Intelligent monitoring system and method for hypotension of dialysis patients based on multi-modal signals

Abstract

The application discloses a dialysis patient hypotension intelligent monitoring system and method based on a multi-modal signal, relates to the technical field of medical physiological monitoring, and comprises a multi-modal data acquisition module, a data preprocessing module, a multi-modal feature fusion module and a hypotension identification and early warning module. The system synchronously acquires physiological signals and dialysis equipment operation parameters of a patient, determines the operation parameters as excitation modal data and the physiological signals as response modal data, extracts excitation-response coupling features representing excitation driving response, and fuses the features with single-modal features to obtain cross-modal fusion features, according to which, a hypotension risk grade is identified and graded early warning is performed. The application utilizes the time sequence coupling structure of ultrafiltration excitation and physiological response as features, identifies circulation compensation reserve failure in advance from the dynamic system identification level, solves the problem that the prior art symmetrically processes each modal and the monitoring lags behind blood pressure drop, and improves the safety of hemodialysis treatment.

Description

Technical Field

[0001] This invention relates to the field of medical physiological monitoring technology, specifically to an intelligent monitoring system and method for hypotension in dialysis patients based on multimodal signals. Background Technology

[0002] Hypotension is one of the most common and serious complications in maintenance hemodialysis patients. During dialysis, the dialysis machine continuously removes excess water from the patient's blood vessels through ultrafiltration. When the rate of decrease in intravascular volume caused by ultrafiltration exceeds the rate of refilling of the blood vessels by interstitial fluid, and the patient's circulatory system's compensatory mechanisms are insufficient to maintain effective circulating blood volume, hypotension occurs during dialysis. Hypotension during dialysis not only causes discomfort for patients during dialysis, but recurrent episodes are also closely related to myocardial damage, insufficient cerebral perfusion, vascular access failure, and even increased mortality. Therefore, accurate and early monitoring and warning of hypotension during dialysis is of significant clinical value in ensuring the safety of hemodialysis treatment.

[0003] Existing dialysis hypotension monitoring and early warning technologies mainly follow two technical routes. The first route is based on threshold determination of a single physiological parameter. Chinese invention patent CN102686252B discloses a device and method for predicting rapid symptomatic hypotension. This method sets a pressure sensor in the extracorporeal blood flow circuit coupled to the subject's cardiovascular system, registers pulse shape parameters related to the subject's pulse generator, and examines whether the pulse amplitude measurement or its statistical deviation measurement calculated from the pulse shape parameters meets the judgment criteria during the measurement period. If it does, an output signal indicating rapid symptomatic hypotension is generated. However, this approach relies solely on the single modal signal of pressure and treats the operation of the dialysis device as an external condition unrelated to physiological response. It does not utilize the intrinsic driving relationship between dialysis ultrafiltration and physiological response. Therefore, it can only provide an indication after blood pressure has already shown an observable downward trend, making it difficult to identify risks in earlier stages before circulatory compensation capacity is exhausted and blood pressure has substantially decreased.

[0004] The second approach is data-driven prediction based on machine learning. US Patent No. US12040092B2 discloses a real-time intradialytic hypotension prediction system. This system segments historical hemodialysis treatment data into positive, negative, and non-hypotension categories based on their temporal proximity to hypotension events, creating a machine learning training dataset. Based on this dataset, a machine learning model is trained to predict the imminent occurrence of hypotension events. However, this approach indiscriminately feeds various physiological data and dialysis machine data as independent input features into the classification model, simply aligning each modality on the time axis and processing them side-by-side. Essentially, it still involves symmetrical fusion of modalities. This approach erases the temporal coupling structure between the dialysis machine's ultrafiltration stimulation and the patient's physiological response during the alignment process. This means the model can only capture patterns from the statistical changes of each signal, failing to characterize the dynamic characteristics of the patient's circulatory system's response to ultrafiltration stimulation. Therefore, its prediction still lags behind reflecting the intrinsic process of compensatory capacity decline.

[0005] In summary, both single-parameter threshold approaches and data-driven approaches share a common, deep-seated technical bottleneck: existing solutions treat dialysis machine operating parameters and patient physiological signals as equals, applying symmetrical fusion or independent processing. They treat the temporal misalignment between ultrafiltration stimulation and physiological response as a synchronization error that needs to be eliminated through timestamp alignment. This erases the crucial information carried by the temporal coupling structure of the dialysis machine's stimulation driving the physiological response—precisely because this temporal coupling structure encodes the patient's circulatory system's volume compensation capacity for ultrafiltration dehydration. Since the decline in compensation capacity precedes the actual decrease in blood pressure, existing solutions, lacking this temporal coupling structure, can only respond passively after blood pressure has already dropped, failing to provide true early warning in the early stages of compensatory reserve depletion. How to identify the depletion of circulatory compensatory reserve in advance from the temporal coupling relationship between dialysis machine ultrafiltration stimulation and the patient's multimodal physiological response, under the constraints of real-time bedside monitoring in the hemodialysis unit, remains a core technical problem that current technologies have not yet solved. Summary of the Invention

[0006] Addressing the core technical bottlenecks in existing technologies for monitoring hypotension in dialysis patients—namely, symmetrical processing of dialysis equipment operating parameters and patient physiological signals, erasing the temporal coupling structure between ultrafiltration stimulation and physiological response, thus failing to provide early warning in the early stages of compensatory reserve depletion—this invention provides an intelligent monitoring system and method for hypotension in dialysis patients based on multimodal signals. By determining the dialysis equipment operating parameters as the excitation mode and the patient physiological signals as the response mode, and extracting the excitation-response coupling features between the two for asymmetric cross-modal fusion, this invention achieves early identification and graded warning of hypotension in dialysis patients from the principle of dynamic system identification between dialysis stimulation and physiological response, without changing the existing dialysis equipment and physiological monitoring hardware.

[0007] The technical solution of this invention is as follows:

[0008] This invention provides an intelligent monitoring system for hypotension in dialysis patients based on multimodal signals, comprising: a multimodal data acquisition module for simultaneously acquiring patient physiological signals and dialysis equipment operating parameters, wherein the patient physiological signals include continuous blood pressure signals, electrocardiogram signals, and relative blood volume signals, and the dialysis equipment operating parameters include ultrafiltration rate and dialysate conductivity; a data preprocessing module for denoising and segmenting the patient physiological signals and the dialysis equipment operating parameters to obtain preprocessed signal segments; and a multimodal feature fusion module for determining the signal segments corresponding to the dialysis equipment operating parameters from the preprocessed signal segments as excitation modal data, and... The signal segments corresponding to the patient's physiological signals are determined as response modal data. Single-modal features are extracted from both the excitation modal data and the response modal data. Excitation-response coupling features between the excitation modal data and the response modal data are also extracted. The single-modal features and the excitation-response coupling features are fused to obtain cross-modal fusion features. These excitation-response coupling features characterize the temporal coupling relationship where changes in the excitation modal data drive changes in the response modal data. A hypotension identification and early warning module is used to identify the hypotension risk level of dialysis patients based on the cross-modal fusion features and output graded early warning signals according to the hypotension risk level.

[0009] This invention also provides a method for intelligent monitoring of hypotension in dialysis patients based on multimodal signals, comprising the following steps: synchronously acquiring patient physiological signals and dialysis equipment operating parameters of dialysis patients, wherein the patient physiological signals include continuous blood pressure signals, electrocardiogram signals, and relative blood volume signals, and the dialysis equipment operating parameters include ultrafiltration rate and dialysate conductivity; denoising and segmenting the patient physiological signals and the dialysis equipment operating parameters respectively to obtain preprocessed signal segments; determining the signal segments corresponding to the dialysis equipment operating parameters as excitation modal data and the signal segments corresponding to the patient physiological signals as response modal data from the preprocessed signal segments; extracting single-modal features from the excitation modal data and the response modal data respectively; extracting excitation-response coupling features between the excitation modal data and the response modal data; and fusing the single-modal features and the excitation-response coupling features to obtain cross-modal fusion features; identifying the hypotension risk level of dialysis patients according to the cross-modal fusion features, and outputting graded early warning signals according to the hypotension risk level.

[0010] The beneficial effects of this invention are as follows:

[0011] First, by defining dialysis equipment operating parameters as excitation modal data and patient physiological signals as response modal data, and extracting excitation-response coupling features characterizing changes in response modal data driven by excitation modal data, this invention enables early characterization of circulatory compensatory capacity decline before a substantial drop in blood pressure. The mechanism is as follows: the fundamental cause of hypotension during dialysis is that the rate of decrease in intravascular volume caused by ultrafiltration exceeds the rate of plasma refilling, resulting in insufficient circulatory compensation. Ultrafiltration is the cause, and physiological changes are the effect; a definite temporal driving relationship exists between the two. Existing technologies that symmetrically process dialysis equipment operating parameters and physiological signals treat the temporal misalignment corresponding to this driving relationship as an error and align it to eliminate it. This solution, however, extracts the temporal coupling structure itself as a feature, shifting the monitoring focus from whether blood pressure values ​​are already low to whether the circulatory system's response to ultrafiltration excitation is declining. Compared to providing an indication only after abnormal blood pressure pulse shape has appeared, this solution can provide risk information at an earlier stage, before compensatory capacity begins to decline and blood pressure has substantially decreased.

[0012] Second, by modeling the mapping from excitation modal data to response modal data as a time-varying response kernel with a response time constant and response gain, and recursively updating the kernel parameters during the dialysis process, this invention achieves continuous tracking of the dynamic changes in compensatory reserves as dehydration progresses. The mechanism is as follows: as dialysis progresses and the cumulative ultrafiltration volume increases, the patient's plasma refilling capacity and circulatory compensatory reserves are gradually depleted, causing the circulatory system's response time constant to ultrafiltration excitation to continuously lengthen and the response kernel to become continuously dulled—an inherently time-varying process. Using fixed-parameter models or uniformly modeling the entire dataset cannot characterize this time-varying characteristic, leading to the response characteristics of the later stages of dialysis being masked by the earlier data. This solution, however, identifies the time-varying response kernel recursively through time windows, ensuring that the description of the response characteristics always matches the current dehydration state. Compared to static classification models trained on entire historical datasets, this solution can distinguish the differences in response characteristics of the same patient at different stages of dialysis, thereby reducing false negatives and false negatives for patients in a critical compensatory state.

[0013] Third, through the synergy of excitation-response coupling feature extraction, time-varying response kernel identification, and response time constant decay rate inflection point detection, this invention generates a synergistic effect exceeding the sum of the individual effects of each feature. Specifically: the estimation of individual physiological response delays is unstable under continuous ultrafiltration modulation and baseline drift in a bedside environment; individual time-varying response kernels only describe the response characteristics at the current moment and lack foresight; and individual decay rates cannot be calculated without a reliable kernel parameter sequence. However, this scheme uses prescription adjustment events that actively adjust the ultrafiltration rate of dialysis prescriptions as known excitation markers to anchor the delay estimation, making the delay estimation stable under bedside perturbations, thereby supporting the reliable recursive identification of time-varying response kernels. Then, the inflection point where the response time constant decay rate changes from linear growth to nonlinear accelerated growth is detected on the kernel parameter sequence, that is, the moment when the nonlinear acceleration index obtained by the normalized second difference of the response time constant decay rate continuously exceeds the set inflection point threshold. This inflection point corresponds to the critical turning point of the cyclic compensatory reserve towards depletion, which often occurs before the actual decrease in blood pressure. The three elements are interconnected, enabling this scheme to achieve a forward-looking identification capability for the depletion of compensatory reserves that cannot be achieved by any single feature.

[0014] Fourth, by constructing physical constraint terms based on plasma refill kinetics and using them as regularization terms to constrain the mapping of cross-modal fusion features to hypotension risk levels, this invention embeds known physiological and physical laws into the data-driven model. The mechanism is that the decay rate of the response time constant and the cumulative ultrafiltration volume should satisfy a monotonically increasing relationship determined by the plasma refill mechanism. Purely data-driven mappings are prone to fitting spurious associations that violate this physical law when the sample size is limited. However, by introducing physical constraint regularization terms, the mapping is constrained within a solution space that conforms to the refill kinetics, transforming the model from a pure data black box into a gray box model constrained by physical laws, resulting in more stable recognition performance among different dialysis patients. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the structure of the intelligent monitoring system for hypotension in dialysis patients based on multimodal signals provided in an embodiment of the present invention;

[0016] Figure 2 This is a flowchart illustrating the intelligent monitoring method for hypotension in dialysis patients based on multimodal signals provided in this embodiment of the invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0018] like Figure 1 As shown, this embodiment provides an intelligent monitoring system for hypotension in dialysis patients based on multimodal signals. Deployed at the bedside in the hemodialysis room, it is connected to the dialysis equipment and the physiological signal acquisition front-end. The system includes a multimodal data acquisition module 1, a data preprocessing module 2, a multimodal feature fusion module 3, a hypotension identification and early warning module 4, a human-computer interaction terminal 5, and a storage module 6. The specific implementation methods of each module are described below.

[0019] The multimodal data acquisition module 1 is used to simultaneously acquire patient physiological signals and dialysis equipment operating parameters from dialysis patients. These physiological signals include continuous blood pressure signals, electrocardiogram (ECG) signals, and relative blood volume signals. Continuous blood pressure signals are acquired by a continuous non-invasive blood pressure cuff or finger-type continuous blood pressure sensor worn on the non-arteriovenous fistula side of the patient's upper limb, with a sampling frequency of 100Hz. ECG signals are acquired by a three-lead ECG electrode attached to the patient's chest, with a sampling frequency of 1000Hz. Relative blood volume signals are acquired by an optical blood volume sensor located at the arterial end of the dialysis extracorporeal circulation tubing, measuring transmittance, with a sampling frequency of 100Hz. The optical blood volume sensor emits near-infrared light into the blood flowing through the arterial end of the tubing using a light-emitting element, and receives the light intensity transmitted through the blood using a light-receiving element. Because the hemoglobin concentration in the blood relatively increases as the intravascular volume decreases, the transmittance decreases accordingly. Based on this, the optical blood volume sensor outputs a relative blood volume signal reflecting the change in relative intravascular blood volume.

[0020] The operating parameters of the dialysis equipment include ultrafiltration rate and dialysate conductivity. The multimodal data acquisition module 1 communicates with the control unit of the dialysis equipment via a data interface, reading the current ultrafiltration rate and dialysate conductivity in real time at 1-second sampling intervals. The ultrafiltration rate ranges from 0 L / h to 2.5 L / h, and the dialysate conductivity ranges from 13.0 mS / cm to 15.0 mS / cm. The multimodal data acquisition module 1 assigns a uniform timestamp to each of the acquired patient physiological signals and dialysis equipment operating parameters to ensure the correspondence of signals on the time axis during subsequent processing, but it does not eliminate the inherent time difference between excitation and response.

[0021] It should be noted that the patient physiological signals acquired by the multimodal data acquisition module 1 and the dialysis equipment operating parameters are fundamentally different in physical origin: the ultrafiltration rate in the dialysis equipment operating parameters is an external driving quantity actively applied by the dialysis equipment and controllable by the dialysis prescription, while the patient physiological signals are passive results of the patient's circulatory system responding to this external driving quantity. This difference is the physical basis for subsequently identifying the two as the excitation mode and response mode, respectively, in this system.

[0022] The data preprocessing module 2 is used to denoise and segment the patient's physiological signals and the dialysis equipment operating parameters to obtain preprocessed signal segments.

[0023] For denoising, bandpass filtering with a passband of 0.5Hz to 40Hz is used for ECG signals to remove baseline drift and power frequency interference; low-pass filtering with a cutoff frequency of 20Hz is used for continuous blood pressure signals to suppress high-frequency noise; for relative blood volume signals, since their effective changes are in the low-frequency band while optical measurement noise is in the high-frequency band, wavelet thresholding based on discrete wavelet transform is used to decompose the signal to the 5th level and reconstruct it after applying soft thresholding to each detail coefficient; for ultrafiltration rate and dialysate conductivity, since they are set quantities and slowly varying measurement quantities output by the dialysis equipment, only median filtering is used to remove isolated outliers caused by communication jitter. The bandpass filtering, low-pass filtering, wavelet thresholding denoising, and median filtering involved in the above denoising are all signal processing methods known in the art. The appropriate passband, cutoff frequency, and wavelet decomposition level can be selected according to the spectral characteristics of each signal, and their calculation details will not be elaborated further.

[0024] For segmentation, data preprocessing module 2 uses a sliding time window with a length of 120s and a step size of 30s between adjacent windows to divide the denoised signals into overlapping signal segments along the time axis. Each signal segment carries its window start and end timestamps, thus obtaining the preprocessed signal segments. The sliding time window length of 120s is chosen to ensure that each signal segment can completely cover the main time history of a typical ultrafiltration excitation and the resulting physiological response; the step size of 30s is chosen to control the redundant computational load between adjacent windows while ensuring timely risk updates.

[0025] The multimodal feature fusion module 3 is the core of this system. It is used to determine the excitation modal data and response modal data from the preprocessed signal segment, extract single-modal features and excitation-response coupling features, and fuse them to obtain cross-modal fusion features.

[0026] First, the multimodal feature fusion module 3 identifies the signal segments corresponding to the dialysis equipment's operating parameters as excitation modal data and the signal segments corresponding to the patient's physiological signals as response modal data from the preprocessed signal segments. As mentioned earlier, the ultrafiltration rate is an external driving force actively applied by the dialysis equipment, so the signal segments corresponding to the operating parameters are used as excitation modal data; the patient's physiological signals are the result of the circulatory system's response to this driving force, so the signal segments corresponding to the physiological signals are used as response modal data.

[0027] Secondly, the multimodal feature fusion module 3 extracts single-modal features from the excitation modal data and the response modal data respectively. For the excitation modal data, it extracts the mean, number of step changes, and amplitude of the ultrafiltration rate within the signal segment, as well as the mean and drift of the dialysate conductivity. For the response modal data, it extracts the systolic and diastolic blood pressure of the continuous blood pressure signal and its descent slope within the segment, the heart rate and heart rate variability index obtained after R-wave detection of the electrocardiogram signal, and the descent amplitude and descent slope of the relative blood volume signal within the segment. The extraction of the above single-modal features adopts conventional statistical feature calculation and time-frequency feature calculation methods in this field, mapping each signal segment to a corresponding single-modal feature vector.

[0028] Next, the multimodal feature fusion module 3 extracts the excitation-response coupling features between the excitation modal data and the response modal data. The excitation-response coupling features are used to characterize the temporal coupling relationship in which changes in the excitation modal data drive changes in the response modal data. The extraction includes physiological response delay estimation, compensation response time ratio calculation, time-varying response kernel identification, and response time constant decay rate inflection point detection, which will be explained below.

[0029] Event-anchored estimation of physiological response delay. The multimodal feature fusion module 3 determines the excitation moment when the ultrafiltration rate undergoes a step change in the excitation modality data and the response moment when the relative blood volume signal in the response modality data shows a corresponding change. The time difference between the excitation moment and the response moment is defined as the physiological response delay. Considering that the ultrafiltration rate is often continuously fine-tuned in the bedside environment of a hemodialysis unit, and that patient movement and positional changes introduce baseline drift, the physiological response delay obtained by directly estimating the cross-correlation of the entire signal is not stable. Therefore, the multimodal feature fusion module 3 acquires prescription adjustment events where the dialysis device actively adjusts the ultrafiltration rate according to a preset dialysis prescription. Using these prescription adjustment events as known excitation markers, the physiological response delay is estimated only within a preset time window starting from the prescription adjustment event. Specifically, within a time window starting from the excitation moment and with a length consistent with the sliding time window, a normalized cross-correlation is calculated between the change sequence of the excitation modality data and the change sequence of the response modality data. The time shift that maximizes the normalized cross-correlation is taken as the physiological response delay. The calculation formula is as follows:

[0030] ,in: Let be the physiological response delay, and be a scalar with a value range of . The unit is s, which is calculated by taking the time shift corresponding to the maximum value of the normalized cross-correlation in this formula. It characterizes the capacity of the circulatory system of dialysis patients to compensate for ultrafiltration dehydration. The larger the value, the slower the compensation response and the worse the capacity to compensate. The operator that takes the value of the argument that maximizes the objective function; Let be the time-shift variable to be searched, be a scalar, and take values ​​ranging from 1 to 2. The unit is s, which is determined by iterating through the search interval value by value. The lower bound of the time delay search is 0s, which is a scalar. In this embodiment, it is taken as 0s, and the unit is s. It is determined by the shortest physiological time course of blood volume change caused by dialysis ultrafiltration. The upper bound of the time delay search is a scalar value, which is 60s in this embodiment. The unit is s, and it is determined by the length of the sliding time window and the typical compensation time course. If the value is too small, it will truncate the true time delay of the slow patient; if the value is too large, it will introduce irrelevant fluctuations. For discrete time variables within a time window, it is a scalar with the unit of seconds, determined by the acquisition timestamp; The excitation time is denoted by s, which is a scalar quantity in seconds and is determined by the occurrence time of the prescription adjustment event. The time window length is a scalar value; in this embodiment, it is taken as 120 seconds, and the unit is seconds, which is consistent with the sliding time window length. To stimulate modal data in The change in ultrafiltration rate at time t relative to the excitation time is a scalar quantity in L / h, calculated from the ultrafiltration rate sequence in the excitation modal data. In response to modal data in The change in relative blood volume signal at time t is relative to the excitation time. It is a scalar, dimensionless quantity, and is calculated from the relative blood volume signal sequence in the response modal data. Indicates to since to Summing each one individually; The operator that takes the arithmetic square root.

[0031] Calculation of the compensatory response time-series ratio. The multimodal feature fusion module 3, based on the response onset delay, decomposes the response modal data within the time window into a fast response component and a slow response component. The fast response component corresponds to the rapid compensatory process dominated by baroreceptor reflex and vasoconstriction, while the slow response component corresponds to the slow compensatory process dominated by interstitial fluid refilling into blood vessels; both have different response onset delays. The multimodal feature fusion module 3 fits the response curve within the time window as the superposition of the fast and slow components, obtaining the onset delays of the fast and slow response components respectively, and determines the ratio of the two as the compensatory response time-series ratio. The calculation formula is as follows:

[0032] ,in: The compensatory response time ratio is a scalar with a value range of [value range missing]. , dimensionless, is calculated by this formula, and represents the relative relationship between the fast compensation process and the slow compensation process in terms of response start time. The closer its value is to 1, the more it indicates that the fast compensation process is relatively lagging and suggests that the fast compensation pathway function is weakened. Let be the start delay of the fast response component, and let be a scalar with a value range of . The unit is seconds (s), which is obtained by taking the fast response component after decomposing the fast and slow components of the response curve and taking the time difference between the moment when the fast response component deviates from the baseline by a set proportion and the excitation time. Let be the initial delay of the slow response component, and let be a scalar with a value range of . The unit is seconds (s), which is obtained by taking the time difference between the moment when the slow response component deviates from the baseline by a set proportion after decomposing the fast and slow components of the response curve and the excitation time.

[0033] Online recursive identification of time-varying response kernels. The multimodal feature fusion module 3 models the mapping from excitation modal data to response modal data as a time-varying response kernel. The kernel parameters of the time-varying response kernel include the response time constant and response gain. Considering that as dialysis progresses and the cumulative ultrafiltration volume increases, the patient's circulatory compensatory reserve is gradually depleted, and the response characteristics of the circulatory system to ultrafiltration excitation are time-varying, the multimodal feature fusion module 3 discretizes this mapping into a first-order difference model with time-varying coefficients, and recursively updates the kernel parameters based on the excitation modal data and response modal data within each time window, obtaining a kernel parameter sequence that changes with the dialysis process. The recursive update of the kernel parameters uses recursive least squares with a forgetting factor, and its update formula is:

[0034] ,

[0035] ,

[0036] ,

[0037] in: For the first The kernel parameter estimation vector at each sampling time point is a two-dimensional vector, whose two components are the autoregressive coefficients of the discrete first-order model. With input coefficients , dimensionless, is calculated by this recursive formula, and represents the mapping relationship from the excitation modal data to the response modal data at the current moment; For the first The kernel parameter estimation vector at each sampling time point, its meaning, type, and... The same, given by the recursive result of the previous sampling time; The sampling time sequence number is a positive integer, dimensionless, and is determined by the order of sampling points within the time window. For the first The gain vector at each sampling time point is a two-dimensional, dimensionless vector, calculated by this formula, and is used to determine the correction magnitude of the new observation to the nuclear parameter estimate. In response to modal data at the first The relative blood volume signal sampled at each sampling time is a scalar, dimensionless quantity, given by the response mode data; For the first The regression vector at each sampling time is a two-dimensional vector, whose two components are the response sample value at the previous time step. and lag Excitation sampled values ​​of each sampling point The values ​​are given by the response modal data and the excitation modal data; To excite modal data in the first The ultrafiltration rate sampled at each sampling time is a scalar, with units of L / h, and is given by the excitation modal data; The number of sampling points corresponding to the time delay is a non-negative integer, dimensionless, and is determined by the physiological response time delay. Divide by the sampling period and round down; superscript The transpose operator for a vector or matrix; For the first The covariance matrix at each sampling time point is a dimensionless 2x2 matrix, which is obtained recursively from this formula and is used to characterize the uncertainty of kernel parameter estimation. For the first The covariance matrix at each sampling time, its meaning, type and... The same, given by recursion from the previous sampling time; Let be the forgetting factor, and be a scalar with a range of values ​​of . In this embodiment, 0.98 is used. It is dimensionless and is determined by the need to track the time-varying rate of the response characteristics during the dialysis process. If the value is too small, the estimation will be overly sensitive to noise; if the value is too large, the estimation will lag in tracking the time-varying characteristics. The sampling period is denoted by , and is a scalar; in this embodiment, it is taken as 1 second, with the unit being seconds, determined by the sampling interval of the excitation modal data. After obtaining... Then, the response time constant and response gain are respectively determined by and The conversion yields, where The response time constant is expressed in seconds. For response gain, dimensionless; the gain of each time window and Arranging them in chronological order constitutes the kernel parameter sequence.

[0038] Inflection point detection of response time constant decay rate. The multimodal feature fusion module 3 calculates the rate of change of the response time constant in the kernel parameter sequence to obtain the response time constant decay rate. When the cyclic compensation reserve depletes, the response time constant decay rate changes from an approximately linear increase to a nonlinear accelerated increase. The multimodal feature fusion module 3 detects this transition by taking the normalized second-order difference of the response time constant decay rate. The formulas for calculating the response time constant decay rate and its nonlinear acceleration index are as follows:

[0039] ,

[0040] ,

[0041] in: For the first The decay rate of the response time constant at each sampling time is a scalar, dimensionless quantity, calculated by this formula, and characterizes how fast the response time constant changes with the dialysis process. For the first The response time constant at each sampling time is defined as the conversion formula in the aforementioned recursive identification, and its unit is seconds. For the first The response time constant at each sampling time, in seconds; The definition is the same as the aforementioned recursive formula, and the unit is s; For the first The nonlinear acceleration index at each sampling time is a scalar, and its value range is... , dimensionless, is calculated by this formula, and characterizes the degree to which the decay rate of the response time constant deviates from linear growth. The larger the value, the more nonlinear the decay accelerates. Operators that take the absolute value; , The first , The decay rate of the response time constant at each sampling moment, its meaning, type and same; is the baseline statistical window length, a positive integer, which is 10 in this embodiment. It is dimensionless and is determined by the need for a stationary estimation of the baseline level of decay rate. The summation sequence variable within the baseline statistical window is a positive integer and dimensionless. Indicates to since to Summing each one individually; To prevent small positive numbers with a denominator of zero from being scalars, this embodiment takes... Dimensionless. When the nonlinear acceleration index When the threshold value is exceeded continuously, the multimodal feature fusion module 3 marks the moment of the first exceedance as the precursor point of compensatory reserve depletion, and uses the response time constant decay rate sequence and the precursor point of compensatory reserve depletion together as the hypotension precursor channel in the cross-modal fusion features.

[0042] Excitation-response coupling fusion. Multimodal feature fusion module 3 combines physiological response delay, compensatory response timing ratio, kernel parameter sequence, and hypotension precursor channels to form an excitation-response coupling feature vector, and fuses single-modal features with the excitation-response coupling features to obtain cross-modal fusion features. Unlike fusion methods that symmetrically concatenate modalities, this embodiment employs a fusion method that uses the excitation mode as a gate and asymmetrically modulates the response mode: A gate weight is calculated using the excitation mode's single-modal feature vector and the excitation-response coupling feature vector. This gate weight is then used to weight the response mode's single-modal feature vector, and finally concatenated with the excitation mode's single-modal feature vector and the excitation-response coupling feature vector. The calculation formula is as follows:

[0043] ,

[0044] ,

[0045] in: Let be the coupling gate vector, and be A dimensional vector, where each component takes values ​​ranging from 1 to 2. , dimensionless, is calculated by this formula and is used to perform asymmetric weighting of each characteristic component of the response mode based on the excitation mode and coupling relationship, reflecting the physical direction of the excitation-driven response; The normalized exponential function is defined as taking the exponent of each component of the input vector and dividing it by the sum of the exponents of all components, so that the output components are non-negative and their sum is 1. Here is the gated weight matrix, and here is... A dimensionless matrix, learned through the training process; For the excitation mode single-mode eigenvector, for A dimensionless vector, obtained by normalizing the single-modal features of the excitation modal data; Let be the excitation-response coupled feature vector, and be... The dimensionless vector is obtained by concatenating and normalizing the physiological response delay, the compensatory response timing ratio, the kernel parameter sequence and the hypotension precursor channel. To respond to the single-mode eigenvector, for A dimensionless vector, obtained by normalizing the single-modal features of the response modal data; This represents the vector concatenation operator; The Hadamard product operator represents the multiplication of corresponding vector components; For cross-modal fusion feature vectors, The dimensionless vector, calculated by this formula, serves as the input for the hypotension risk level identification module 4. , , , respectively, are the dimensions of the excitation mode single-mode features, the response mode single-mode features, and the excitation-response coupling features, all of which are positive integers and dimensionless.

[0046] The hypotension identification and early warning module 4 is used to identify the hypotension risk level of dialysis patients based on cross-modal fusion features and output graded early warning signals according to the hypotension risk level. The hypotension identification and early warning module 4 inputs the cross-modal fusion feature vector into the hypotension risk identifier, which outputs a hypotension risk value and determines the hypotension risk level according to the range of the hypotension risk value. In this embodiment, the hypotension risk level includes low risk, medium risk, and high risk. A hypotension risk value less than 0.3 is considered low risk; a hypotension risk value not less than 0.3 and not greater than 0.7 is considered medium risk; and a hypotension risk value greater than 0.7 is considered high risk.

[0047] To improve the stability of hypotension risk identification across different patients, the hypotension identification and early warning module 4 constructs a physical constraint term based on plasma refill kinetics during the training of the hypotension risk identifier. This physical constraint term is used as a regularization term to constrain the mapping from cross-modal fusion features to hypotension risk levels. The physical constraint term characterizes the monotonically increasing relationship between the response time constant decay rate and the cumulative ultrafiltration volume; that is, as the cumulative ultrafiltration volume increases, the response time constant decay rate should remain monotonically constant. The physical constraint term penalizes violations of this monotonic relationship, and its calculation formula is as follows:

[0048] ,in: Let be the total loss during training of the low blood pressure risk identifier, and let be a scalar with a value range of . , dimensionless, is calculated by this formula and is used to guide the parameter optimization of the hypotension risk identifier; Let be the classification loss, and let be a scalar with a value range of . , dimensionless, is calculated from the cross-entropy between the output of the hypotension risk identifier and the hypotension event label; Here, represents the physical constraint weight, and is a scalar with a value range of . In this embodiment, 0.1 is used, which is dimensionless and determined by the relative importance of the classification loss and the physical constraint term. If the value is too small, the physical constraint will not be effective enough; if the value is too large, it will weaken the discriminative information of the data itself. Indicates the sampling time sequence number From 2 to Summing each one individually; The total number of sampling moments during the dialysis process is a positive integer, dimensionless, and is determined by the dialysis duration and sampling cycle. An operator that takes the larger of zero and the value within the parentheses; , The decay rate is the response time constant, defined as in the aforementioned inflection point detection formula, and is dimensionless. This indicates an indicator function that takes the value 1 if the condition within the square brackets is true, and 0 otherwise. For the first The cumulative ultrafiltration amount at each sampling time is a scalar, and its value ranges from [value missing]. The unit is mL, and it is obtained by integrating the ultrafiltration rate over time. For the first The cumulative ultrafiltration volume at each sampling time, in mL.

[0049] The hypotension identification and early warning module 4 outputs graded early warning signals according to the hypotension risk level, which are then presented via the human-machine interface terminal 5. The human-machine interface terminal 5 is used to present graded early warning signals according to the hypotension risk level. When the hypotension risk level is low, the human-machine interface terminal 5 continuously displays the current hypotension risk value and various physiological signal waveforms on the display interface in a normal manner. When the hypotension risk level is medium, the human-machine interface terminal 5 outputs a display warning, highlighting the hypotension risk value in a prominent color on the display interface and prompting medical staff to pay closer attention. When the hypotension risk level is high, the human-machine interface terminal 5 outputs an audible and visual warning, and simultaneously sends an ultrafiltration rate reduction command to the dialysis equipment to assist medical staff in timely intervention. It should be noted that the ultrafiltration rate reduction command only applies to the operating parameter settings of the dialysis equipment, and the final confirmation of the dialysis prescription is made by medical staff.

[0050] Storage module 6 stores the stimulus-response coupling features from the patient's historical dialysis process, obtaining the patient's individual coupling feature baseline. For each patient, storage module 6 aligns the stimulus-response coupling features extracted from multiple historical dialysis processes according to the dialysis progress and takes the statistical center as the patient's individual coupling feature baseline. Hypotension identification and early warning module 4, when identifying the hypotension risk level, further compares the current stimulus-response coupling features with the patient's individual coupling feature baseline. When the current physiological response delay is significantly longer than the patient's individual coupling feature baseline, or the response time constant decay rate is significantly higher than the patient's individual coupling feature baseline, the hypotension risk level is adjusted upward accordingly. By introducing the comparison of the patient's individual coupling feature baseline, this system can adapt to individual differences in dry weight, cardiac function, and autonomic nervous function among different patients, further improving the adaptability of hypotension risk identification.

[0051] During a complete hemodialysis treatment, the multimodal data acquisition module 1 continuously and synchronously acquires the patient's physiological signals and the dialysis equipment's operating parameters; the data preprocessing module 2 denoises and segments the acquired signals, continuously outputting new preprocessed signal segments in 30-second increments; the multimodal feature fusion module 3 determines the excitation modal data and response modal data for each new preprocessed signal segment, extracts single-modal features and excitation-response coupling features, and fuses them to obtain cross-modal fusion features; the hypotension identification and early warning module 4 identifies the current hypotension risk level and outputs a graded early warning signal; the human-computer interaction terminal 5 presents the information, and the storage module 6 synchronously accumulates excitation-response coupling features to update the patient's individual coupling feature baseline. Thus, this system achieves real-time, early, and graded monitoring and early warning of hypotension risk in dialysis patients at the bedside.

[0052] like Figure 2 As shown, this embodiment provides an intelligent monitoring method for hypotension in dialysis patients based on multimodal signals. This method can be executed by the system described in the aforementioned system embodiment. The specific implementation methods of denoising, segmentation, feature extraction, stimulus-response coupling feature extraction, and hypotension risk identification involved in each step of the method are the same as the implementation methods of each module in the system embodiment, and the similarities will not be repeated. The method includes the following steps.

[0053] Step S1 involves simultaneously acquiring the patient's physiological signals and the dialysis equipment's operating parameters. The patient's physiological signals include continuous blood pressure, electrocardiogram (ECG), and relative blood volume signals, acquired by a continuous blood pressure sensor, a three-lead ECG electrode, and an optical blood volume sensor located at the arterial end of the extracorporeal circulation tubing, respectively. The dialysis equipment's operating parameters include ultrafiltration rate and dialysate conductivity, read from the dialysis equipment via a data interface. Each signal is labeled with a uniform acquisition timestamp, but the inherent time difference between excitation and response is not eliminated.

[0054] Step S2 involves denoising and segmenting the patient's physiological signals and dialysis equipment operating parameters to obtain preprocessed signal segments. Specifically, bandpass filtering with a passband of 0.5Hz to 40Hz is used for the electrocardiogram signal, low-pass filtering with a cutoff frequency of 20Hz is used for the continuous blood pressure signal, wavelet threshold denoising based on discrete wavelet transform is used for the relative blood volume signal, and median filtering is used for the ultrafiltration rate and dialysate conductivity. Segmentation uses a sliding time window with a length of 120s and a step size of 30s.

[0055] Step S3: From the preprocessed signal segments, the signal segments corresponding to the dialysis equipment operating parameters are identified as excitation modal data, and the signal segments corresponding to the patient's physiological signals are identified as response modal data. Single-modal features of the excitation modal data and response modal data are extracted respectively. Excitation-response coupling features between the excitation modal data and response modal data are extracted, and the single-modal features and excitation-response coupling features are fused to obtain cross-modal fusion features.

[0056] Specifically, when extracting the excitation-response coupling features, firstly, the prescription adjustment event, in which the dialysis device actively adjusts the ultrafiltration rate according to the preset dialysis prescription, is used as a known excitation marker. Within a preset time window starting from the prescription adjustment event, the normalized cross-correlation between the change sequence of the excitation modal data and the change sequence of the response modal data is calculated. The time shift that maximizes the normalized cross-correlation is taken as the physiological response delay, which characterizes the circulatory system's capacity compensation for ultrafiltration dehydration in dialysis patients. Secondly, based on the response onset delay, the response within the time window is decomposed into fast response and slow response components. The ratio of the onset delay of the fast response component to the onset delay of the slow response component is determined as the compensatory response time ratio. Thirdly, the mapping from excitation modal data to response modal data is modeled as a time-varying response kernel. Recursive least squares with a forgetting factor is used to recursively update the response time constant within each time window. The kernel parameters of the response gain and response time are calculated to obtain a kernel parameter sequence. Then, the rate of change of the response time constant in the kernel parameter sequence is calculated to obtain the response time constant decay rate. The normalized second difference of the response time constant decay rate is obtained to obtain a nonlinear acceleration index. When the nonlinear acceleration index continuously exceeds a set inflection point threshold, the moment of the first exceedance is marked as the pre-excitation point of compensatory reserve depletion. The response time constant decay rate sequence and the pre-excitation point of compensatory reserve depletion together constitute a hypotension precursor channel. Finally, the physiological response delay, compensatory response timing ratio, kernel parameter sequence and hypotension precursor channel constitute an excitation-response coupling feature. The gating weight is calculated using the excitation mode single-mode feature vector and the excitation-response coupling feature vector. The response mode single-mode feature vector is weighted with the gating weight and then concatenated with the excitation mode single-mode feature vector and the excitation-response coupling feature vector to obtain a cross-modal fusion feature. The calculation formulas involved in the above steps are the same as those used in the multimodal feature fusion module 3 in the system embodiment.

[0057] Step S4: Identify the hypotension risk level of dialysis patients based on cross-modal fusion features, and output graded warning signals according to the hypotension risk level. Specifically, the cross-modal fusion features are input into the hypotension risk identifier to obtain a hypotension risk value. The low-risk, medium-risk, or high-risk level is determined according to the interval in which the hypotension risk value falls. During training, the hypotension risk identifier uses a physical constraint term constructed based on plasma refill kinetics as a regularization term to constrain the mapping from cross-modal fusion features to hypotension risk levels. When the hypotension risk level is medium-risk, a display warning is output; when the hypotension risk level is high-risk, an audible and visual warning is output, and an ultrafiltration rate reduction command is sent to the dialysis equipment.

[0058] Furthermore, the method also includes: storing the stimulus-response coupling characteristics of dialysis patients during their historical dialysis processes to obtain a patient-specific coupling characteristic baseline; comparing the current stimulus-response coupling characteristics with the patient-specific coupling characteristic baseline; and adjusting the hypotension risk level based on the comparison results to accommodate individual differences among different patients.

[0059] Through steps S1 to S4 above, under the constraint of real-time bedside monitoring in the hemodialysis room, this method can identify the depletion of circulatory compensatory reserve in advance from the temporal coupling relationship between the ultrafiltration stimulation of the dialysis machine and the patient's multimodal physiological response, thereby realizing real-time, early, and graded monitoring and early warning of hypotension risk in dialysis patients.

[0060] 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 and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart monitoring system for hypotension in dialysis patients based on multimodal signals, characterized in that, include: A multimodal data acquisition module is used to simultaneously acquire the patient's physiological signals and the dialysis equipment operating parameters. The patient's physiological signals include continuous blood pressure signals, electrocardiogram signals, and relative blood volume signals. The dialysis equipment operating parameters include ultrafiltration rate and dialysate conductivity. The data preprocessing module is used to denoise and segment the patient's physiological signals and the dialysis equipment's operating parameters to obtain preprocessed signal segments. A multimodal feature fusion module is used to determine the signal segments corresponding to the operating parameters of the dialysis device as excitation modal data and the signal segments corresponding to the patient's physiological signals as response modal data from the preprocessed signal segments. It extracts single-modal features from the excitation modal data and the response modal data respectively, and extracts the excitation-response coupling features between the excitation modal data and the response modal data. The single-modal features and the excitation-response coupling features are fused to obtain cross-modal fusion features. The excitation-response coupling features characterize the temporal coupling relationship where changes in the excitation modal data drive changes in the response modal data. When extracting the excitation-response coupling features, the multimodal feature fusion module determines the excitation moment when the ultrafiltration rate undergoes a step change in the excitation modal data and the response moment when the relative blood volume signal undergoes a corresponding change in the response modal data. The time difference between the excitation moment and the response moment is determined as the physiological response delay, which is used as a component of the excitation-response coupling features. The physiological response delay characterizes the circulatory system's capacity to compensate for ultrafiltration dehydration in dialysis patients. The excitation-response coupling features include physiological response delay estimation, compensatory response timing ratio calculation, time-varying response kernel identification, and response time constant decay rate inflection point detection. The hypotension identification and early warning module is used to identify the hypotension risk level of dialysis patients based on the cross-modal fusion features, and output graded early warning signals according to the hypotension risk level.

2. The intelligent monitoring system for hypotension in dialysis patients based on multimodal signals according to claim 1, characterized in that, The multimodal feature fusion module is also used to: acquire prescription adjustment events in which the dialysis device actively adjusts the ultrafiltration rate according to a preset dialysis prescription, and use the prescription adjustment events as known excitation markers to estimate the physiological response delay within a preset time window starting from the prescription adjustment events; Based on the response start delay, the response modal data within the time window is decomposed into a fast response component and a slow response component. The ratio of the start delay of the fast response component to the start delay of the slow response component is determined as the compensatory response timing ratio, which is used as a supplementary component of the excitation-response coupling feature.

3. The intelligent monitoring system for hypotension in dialysis patients based on multimodal signals according to claim 2, characterized in that, The multimodal feature fusion module models the mapping from the excitation modal data to the response modal data as a time-varying response kernel. The kernel parameters of the time-varying response kernel include the response time constant and the response gain. The multimodal feature fusion module recursively updates the kernel parameters according to the excitation modal data and the response modal data within each time window to obtain a kernel parameter sequence that changes with the dialysis process. The kernel parameter sequence is used as a component of the excitation-response coupling feature.

4. The intelligent monitoring system for hypotension in dialysis patients based on multimodal signals according to claim 3, characterized in that, The multimodal feature fusion module calculates the rate of change of the response time constant in the kernel parameter sequence to obtain the response time constant decay rate. The multimodal feature fusion module calculates the normalized second difference of the response time constant decay rate to obtain a nonlinear acceleration index, which characterizes the degree to which the response time constant decay rate deviates from linear growth. When the nonlinear acceleration index continuously exceeds a set inflection point threshold, the moment when the nonlinear acceleration index first exceeds the inflection point threshold is marked as a precursor point of compensatory reserve depletion, and the response time constant decay rate and the precursor point of compensatory reserve depletion are used as the hypotension precursor channel in the cross-modal fusion features.

5. The intelligent monitoring system for hypotension in dialysis patients based on multimodal signals according to claim 4, characterized in that, The multimodal feature fusion module constructs physical constraint terms based on the plasma refill kinetics relationship. These physical constraint terms characterize the monotonically increasing relationship between the decay rate of the response time constant and the cumulative ultrafiltration volume. When identifying the low blood pressure risk level, the low blood pressure identification and early warning module uses the physical constraint term as a regularization term to constrain the mapping of the cross-modal fusion features to the low blood pressure risk level.

6. The intelligent monitoring system for hypotension in dialysis patients based on multimodal signals according to claim 1, characterized in that, The multimodal data acquisition module includes an optical blood volume sensor, which is disposed at the arterial end of the extracorporeal circulation tubing for dialysis and is used to acquire the relative blood volume signal by measuring light transmittance.

7. The intelligent monitoring system for hypotension in dialysis patients based on multimodal signals according to claim 1, characterized in that, It also includes a human-computer interaction terminal, wherein the low blood pressure risk level includes low risk level, medium risk level and high risk level; the human-computer interaction terminal outputs a display warning when the low blood pressure risk level is medium risk level, outputs an audio-visual warning when the low blood pressure risk level is high risk level and sends an ultrafiltration rate reduction command to the dialysis equipment.

8. The intelligent monitoring system for hypotension in dialysis patients based on multimodal signals according to claim 1, characterized in that, It also includes a storage module, which is used to store the stimulus-response coupling characteristics of dialysis patients in the history of dialysis to obtain the patient's individual coupling characteristic baseline; the hypotension identification and early warning module is also used to compare the current stimulus-response coupling characteristics with the patient's individual coupling characteristic baseline, and adjust the hypotension risk level according to the comparison result.

9. A method for intelligent monitoring of hypotension in dialysis patients based on multimodal signals, comprising the intelligent monitoring system for hypotension in dialysis patients based on multimodal signals as described in any one of claims 1-8, characterized in that, Includes the following steps: The system simultaneously collects physiological signals from dialysis patients and operating parameters of the dialysis equipment. The physiological signals include continuous blood pressure signals, electrocardiogram signals, and relative blood volume signals. The operating parameters of the dialysis equipment include ultrafiltration rate and dialysate conductivity. The patient's physiological signals and the dialysis equipment's operating parameters are denoised and segmented to obtain preprocessed signal segments; From the preprocessed signal segments, the signal segments corresponding to the operating parameters of the dialysis device are identified as excitation modal data, and the signal segments corresponding to the patient's physiological signals are identified as response modal data. Single-modal features are extracted from the excitation modal data and the response modal data, respectively. Excitation-response coupling features between the excitation modal data and the response modal data are extracted, and the single-modal features and the excitation-response coupling features are fused to obtain cross-modal fusion features. The excitation-response coupling features characterize the temporal coupling relationship where changes in the excitation modal data drive changes in the response modal data. When extracting the excitation-response coupling features, the excitation moment when the ultrafiltration rate undergoes a step change in the excitation modal data is determined, and the response moment when the relative blood volume signal in the response modal data shows a corresponding change is determined. The time difference between the excitation moment and the response moment is determined as the physiological response delay, and the physiological response delay is used as a component of the excitation-response coupling features. The physiological response delay characterizes the circulatory system's capacity to compensate for ultrafiltration dehydration in dialysis patients. The low blood pressure risk level of dialysis patients is identified based on the cross-modal fusion features, and a graded early warning signal is output according to the low blood pressure risk level.

Images