation features" rather than independent parameters, solve the problem of high misjudgment rates caused by individual baseline differences in traditional methods, which ultimately leads to limited accuracy in infection risk prediction. This provides infection monitoring with feature expressions and threshold decision-making basis that are closer to the physiological and pathological mechanisms. The organic integration of the three types of features provides high-quality, cross-modal, and multi-dimensional input for subsequent transfer learning models, enabling the system to identify infection signs from three perspectives: "abnormal blood flow, abnormal metabolism, and abnormal coordination." This significantly improves the sensitivity and specificity of early infection detection, while also providing individual feature basis for dynamic threshold adjustment and spatial correlation clues for locating infection foci, thus constructing a complete monitoring system covering pathological mechanism analysis, risk prediction, and precise diagnosis.
[0092] Methods for obtaining stress characteristics include:
[0093] Pressure characteristics include dynamic characteristics and frequency domain characteristics;
[0094] The time-frequency spectrum matrix is obtained from the pressure time-frequency spectrum. The time-frequency spectrum matrix is integrated over the respiratory frequency range to obtain the respiratory modulation intensity signal A(t). The respiratory cycle is segmented according to the peak value of A(t). The pulse pressure sequence is calculated in each respiratory cycle. The mean and standard deviation are calculated. The ratio of the standard deviation to the mean is calculated to obtain the statistical coefficient of variation.
[0095] After differentiating the pressure signal, SIFT is performed to obtain the time spectrum of the derivative. Energy abrupt change points are found in the preset high-frequency region to obtain the rising and falling edges of the contraction period. The corresponding intervals are truncated in the time domain to calculate the slope of the contraction period.
[0096] The kinetic characteristics are obtained by concatenating the statistical coefficient of variation and the slope during contraction.
[0097] Extract the power H1 and DC component power H0 at the fundamental frequency (i.e., the frequency corresponding to heart rate) of each time spectrum in the pressure time spectrum, calculate the ratio of H1 to H0, and obtain the proportion of the first harmonic energy.
[0098] For each time point in the pressure spectrum, the power spectral density is integrated within a preset frequency range to calculate the low-frequency power and obtain the low-frequency power integral.
[0099] The frequency domain characteristics are obtained by splicing together the proportion of first harmonic energy and the low-frequency band power integral.
[0100] By extracting kinetic and frequency domain features from the time-frequency spectrum of pressure, a pressure feature system covering short-term fluctuations and long-term frequency components was constructed. The statistical coefficient of variation, through regularity analysis of pulse pressure fluctuations within the respiratory cycle, sensitively captures circulatory stability abnormalities induced by infection. The systolic slope, detected by energy mutations in the derivative time-frequency spectrum, accurately locates changes in the rate of pressure rise during cardiac systole, reflecting abnormalities in myocardial contractility or vascular resistance caused by infection. The proportion of first harmonic energy quantifies heart rate-related pressure fluctuation components, identifying periodic abnormalities caused by autonomic nervous system dysregulation during infection. Low-frequency power integration focuses on the low-frequency energy distribution related to microcirculation, effectively capturing early vascular endothelial dysfunction in septic shock. These features, through joint time-frequency domain analysis, preserve the instantaneous kinetic characteristics of the pressure signal while uncovering the pathological significance of long-term frequency components. This provides key indicators for multimodal fusion models that reflect hemodynamic changes at different stages of infection, enabling early identification of pathological signals such as microcirculatory disturbances and circulatory instability caused by infection, improving the detection sensitivity of the monitoring system for early infection, and providing core evidence for infection severity assessment, dynamic threshold adjustment, and cross-modal feature correlation analysis based on the pressure dimension.
[0101] Methods for obtaining temperature characteristics include:
[0102] Temperature characteristics include spatial gradient characteristics and dynamic response characteristics;
[0103] The temperatures at the puncture point and the tip of the dialysis catheter, as well as the physical distance between the two points, are obtained. The difference between the tip temperature and the puncture point temperature is calculated to obtain the temperature difference. The temperature gradient is calculated based on the physical distance and the temperature difference, and is used as a spatial gradient feature.
[0104] The peak temperature sequence within a preset time period is obtained, and the first derivative is obtained by the central difference method to obtain the time derivative sequence and the temperature rise rate.
[0105] The pressure signal is low-pass filtered to obtain a denoised signal, and the lowest pressure point, i.e. the trough moment, is located in each cardiac cycle by using the threshold method.
[0106] Extract the time window containing a single pressure trough, extract the corresponding temperature rise rate and pressure signal segment, calculate the cross-correlation function of the temperature rise rate and pressure signal segment, obtain the time delay corresponding to the peak position, and calculate the phase difference of the pressure trough based on the time delay.
[0107] The phase difference between the rate of temperature rise and the pressure trough is spliced together to form a dynamic response feature.
[0108] By extracting spatial gradient and dynamic response features from temperature characteristics, a temperature-dimensional analysis method covering spatial distribution and dynamic coupling can be provided for blood multimodal infection monitoring. The temperature gradient, calculated using the temperature difference and physical distance between the puncture point and the tip, accurately captures local thermometabolic abnormalities caused by infection, providing direct evidence for spatial localization of infection foci. The rate of temperature rise, calculated using first-order differentials, sensitively identifies dynamic changes in body temperature in the early stages of infection, improving the detection capability for occult infections. The pressure trough phase difference, through cross-correlation analysis of temperature and pressure signals, reveals the time delay relationship between temperature changes and the cardiac cycle, reflecting the interference of infection on the cardiovascular-metabolic coupling mechanism. These features preserve the spatial heterogeneity of temperature signals while analyzing their dynamic correlation with pressure signals, providing key indicators for multimodal fusion models that reflect the spatial localization of infection pathology and abnormal physiological mechanisms. This effectively assists in early infection screening, precise localization of infection foci, and cross-modal pathological mechanism analysis, enhancing the monitoring system's multidimensional perception capability of infection from a temperature perspective.
[0109] Methods for obtaining cross features include:
[0110] The pressure and temperature signals are denoised and normalized to obtain standard pressure and temperature signals.
[0111] The standard pressure signal value range is divided into N1 intervals, and the standard temperature signal value range is divided into N2 intervals. Each interval contains an equal number of data points. The standard pressure signal and standard temperature signal are mapped to the corresponding intervals to obtain discrete pressure sequence and discrete temperature sequence.
[0112] The frequency of standard pressure signals and standard temperature signals falling into each interval combination (i,j) at the same time is statistically analyzed, where i∈[1,N1] and j∈[1,N2]. After normalization, the joint probability distribution is obtained, the marginal probability is calculated, and the mutual information entropy is obtained based on the joint probability distribution and the marginal probability.
[0113] The standard pressure signal and standard temperature signal are converted into an angle sequence using the inverse cosine function;
[0114] The angle sequence is converted into a two-dimensional matrix of length M using trigonometric identities, thus obtaining the GAF matrix of the pressure signal and the GAF matrix of the temperature signal.
[0115] The GAF matrix of the pressure signal and the GAF matrix of the temperature signal are used as the first two channels of the RGB channel. The third channel can be filled with 0 or the mean value to form an M×M×2 input image and obtain a two-dimensional spatiotemporal correlation image.
[0116] By calculating mutual information entropy and constructing a two-dimensional spatiotemporal correlation image to obtain cross-features, the deep analysis capability of blood multimodal infection monitoring in pressure-temperature cross-modal correlation is improved. Mutual information entropy captures the synergistic anomalies of pressure and temperature signals during infection by quantifying the joint probability distribution of the two signals, effectively distinguishing non-infectious abnormal interference. Gram angle field (GAF) converts one-dimensional time-series signals into two-dimensional images, and retains the amplitude and phase information of the signals through triangular identity mapping, so that the spatiotemporal correlation pattern of pressure and temperature can be automatically extracted by convolutional neural networks. Combined with the spatiotemporal correlation image formed by RGB channel fusion, the ability of deep learning models to identify complex pathological patterns is further improved. Such cross-features not only break through the information limitations of a single modality, but also reveal the synergistic disorder mechanism of hemodynamics and thermometabolism in infection pathology through nonlinear correlation analysis. This can significantly improve the detection sensitivity of the monitoring system for early infection, and provide key support for infection focus localization, cross-domain model generalization, and dynamic threshold optimization, thus constructing a multimodal collaborative intelligent infection monitoring system.
[0117] By acquiring synthetic and clinical data and combining them with transfer learning, a domain adversarial neural network is built to intelligently analyze real-time features and obtain predicted values of patient infection risk probability and inflammatory factor concentration ranges. The analyzed features include pressure features, temperature features, and cross features.
[0118] Methods for obtaining the probability of infection risk and the predicted range of inflammatory factor concentrations in patients include:
[0119] Generative analysis features of the source domain are generated through a preset external pulsating flow system. The preset external pulsating flow system is built by a neural network and combined with data from the user database for analysis and fitting to generate the generative analysis features of the source domain.
[0120] Real monitoring data from R patients were collected for feature extraction to obtain analytical features of the target domain; the monitoring data included pressure data and temperature data.
[0121] Spatiotemporal perturbations are applied to the source domain data, such as adding Gaussian noise, time series truncation, or spatial location offset, and interpolation is performed on the target domain data to complete it, such as linear interpolation, spline interpolation, or prediction interpolation based on adjacent features for missing data, and cross-domain sample pairs are generated through mixed precision training.
[0122] A shared feature extractor is built using 3D ResNet-50. The analyzed features are used as input to the shared feature extractor to obtain high-dimensional feature vectors.
[0123] Methods for obtaining high-dimensional feature vectors of the source domain include:
[0124] Different enhancement results of the same analytical feature in the source domain data are taken as positive samples, and the projected features of any enhancement results of other analytical features are taken as negative samples. The data are divided according to the batch size B. In each batch, one positive sample corresponds to (B-1)×2 negative samples, forming a mixed batch.
[0125] A normalized temperature-scale cross-entropy loss is constructed based on the cosine similarity of positive sample pairs and the temperature hyperparameter (taken as 0.1-0.5);
[0126] Samples are extracted from the mixed batch as input to the encoder to obtain high-dimensional feature vectors of the source domain.
[0127] A domain adversarial neural network was constructed, and source domain data, target domain data, and high-dimensional feature vectors were used as inputs to obtain predicted values of patient infection risk probability and inflammatory factor concentration ranges.
[0128] Training methods for domain adversarial neural networks include:
[0129] Construct a discriminator with dual task headers, including a domain classification header and an auxiliary task header;
[0130] The domain classification head outputs the domain discrimination probability, which is used to determine whether a high-dimensional feature vector comes from the source domain or the target domain;
[0131] The auxiliary task header is used to predict infection labels from source domain data;
[0132] A gradient inversion layer is inserted between the shared feature extractor and the discriminator. During training, gradient inversion is applied to the domain classification loss so that the shared feature extractor learns domain-invariant features, namely the infection features common to synthetic and clinical data.
[0133] The overall task loss function is designed to update the discriminator and shared feature extractor until a preset optimal performance condition is reached. The domain adversarial neural network corresponding to the optimal performance condition is then used as the final output domain adversarial neural network. The overall task loss function is calculated from the task loss, domain adversarial loss, and auxiliary loss. The task loss is calculated as the total task loss of the target domain and the source domain. The task loss is calculated by quantifying the deviation between the predicted probability and the true label using cross-entropy loss, and then adding the mean squared error loss that quantifies the deviation between the predicted interval and the true value. The domain adversarial loss is calculated by quantifying the error of domain classification using binary cross-entropy loss. The auxiliary loss is calculated by quantifying the error of infection prediction within the source domain using cross-entropy loss.
[0134] By combining synthetic and clinical data and employing transfer learning to construct a domain adversarial neural network, the cross-domain generalization ability of blood multimodal infection monitoring can be improved. Pre-training the model using synthetic data addresses the scarcity of clinically labeled data. By simulating the stress and temperature spatiotemporal patterns of infection scenarios, the model learns pathological feature associations in advance. Domain adversarial training forces the shared feature extractor to learn cross-domain invariant features through a gradient inversion layer, eliminating distribution differences between the source and target domains and ensuring the model accurately identifies infection-related features in real clinical environments. After fine-tuning with clinical data, the output infection risk probability and inflammatory factor concentration range prediction values retain the prior pathological patterns of the synthetic data while adapting to the physiological characteristics of the target patient group. This allows for early warning of infection risks and improves generalization accuracy across different hospitals and devices. This method effectively alleviates the dependence of medical AI on massive amounts of labeled data, achieving a leap from data simulation learning to accurate clinical prediction. It provides an intelligent analysis tool that is both efficient and reliable for early infection diagnosis, and is particularly suitable for rapid deployment and accurate decision-making in rare infection cases or new equipment scenarios.
[0135] The system acquires the patient's baseline data and, based on the patient's infection risk probability and the predicted values of inflammatory factor concentration ranges, dynamically adjusts the pressure fluctuation tolerance bandwidth and temperature rise slope threshold using reinforcement learning to generate personalized alarm commands.
[0136] Methods for dynamically adjusting the pressure fluctuation tolerance bandwidth and temperature rise slope threshold to generate personalized alarm commands include:
[0137] Define a state space, which includes a state s. The state s consists of real-time features, historical data, and patient baseline. Real-time features include pressure features, temperature features, and cross features. Historical data includes the number of alarms, true positives, false positives, and false negatives within a historical time period. Patient baseline includes patient age, baseline blood pressure, and baseline body temperature.
[0138] Define an action space, which includes action vectors. Each action vector consists of a pressure fluctuation tolerance bandwidth and a temperature rise slope threshold. The pressure fluctuation tolerance bandwidth is set to an E1 level adjustment based on the current patient baseline; the temperature rise slope threshold is set to an E2 level adjustment based on the current patient baseline. For example, the pressure threshold characteristic can include 7 adjustment values: -10, -5, -2, 0, +2, +5, and +10; the temperature threshold adjustment can include 7 adjustment values: -0.2, -0.1, -0.05, 0, +0.05, +0.1, and +0.2. Therefore, a total of 7 × 7 = 49 combinations can be formed.
[0139] The threshold adjustment is triggered once every preset time interval, and the action is selected according to the current status to update the alarm threshold;
[0140] A reward function is defined, which is obtained by weighting accuracy reward, stability reward and timeliness reward. Among them, accuracy reward is obtained by weighting the number of true positives, false positives and false negatives with preset true positive reward, false positive penalty and false negative penalty; stability reward is obtained by cosine similarity between current feature and feature of previous time step; timeliness reward is obtained by time difference from the occurrence of real infection to system alarm.
[0141] Based on real-time features, historical data, and the real-time features corresponding to historical data, a system is designed to encourage thresholds to remain stable for normal patterns. This system can be obtained by setting positive rewards for the number of true positives and negatives, and calculating the cosine similarity between real-time features and real-time features corresponding to historical data.
[0142] An Actor-Critic architecture is adopted, where state s serves as the input to the Actor network to obtain the action probability distribution; state s also serves as the input to the Critic network to obtain the state value; the Actor network uses log probability to measure the policy probability of choosing action a in state s, and uses an advantage function to evaluate the advantage of choosing action a in state s compared to the average case; the Critic network uses the sum of discounted estimates of the current reward and the future cumulative reward as the target of the state value, calculates the squared difference between the current state value estimate and the target, and measures the prediction error of the value network; the action with the highest state value is selected as the final action to be executed, and the stress fluctuation tolerance bandwidth and temperature rise slope threshold are obtained.
[0143] The above method integrates baseline data such as patient age and baseline blood pressure / body temperature, and automatically generates exclusive thresholds for special groups such as patients with hypertension and patients with low fever, thereby reducing the false positive rate and false negative rate and solving the misjudgment problem caused by the traditional fixed threshold "one-size-fits-all" approach.
[0144] While baseline temperature and pressure vary significantly among patients, the coordinated abnormalities in temperature and pressure during infection share commonalities across individuals. Cross-features, through standardized correlation metrics, eliminate the interference of individual baseline differences, allowing threshold adjustments to focus on pathological correlations rather than absolute values. For example, a patient with a baseline pressure of 150 mmHg might have a pressure fluctuation of 10 mmHg within the normal range, but if this is accompanied by a 0.3°C increase in temperature and an increase in mutual information entropy, it suggests infection risk, requiring a tightening of the pressure bandwidth threshold. The temperature and pressure correlation patterns differ at different stages of infection; cross-features can reflect pathological progression in real time, guiding dynamic threshold adjustments as the disease progresses, whereas a single feature cannot achieve this stage-specific adaptation.
[0145] By incorporating cross-features, the state space can characterize the dynamic coupling of pressure and temperature, enabling the model to effectively learn the key patterns of coordinated changes in temperature and pressure characteristics during infection. Adjusting the pressure fluctuation tolerance bandwidth and temperature rise slope threshold requires comprehensive consideration of the correlation strength between features. For example, when mutual information entropy is high, the system should increase its sensitivity to synchronous anomalies in temperature and pressure; if mutual information entropy is low, it may be judged as independent noise. Without cross-features, the model cannot achieve this dynamic weight adjustment, and threshold optimization may deviate from clinical realities. Cross-features upgrade "pressure-temperature" from independent variables to correlated variables, giving threshold adjustment the ability to perceive pathological mechanisms. For example, when cross-features show that the mutual information entropy of temperature and pressure is 1.5 times higher than the normal mean, the reinforcement learning model can automatically trigger a coordinated abnormal response mode, reducing the pressure bandwidth threshold by 20% and the temperature rise slope threshold by 0.1℃ / h. This decision-making logic is impossible for a single-modal threshold system. Cross-features, through a paradigm shift from unimodal independent analysis to multimodal correlation modeling, provide a more pathologically accurate feature representation for infection monitoring by incorporating the dynamic interaction patterns of pressure and temperature. Threshold effectiveness is evaluated in real-time using historical alarm data, with threshold optimization triggered at preset intervals. This effectively addresses changes in equipment noise and fluctuations in patient condition, ensuring the stability of the monitoring system in different scenarios such as ICUs and general wards, and improving alarm accuracy. The cosine similarity constraint introduced into the reward function ensures that the threshold remains stable to normal physiological fluctuations while remaining sensitive to infection-related abnormal features, forming a balanced mechanism of "stable baseline - anomaly detection." Combining infection risk probability and inflammatory factor prediction values, the system can dynamically generate tiered alarm instructions to assist doctors in quickly locating the infection stage and shortening diagnosis time. This method, through the "state-action-reward" closed loop of reinforcement learning, transforms clinical experience into a computable optimization objective, achieving a leap from experience-driven threshold setting to data-driven intelligent decision-making. This significantly improves the clinical practicality and reliability of infection monitoring systems, especially suitable for the management of critically ill patients with complex conditions and significant individual differences.
[0146] Example 2:
[0147] Please see Figure 3 As shown, this embodiment, based on embodiment 1, further provides a method for locating suspected foci of infection, including the following steps:
[0148] During the feature extraction stage, the feature weight of each spatial location is recorded;
[0149] Calculate the deviation between the real-time feature value and the patient's baseline value for each spatial location;
[0150] An abnormal threshold is set based on the patient's baseline values, and abnormal sites are detected and marked.
[0151] Establish a mapping relationship between catheter sites and vascular segments, and obtain suspected infection foci based on the location of abnormal sites.
[0152] Example 3:
[0153] Please see Figure 4 As shown, this embodiment provides a blood multimodal infection monitoring system, including:
[0154] Acquisition and Analysis Module: The dialysis catheter with integrated intelligent sensing elements in the interventional blood vessel intelligently acquires and analyzes pressure and temperature data in real time to obtain pressure time spectrum and temperature correction matrix.
[0155] Feature extraction module: Extracts features from the pressure-time spectrum and temperature correction matrix to obtain analytical features, including pressure features, temperature features, and cross features;
[0156] Monitoring and Analysis Module: Acquires synthetic and clinical data, combines transfer learning to build a domain adversarial neural network, performs intelligent analysis on real-time features, and obtains predicted values of patient infection risk probability and inflammatory factor concentration range.
[0157] Dynamic alarm module: Obtain patient baseline, predict patient infection risk probability and inflammatory factor concentration range based on reinforcement learning, dynamically adjust pressure fluctuation tolerance bandwidth and temperature rise slope threshold, and generate personalized alarm commands.
[0158] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
[0159] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A blood-based multimodal infection surveillance method, characterized in that, Includes the following steps: The dialysis catheter, which integrates intelligent sensing elements into the interventional blood vessel, intelligently collects and analyzes pressure and temperature data in real time to obtain the pressure time spectrum and temperature correction matrix. Feature extraction is performed on the pressure-time spectrum and temperature correction matrix to obtain analytical features, including pressure features, temperature features, and cross features. By acquiring synthetic and clinical data and combining them with transfer learning, a domain adversarial neural network is built to intelligently analyze real-time features and obtain predicted values of patient infection risk probability and inflammatory factor concentration ranges. Methods for obtaining the probability of infection risk and the predicted range of inflammatory factor concentrations in patients include: Generative analysis features of the source domain are generated by pre-setting an external pulsating flow system; Collect monitoring data and extract features to obtain analytical features of the target domain; The source and target domain data are preprocessed, and cross-domain sample pairs are generated through mixed-precision training. A shared feature extractor is built using 3D ResNet-50. The analyzed features are used as input to the shared feature extractor to obtain high-dimensional feature vectors. A domain adversarial neural network was constructed, and source domain data, target domain data, and high-dimensional feature vectors were used as inputs to obtain predicted values of patient infection risk probability and inflammatory factor concentration ranges. Methods for obtaining high-dimensional feature vectors of the source domain include: Different enhancement results of the same analytical feature in the source domain data are used as positive samples, and the projected features of any enhancement results of other analytical features are used as negative samples, forming a mixed batch according to a preset batch size and preset ratio. A normalized temperature-scale cross-entropy loss is constructed based on the cosine similarity of positive sample pairs and the temperature hyperparameter. Samples are drawn from the mixed batch as input to the encoder to obtain a high-dimensional feature vector of the source domain; The system acquires the patient's baseline data and, based on the patient's infection risk probability and the predicted values of inflammatory factor concentration ranges, dynamically adjusts the pressure fluctuation tolerance bandwidth and temperature rise slope threshold using reinforcement learning to generate personalized alarm commands.
2. The blood multimodal infection monitoring method according to claim 1, characterized in that, Methods for obtaining the spectrum of pressure include: The original pressure signal is subjected to discrete Fourier transform to obtain the pressure signal, and the pressure signal is divided into physiological fluctuation frequency band and pathological low frequency band by spectrum segmentation. Scaling functions and wavelet functions are constructed, and pressure signals are combined to decompose the physiological fluctuation frequency band and the pathological low frequency band to obtain physiological and pathological components. The baseline value is defined as the state vector, the pathological component is defined as the observation value, and the observation noise is defined in combination with the sensor accuracy. The current pathological component is predicted based on the state vector of the previous moment by the Kalman filter algorithm. The prediction is then updated by combining the observation value and the observation noise to obtain the corrected pathological component. The physiological component and the corrected pathological component are superimposed in the time domain to obtain the reconstructed signal. The reconstructed signal is windowed and framed, and then a discrete Fourier transform is performed on each frame of the reconstructed signal to obtain the pressure time spectrum.
3. The blood multimodal infection monitoring method according to claim 2, characterized in that, Methods for obtaining the temperature correction matrix include: Based on the original sensitivity coefficient, temperature drift deviation, and random noise of the temperature sensor, the sensor measurement value at each spatial location is modeled against the actual physical quantity to obtain the spatial temperature deviation. Under the premise of preset constant physical quantities, measurements at different temperatures are collected at each spatial location, and a dataset is established for each spatial location; For each spatial location, the temperature drift deviation is set to an L-order polynomial. The spatial temperature deviation is replaced, and the replaced spatial temperature deviation is solved by the least squares method to obtain the calibration spatial temperature deviation value and the calibration sensitivity coefficient. The temperature drift correction matrix is obtained by subtracting the temperature drift estimate from the calibration space temperature deviation value. The sensitivity matrix is obtained by calculating the calibration sensitivity coefficient and the normalized sensitivity, and the temperature correction matrix is obtained by calculating the sensitivity matrix and the temperature drift correction matrix.
4. The blood multimodal infection monitoring method according to claim 3, characterized in that, Methods for obtaining stress characteristics include: Pressure characteristics include dynamic characteristics and frequency domain characteristics; The time-frequency spectrum matrix is obtained from the pressure time-frequency spectrum. The time-frequency spectrum matrix is integrated over the respiratory frequency range to obtain the respiratory modulation intensity signal A(t). The respiratory cycle is segmented according to the peak value of A(t). The mean and standard deviation of the pulse pressure sequence in each respiratory cycle are statistically analyzed, and the statistical coefficient of variation is calculated. After differentiating the pressure signal, SIFT is performed to obtain the time spectrum of the derivative. Energy abrupt change points are found in the preset high-frequency region to obtain the rising and falling edges of the contraction period. The corresponding intervals are truncated in the time domain to calculate the slope of the contraction period. The kinetic characteristics are obtained by concatenating the statistical coefficient of variation and the slope during contraction. Extract the power H1 at the fundamental frequency and the power H0 of the DC component of each time spectrum in the pressure time spectrum, and calculate the proportion of the first harmonic energy; For each time point in the pressure spectrum, the power spectral density is integrated within a preset frequency range to calculate the low-frequency power and obtain the low-frequency power integral. The frequency domain characteristics are obtained by splicing together the proportion of first harmonic energy and the low-frequency band power integral.
5. The blood multimodal infection monitoring method according to claim 4, characterized in that, Cross features include mutual information entropy and two-dimensional spatiotemporal correlation images. Methods for obtaining cross features include: The pressure and temperature signals are denoised and normalized to obtain standard pressure and temperature signals. The standard pressure signal value range is divided into N1 intervals, and the standard temperature signal value range is divided into N2 intervals. Each interval contains an equal number of data points. The standard pressure signal and standard temperature signal are mapped to the corresponding intervals to obtain discrete pressure sequence and discrete temperature sequence. The frequency at which the statistical standard pressure signal and the standard temperature signal fall into each interval combination (i,j) at the same time is calculated. , After normalization, the joint probability distribution is obtained, the marginal probability is calculated, and the mutual information entropy is obtained based on the joint probability distribution and the marginal probability. The standard pressure signal and standard temperature signal are converted into an angle sequence using the inverse cosine function; The angle sequence is converted into a two-dimensional matrix of length M using trigonometric identities, thus obtaining the GAF matrix of the pressure signal and the GAF matrix of the temperature signal. The GAF matrix of the pressure signal and the GAF matrix of the temperature signal are used as the first two channels of the RGB channel to form an M×M×2 input image, thus obtaining a two-dimensional spatiotemporal correlated image.
6. The blood multimodal infection monitoring method according to claim 1, characterized in that, Training methods for domain adversarial neural networks include: Construct a discriminator with dual task headers, including a domain classification header and an auxiliary task header; The domain classification head outputs the domain discrimination probability, which is used to determine whether a high-dimensional feature vector comes from the source domain or the target domain; The auxiliary task header is used to predict infection labels from source domain data; A gradient inversion layer is inserted between the shared feature extractor and the discriminator. During training, gradient inversion is applied to the domain classification loss so that the shared feature extractor learns domain-invariant features. The overall task loss function is designed to update the discriminator and shared feature extractor until a preset optimal performance condition is reached. The domain adversarial neural network corresponding to the optimal performance condition is then used as the final output domain adversarial neural network. The overall task loss function is calculated from the task loss, domain adversarial loss, and auxiliary loss. The task loss is calculated as the total task loss of the target domain and the source domain. The task loss is calculated by quantifying the deviation between the predicted probability and the true label using cross-entropy loss, and then adding the mean squared error loss that quantifies the deviation between the predicted interval and the true value. The domain adversarial loss is calculated by quantifying the error of domain classification using binary cross-entropy loss. The auxiliary loss is calculated by quantifying the error of infection prediction within the source domain using cross-entropy loss.
7. The blood multimodal infection monitoring method according to claim 6, characterized in that, Methods for generating personalized alarm commands include: Define a state space, which includes a state s. The state s consists of real-time features, historical data, and patient baseline. Real-time features include pressure features, temperature features, and cross features. Historical data includes the number of alarms, true positives, false positives, and false negatives within a historical time period. Define the action space, which includes action vectors. The action vectors consist of the pressure fluctuation tolerance bandwidth and the temperature rise slope threshold. The pressure fluctuation tolerance bandwidth is set to E1 level based on the current patient baseline. The temperature rise slope threshold is set to E2 level based on the current patient baseline. The threshold adjustment is triggered once every preset time interval, and the action is selected according to the current state s to update the alarm threshold; A reward function is defined, which is obtained by weighting accuracy reward, stability reward and timeliness reward. Among them, accuracy reward is obtained by weighting the number of true positives, false positives and false negatives with preset true positive reward, false positive penalty and false negative penalty; stability reward is obtained by cosine similarity between current feature and feature of previous time step; timeliness reward is obtained by time difference from the occurrence of real infection to system alarm. An Actor-Critic architecture is adopted, where state s serves as the input to the Actor network to obtain the action probability distribution; state s also serves as the input to the Critic network to obtain the state value; the Actor network uses log probability to measure the policy probability of choosing action a in state s, and uses an advantage function to evaluate the advantage of choosing action a in state s compared to the average case; the Critic network uses the sum of discounted estimates of the current reward and the future cumulative reward as the target of the state value, calculates the squared difference between the current state value estimate and the target, and measures the prediction error of the value network; the action with the highest state value is selected as the final action to be executed, and the stress fluctuation tolerance bandwidth and temperature rise slope threshold are obtained.
8. A blood multimodal infection monitoring system, implementing the blood multimodal infection monitoring method according to any one of claims 1-7, characterized in that, include: Acquisition and Analysis Module: The dialysis catheter with integrated intelligent sensing elements in the interventional blood vessel intelligently acquires and analyzes pressure and temperature data in real time to obtain pressure time spectrum and temperature correction matrix. Feature extraction module: Extracts features from the pressure-time spectrum and temperature correction matrix to obtain analytical features, including pressure features, temperature features, and cross features; Monitoring and Analysis Module: Acquires synthetic and clinical data, combines transfer learning to build a domain adversarial neural network, performs intelligent analysis on real-time features, and obtains predicted values of patient infection risk probability and inflammatory factor concentration range. Methods for obtaining the probability of infection risk and the predicted range of inflammatory factor concentrations in patients include: Generative analysis features of the source domain are generated by pre-setting an external pulsating flow system; Collect monitoring data and extract features to obtain analytical features of the target domain; The source and target domain data are preprocessed, and cross-domain sample pairs are generated through mixed-precision training. A shared feature extractor is built using 3D ResNet-50. The analyzed features are used as input to the shared feature extractor to obtain high-dimensional feature vectors. A domain adversarial neural network was constructed, and source domain data, target domain data, and high-dimensional feature vectors were used as inputs to obtain predicted values of patient infection risk probability and inflammatory factor concentration ranges. Methods for obtaining high-dimensional feature vectors of the source domain include: Different enhancement results of the same analytical feature in the source domain data are used as positive samples, and the projected features of any enhancement results of other analytical features are used as negative samples, forming a mixed batch according to a preset batch size and preset ratio. A normalized temperature-scale cross-entropy loss is constructed based on the cosine similarity of positive sample pairs and the temperature hyperparameter. Samples are drawn from the mixed batch as input to the encoder to obtain a high-dimensional feature vector of the source domain; Dynamic alarm module: Obtain patient baseline, predict patient infection risk probability and inflammatory factor concentration range, dynamically adjust pressure fluctuation tolerance bandwidth and temperature rise slope threshold based on reinforcement learning, and generate personalized alarm commands.
9. A dialysis catheter, characterized in that, Applied to the blood multimodal infection monitoring method as described in any one of claims 1-7.