Multi-parameter intelligent early warning method and system for hemodialysis hypotension risk

Abstract

This invention discloses a multi-parameter intelligent early warning method and system for the risk of hypotension during hemodialysis, belonging to the field of intelligent medical data analysis and clinical decision support technology. The method addresses the inherent contradiction between prediction lead time and accuracy caused by the heterogeneity of multiple etiologies in dialysis-related hypotension events. The perception layer reconstructs the phase space of the hemodynamic time series and detects the state transition signals of the hemodynamic system in real time by calculating kinetic indicators. The cognition layer inputs the kinetic indicators as novel observational evidence into the causal inference network, infers the etiological type of hypotension in real time, and outputs the corresponding optimal prediction window. The simulation layer dynamically adjusts the weight configuration of the physiological constraint equations in the physical information neural network based on the etiological classification results, constructing an individualized digital twin for forward simulation prediction. The system can achieve a dynamic optimal balance between prediction lead time and accuracy, and outputs a graded early warning and targeted intervention suggestions containing three layers of interpretable evidence.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data analysis and clinical decision support technology, specifically to a multi-parameter intelligent early warning method and system for the risk of hypotension during hemodialysis. Background Technology

[0002] Hemodialysis is one of the main treatment methods for maintaining the lives of patients with end-stage renal disease, with more than 3 million people worldwide receiving regular hemodialysis treatment. During dialysis, due to the reduction in effective circulating blood volume caused by ultrafiltration and dehydration, coupled with individual differences in patients' cardiovascular compensatory capacity, intradialytic hypotension (IDH) is the most common acute complication, with an incidence of approximately 20% to 30%. Clinically, intradialytic hypotension is usually defined as a decrease in systolic blood pressure of more than 20 mmHg or a decrease in mean arterial pressure of more than 10 mmHg, accompanied by clinical symptoms. Intradialytic hypotension not only causes discomfort such as dizziness, nausea, convulsions, and loss of consciousness, but can also lead to serious consequences such as myocardial ischemia, cerebral insufficiency, mesenteric ischemia, and vascular thrombosis, making it a key factor affecting dialysis quality and long-term patient prognosis.

[0003] From a pathophysiological perspective, the mechanisms of hypotension during dialysis involve multiple etiologies. The first is volume depletion type, where the ultrafiltration rate exceeds the vascular refill rate, leading to a progressive decrease in effective circulating blood volume, a drop in cardiac output, and ultimately, a drop in blood pressure. This type typically presents as a gradual process with early warning signs. The second is vasodilatory type, mainly related to peripheral vasodilation and decreased vascular resistance caused by excessively high dialysate temperature; its development can be rapid. The third is neuroregulatory disorder type, commonly seen in patients with diabetic autonomic neuropathy. Due to impaired baroreflex function, the body cannot effectively compensate for decreased blood volume by increasing heart rate and vascular resistance, resulting in a sharp drop in blood pressure. In clinical practice, the same patient may exhibit different etiologies of hypotension at different dialysis sessions, and even multiple etiologies may overlap during the same dialysis session.

[0004] Currently, the prevention of hypotension during dialysis in clinical practice mainly relies on the following methods: First, the subjective judgment of medical staff based on clinical experience, observing the patient's condition by regularly measuring blood pressure. However, conventional cuff blood pressure monitoring is usually measured every 15 to 30 minutes, which has a monitoring blind spot and may miss the critical period of rapid blood pressure drop. Second, empirical intervention based on online blood volume monitoring, which triggers intervention when the relative blood volume decreases beyond a certain fixed threshold. However, the fixed threshold cannot adapt to individual differences among different patients. Third, prediction methods based on machine learning, which use historical dialysis records to train classification or regression models to predict the risk of hypotension.

[0005] Existing machine learning-based techniques for predicting hypotension during dialysis have the following prominent problems:

[0006] First, there is a contradiction between fixed prediction windows and heterogeneity across multiple etiologies. Most existing methods use fixed 15-minute or 30-minute prediction windows for binary prediction. However, different etiologies of hypotension events have different rates of progression and optimal intervention timing. Volume depletion hypotension develops gradually, with precursor signals detectable 25 to 30 minutes before the event, making a longer prediction window suitable; while vasodilatory hypotension develops rapidly, with an optimal prediction window of only 10 to 15 minutes. Fixed prediction windows lead to premature predictions (more noise, lower accuracy) for some types of hypotension events and delayed predictions (insufficient intervention time) for others, failing to simultaneously achieve the optimal balance between prediction lead time and accuracy.

[0007] Second, reliance on surface features leads to insufficient detection of precursor signals. Existing models mainly utilize traditional features such as changes in blood pressure amplitude, relative changes in blood volume, and cumulative ultrafiltration volume for prediction. These features essentially reflect the surface state variables of the hemodynamic system. However, before a hypotensive event occurs, the hemodynamic system often undergoes changes in its deeper dynamic structure (such as the transition from a steady state to an unstable state). These changes in deeper structure precede significant changes in blood pressure amplitude. Traditional methods lack effective means to detect precursor signals at the level of system dynamic structure. When a significant downward trend in blood pressure is observed, there is often insufficient time left for clinical intervention.

[0008] Third, the lack of interpretability hinders clinical adoption. While deep learning black-box models may achieve high predictive performance, their decision-making process lacks transparency. When the system issues a hypotension warning, clinicians need to know the main risk contributing factors in order to take targeted interventions. For example, if the main cause of hypotension is excessively rapid ultrafiltration, the ultrafiltration rate should be reduced; if the main cause is vasodilation, the dialysate temperature should be lowered. Black-box models cannot provide this etiological attribution information, leading to insufficient trust in the warning results among healthcare workers, who often rely on their own experience and ignore system warnings. Studies have shown that dialysis nurses' compliance rate with warning systems lacking interpretation is less than 50%, severely limiting the clinical application value of warning technologies.

[0009] Fourth, the fixed nature of physical constraints leads to insufficient model specificity. In recent years, some studies have attempted to apply Physics-Informed Neural Networks (PINNs) to hemodynamic modeling, embedding physiological differential equations as constraints into the loss function of the neural network. However, existing methods set the constraint weights of each physiological sub-equation as fixed hyperparameters, failing to adaptively adjust the emphasis of physical constraints based on the actual etiological types exhibited by different patients during dialysis. For example, when a patient's hypotension is mainly caused by blood volume depletion, the constraints of the capillary refill equation and the Frank-Starling equation should be strengthened; while when it is mainly caused by vasodilation, the constraints of the vascular compliance equation should be strengthened. Fixed weights prevent the model from automatically focusing on the most relevant physical processes based on real-time etiological judgment, limiting further improvements in prediction accuracy.

[0010] Fifth, setting individualized parameters for phase space reconstruction methods is difficult. Phase space reconstruction technology in nonlinear dynamics theory can extract state transition information from the deep dynamic structure of time series, theoretically enabling ultra-early detection of hypotension events. However, the setting of key parameters of this method (such as embedding dimension, delay time, Lyapunov exponential warning threshold, etc.) largely depends on the physiological characteristics of individual patients. Differences in vascular compliance and autonomic nervous function among different patients significantly affect the distribution of phase space dynamic characteristics. Fixed statistical prior parameters are difficult to adapt to all patients, leading to excessive false positives or false negatives for some patients.

[0011] Therefore, how to determine the cause of hypotension in real time during dialysis, and adaptively select the optimal prediction window and physical constraint configuration based on the cause, while providing interpretable early warning evidence and targeted intervention suggestions, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0012] Technical Objective: To address the issue of hypotension during hemodialysis due to its multi-etiological heterogeneity, with different etiologies exhibiting varying optimal prediction windows and precursor signal characteristics, existing technologies employing fixed prediction windows and physical constraint weights fail to simultaneously achieve an optimal balance between prediction lead and accuracy. Furthermore, they lack the ability to detect precursor signals at the system dynamics level and provide interpretability support. This invention discloses a multi-parameter intelligent early warning method and system for hypotension risk in hemodialysis, constructing a three-layer, dual-loop feedback prediction architecture consisting of a perception layer, a cognitive layer, and a simulation layer. The perception layer extracts state transition signals from the deep dynamic structure of hemodynamic time series, enabling ultra-early detection of hypotension events. The cognitive layer inputs the dynamic indicators output from the perception layer along with traditional monitoring features into a causal inference network to determine the etiology of hypotension in real time and adaptively select the optimal prediction window. The simulation layer dynamically adjusts the weights of the physiological constraint equations in the physical information neural network based on the etiology classification results from the cognitive layer, constructing an individualized digital twin for forward simulation prediction. Dynamic collaborative parameter calibration is achieved between the three layers through a dual-loop feedback mechanism consisting of a forward feedback loop and a backward calibration loop.

[0013] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0014] A multi-parameter intelligent early warning method for the risk of hypotension during hemodialysis includes the following steps:

[0015] Step S1, Data Acquisition and Preprocessing:

[0016] Multi-source monitoring data is collected in real time during dialysis. In one specific embodiment, the multi-source monitoring data includes, but is not limited to: continuous blood pressure monitoring data (acquired via a cuff-type blood pressure monitor or online continuous blood pressure monitoring device, including systolic blood pressure, diastolic blood pressure, and mean arterial pressure), online blood volume monitoring data (relative changes in blood volume monitored in real time by an optical sensor based on the principle of blood concentration, reflecting the trend of changes in effective circulating blood volume), ultrafiltration rate and cumulative volume data (the ultrafiltration rate setpoint and the completed ultrafiltration cumulative volume output in real time by the dialysis device), dialysate parameter data (including setpoints or measured values ​​of dialysate temperature and dialysate sodium concentration), and patient baseline characteristics (including age, dry weight, predialysis weight, diabetic status, cardiac function classification, history of hypotension, etc.).

[0017] The preprocessing process includes: using an adaptive resampling engine to unify data streams with different sampling rates to a preset time resolution (every 30 seconds in one specific implementation); using a wavelet decomposition-based motion artifact detector to identify and label abnormal data segments caused by patient limb movements, coughing, postural adjustments, etc.; and using an adaptive Z-Score normalization method based on historical distribution to eliminate dimensional differences. The calculation formula for the adaptive Z-Score normalization method is as follows: ,in This represents the normalized current parameter value, where x represents the original parameter value collected at the current time. and These represent the mean and standard deviation of the corresponding parameters from the patient's historical dialysis records. After preprocessing, a multi-parameter time-series matrix with a unified time axis is formed. Where t represents time, This indicates that the matrix belongs to a The system outputs a real matrix space of dimension T, where T is the time step and D is the parameter dimension, and also outputs a data quality confidence vector.

[0018] Step S2, Phase Space Dynamics Sensing:

[0019] This step is the sensing layer of the present invention, which is responsible for detecting state transition signals from the deep dynamic structure of hemodynamic time series, so as to realize ultra-early detection before changes in blood pressure amplitude.

[0020] Specifically, based on Takens' delayed embedding theorem, the blood pressure time series {SBP(t)} is reconstructed into an m-dimensional phase space trajectory. Where m is the embedding dimension and τ is the delay time. The embedding dimension m is automatically determined using the mutual information method: the mutual information function I(τ) of the blood pressure time series under different delays is calculated, and the τ value corresponding to the first local minimum of I(τ) is selected as the optimal delay time. The delay time τ is automatically optimized using the spurious nearest neighbor method: the embedding dimension is gradually increased, and the proportion of spurious nearest neighbors FNN(m) under each dimension is calculated. When FNN(m) drops below a preset threshold, the corresponding m value is the optimal embedding dimension.

[0021] Simultaneously, the blood volume change rate dBV / dt is incorporated as an additional embedding dimension into the phase space, forming an enhanced phase space trajectory. The introduction of this enhanced dimension enables the phase space trajectory to simultaneously reflect the combined state of blood pressure dynamics and blood volume dynamics, providing richer precursor information.

[0022] Three core dynamic metrics are calculated in real time within a sliding window:

[0023] (1) Maximum Lyapunov exponent : Measures the exponential divergence rate of adjacent trajectories in phase space, reflecting the system's sensitivity to small disturbances. When A change from a negative or near-zero value to a significantly positive value indicates a transition from a steady to an unstable state in the hemodynamic system, a transition that precedes a significant decrease in blood pressure amplitude. The calculation employs the Rosenstein algorithm, tracking the separation rate of nearest-neighbor trajectory pairs in phase space.

[0024] (2) Phase space trajectory curvature abrupt change rate The curvature abrupt change is calculated at each time step of the phase space trajectory. When the curvature changes drastically (abruptly) within a short period of time, it indicates that the dynamic structure of the system is undergoing a transformation. The curvature abrupt change rate is defined as the ratio of the absolute value of the time derivative of curvature to its historical moving average.

[0025] (3) Rate of change of attractor dimension The fractal dimension of the phase space attractor was calculated using the box counting method. A significant change in the attractor dimension within a short period indicates a sudden change in system complexity. Stable hemodynamic systems typically have a relatively fixed attractor dimension; when it begins to change significantly, it foreshadows an impending instability.

[0026] when Exceeding the individualization threshold or When a sudden change occurs, a dynamic state transition signal (DSTS) is generated, and the time, intensity, and direction information of the state transition are output. The individualized threshold is based on the prior distribution learned by the offline stage quantile regression method and is dynamically adjusted by the forward feedback loop in step S5.

[0027] Step S3: Causal Inference, Etiological Classification, and Prediction Window Optimization

[0028] This step is the cognitive layer of the present invention, which is responsible for integrating dynamic indicators and traditional monitoring data, diagnosing the most likely cause of hypotension in real time, and selecting the optimal physical model configuration for the digital twin.

[0029] In one specific implementation, the causal inference network is implemented using a causal Bayesian network. A causal directed acyclic graph (DAG) is constructed based on expert knowledge in clinical pathophysiology, defining three core causal links:

[0030] First causal link (blood volume depletion type): This pathway describes the causal mechanism by which excessively high ultrafiltration rates lead to a progressive reduction in effective circulating blood volume, decreased cardiac output, and ultimately, a drop in blood pressure.

[0031] Second causal link (vasodilatory type): This pathway describes the causal mechanism by which higher dialysate temperatures lead to peripheral vasodilation, decreased vascular resistance, and consequently, a drop in blood pressure.

[0032] Third causal link (neurological regulation disorder type): This link describes the causal mechanism by which autonomic dysfunction leads to impaired baroreflex arc function, resulting in an inability to effectively compensate for decreased blood volume and causing a sharp drop in blood pressure.

[0033] A key innovation of this invention is that the dynamic index output in step S2 is inserted as a novel enhanced observational evidence node into the aforementioned causal directed acyclic graph. There is a causal relationship between the increase and the increase in system instability. There is a causal relationship between mutations and dynamic state transitions. This new evidence can provide early observational information for causal inference even before significant changes in blood pressure amplitude occur, significantly advancing the identifiable time window for etiological subtyping.

[0034] A variational inference algorithm is used to calculate the posterior probability P(etiology|evidence) of three etiological types in real time. The optimal prediction window is determined based on the etiological type with the highest posterior probability. The blood volume depletion type corresponds to a long window of 25 to 30 minutes, the vasodilation type corresponds to a short window of 10 to 15 minutes, and the neural regulation disorder type corresponds to a medium window of 15 to 20 minutes. Simultaneously, output physical model configuration commands. This refers to the allocation coefficients of the constraint weights for each sub-equation. When the probability of etiological subtyping changes dynamically during dialysis, the prediction window is also dynamically adjusted accordingly.

[0035] Step S4: Adaptive forward simulation of the causes of physical information digital twins:

[0036] This step is the simulation layer of the present invention, which is responsible for constructing an individualized hemodynamic digital twin based on the etiological classification results of the cognitive layer, and performing forward simulation prediction.

[0037] The Physical Information Neural Network (PINN) employs a dual-branch architecture: the data-driven branch, in one specific implementation, uses a 4-layer LSTM encoder plus a 2-layer fully connected decoder architecture to fit measured monitoring data; the physical constraint branch embeds a set of hemodynamic differential equations, including four sub-equations:

[0038] Frank-Starling's output equation: This describes the relationship between cardiac pumping function and preload, among which... Indicates cardiac output. This represents the functional relationship corresponding to the Frank-Starling mechanism. Preload refers to the heart's preload, and contractility refers to the myocardial contractility.

[0039] Vascular compliance equation: This describes the relationship between blood vessel wall elasticity and pressure-volume, where V represents the blood vessel volume. Where is the vascular compliance coefficient (unit: mL / mmHg), and P is the intravascular pressure. It is a zero-pressure volume.

[0040] Capillary refill equation: This describes the dynamic process of interstitial fluid returning to the blood vessels, in which... This indicates the refilling flow rate of capillaries. The refill factor is expressed as (unit: mL / min / mmHg). Capillary hydrostatic pressure, For interstitial fluid pressure, The reflection coefficient, This represents the osmotic pressure difference of the colloid.

[0041] Pressure Reflection Regulation Equation: This describes the feedback regulation mechanism of the autonomic nervous system on heart rate and vascular resistance, in which... This represents the rate of change of heart rate over time. Pressure reflection gain (unit: bpm / mmHg). For blood pressure set point, This is the actual systolic pressure.

[0042] Joint loss function ,in The fitting error is for the data-driven branch. The residuals of the differential equations for the physical constraint branches, These serve as boundary condition constraints. Another key innovation of this invention lies in... The constraint weights of each sub-equation are no longer fixed hyperparameters, but are determined by the physical model configuration instructions output in step S3. Dynamic adjustment, determined by weighted probability vectors of etiological subtyping. When the causal Bayesian network determines the blood volume depletion type, the constraint weights of the capillary refill equation and the Frank-Starling equation automatically increase; when it determines the vasodilation type, the constraint weights of the vascular compliance equation automatically increase; and when it determines the neuromodulation disorder type, the constraint weights of the baroreflex regulation equation automatically increase.

[0043] Online calibration of individual patient physiological parameter sets using monitoring data from the first 10 minutes after dialysis begins. Including vascular compliance coefficient Pressure reflection gain and capillary refill coefficient In the optimal prediction window Internal forward simulation is performed to generate a prediction of future blood pressure trajectory. Its confidence interval and the time distribution of the probability of hypotension.

[0044] Step S5, Dual-loop feedback dynamic calibration:

[0045] This step is the dual-loop feedback linkage control mechanism of the present invention, which coordinates the information flow and parameter feedback between the three layers.

[0046] Forward feedback loop (simulation layer → perception layer): This loop uses the individual physiological parameter set calibrated online in step S4. Feedback is sent to the phase space reconstruction module in step S2. The specific adjustment rule is: based on the vascular compliance coefficient. The value of the embedding dimension m is adjusted. Patients with lower vascular compliance have higher hemodynamic system complexity and require a higher embedding dimension to fully reconstruct the phase space. When the value is below a preset lower limit (0.5 mL / mmHg in one specific embodiment), the embedding dimension is increased by 1 to 2 dimensions; based on the pressure reflection gain... The value is adjusted to the individualized threshold of the maximum Lyapunov index. Patients with lower baroreflex gain have weaker autonomic regulatory compensation ability and are more prone to state transitions in phase space dynamics. A more conservative detection threshold should be used for earlier detection. When the threshold is lower than the preset lower limit (0.3 bpm / mmHg in one specific embodiment), the threshold is reduced by 15% to 25%.

[0047] Backward calibration loop (perception layer → cognition layer → simulation layer): The continuously updated dynamic indicators in step S2 are passed to the causal inference network in step S3, serving as new observational evidence to update the posterior probability of etiological subtyping in real time. The updated etiological probability and optimal window are immediately passed to step S4 to adjust the physical constraint weights, achieving dynamic self-calibration of the entire prediction system.

[0048] A complete double-loop feedback iteration is executed every 30 to 120 seconds (60 seconds in one specific implementation) to continuously optimize the accuracy of the three-layer prediction as the dialysis process progresses.

[0049] Step S6, Clinical Decision Output:

[0050] Multi-level early warning output is generated by combining the prediction results from the three layers:

[0051] Low-risk alert (blue): When the dynamic indicators show a slight changing trend but have not yet reached the state transition threshold, it is suggested to increase the monitoring frequency.

[0052] Medium-risk warning (yellow): When the kinetic state transition signal has been generated and the probability of a certain etiology type output by the causal inference network exceeds the preset medium-risk probability threshold, the output etiology attribution and predicted blood pressure decline trend are provided, and it is recommended to adjust the ultrafiltration rate or dialysate parameters.

[0053] High-risk emergency alert (red): When the digital twin predicts that blood pressure will drop below the low blood pressure threshold within the optimal prediction window and the confidence level exceeds the preset high-risk confidence level threshold, an emergency warning will be output and targeted intervention suggestions will be generated.

[0054] Each warning level simultaneously outputs three layers of interpretable evidence: a description of system stability changes at the kinetic level, an etiological attribution analysis at the causal level, and a blood pressure trajectory prediction at the simulation level. Targeted intervention recommendations are automatically generated based on the etiological classification results: for cases of blood volume depletion, a reduction in the ultrafiltration rate is recommended; for cases of vasodilation, a reduction in dialysate temperature is recommended; and for cases of neuromodulation disorders, supplementation with normal saline or an increase in dialysate sodium concentration is recommended. Intervention recommendations undergo safety checks for dialysis adequacy targets and exclusion checks for patient contraindications before being output.

[0055] Beneficial Effects: The multi-parameter intelligent early warning method and system for the risk of hypotension during hemodialysis provided by this invention have the following beneficial effects:

[0056] (1) It solves the inherent contradiction between prediction lead time and accuracy: the cause type of hypotension is determined in real time by causal inference network, and the optimal prediction window is selected adaptively according to the cause type. For blood volume depletion type, a long window is used to ensure sufficient intervention time, and for vasodilation type, a short window is used to ensure prediction accuracy, thus achieving dynamic optimal balance.

[0057] (2) Significantly advances the detectable time window of precursor signals: By reconstructing phase space dynamics, the state transition can be detected from the deep structural level of the hemodynamic system. Precursor information can be captured before the blood pressure amplitude changes significantly, advancing the precursor signal detection window by 10 to 20 minutes.

[0058] (3) The physical constraint dynamic adjustment of etiology is realized: the physical constraint weight of the digital twin is no longer a fixed hyperparameter, but is determined in real time by the etiology classification probability vector of the causal inference network, so that the physical information neural network can automatically focus on the most relevant physical process and significantly improve the prediction accuracy.

[0059] (4) Dynamic collaborative calibration of individualized parameters is achieved through dual-loop feedback: the physiological parameters calibrated online by the simulation layer are fed back to the sensing layer to adjust the embedded parameters and dynamic thresholds, which solves the problem of individualized parameter threshold setting in the phase space reconstruction method.

[0060] (5) Provides a highly credible early warning with triple interpretable evidence fusion: Each early warning outputs interpretable evidence at the dynamic, causal and simulation levels. The mutual corroboration of the triple evidence significantly improves clinical trust and the response rate of medical staff. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0062] Figure 1 This is a schematic diagram of the overall process of the intelligent early warning method for the risk of hypotension in hemodialysis according to the present invention.

[0063] Figure 2 This is a system block diagram of the three-layer linkage dual-loop feedback architecture of the perception layer, cognition layer, and simulation layer of the present invention.

[0064] Figure 3 This is a schematic diagram of the directed acyclic graph structure of the causal Bayesian network of the present invention, showing three core causal links and dynamic enhancement observation nodes.

[0065] Figure 4 This is a schematic diagram of the dual-branch architecture of the physical information neural network of the present invention.

[0066] Figure 5 This is a schematic diagram of the data flow of the dual-loop feedback linkage controller of the present invention.

[0067] Figure 6 This is a schematic diagram illustrating the combination of multi-level early warning output and triple interpretable evidence in this invention.

[0068] Figure 7 This is a graph showing the AUC comparison results of different methods in the embodiments of the present invention. Detailed Implementation

[0069] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0070] Example 1: Method Example

[0071] like Figure 1 and Figure 2 As shown in the figure, this embodiment provides a multi-parameter intelligent early warning method for the risk of hypotension during hemodialysis, including an offline training phase and an online operation phase.

[0072] Offline training phase:

[0073] (1) Knowledge construction and parameter learning of causal Bayesian networks:

[0074] Clinical experts in nephrology and dialysis nursing participated in the construction of a causal directed acyclic graph topology based on clinical pathophysiology knowledge. For example... Figure 3 As shown, the directed acyclic graph contains the following nodes: ultrafiltration rate ( ), cumulative ultrafiltration volume to body weight ratio ( ), relative changes in blood volume ( Cardiac output (CO), peripheral vascular resistance (PVR), dialysate temperature deviation ( ), deviation in sodium concentration of dialysate ( ), autonomic nervous system function status ( ), pressure reflection gain ( ), systolic blood pressure (SBP), maximum Lyapunov index ( ), trajectory curvature abrupt change rate ( ), etiological types ( ).

[0075] The directed edge relationships of the three core causal links are as follows: First link Second link Third link and Dynamically enhanced observation nodes and Establish causal relationships with system instability and dynamic state transitions, respectively.

[0076] A historical dialysis record dataset (including confirmed hypotension events and their etiological annotations) was used for Bayesian parameter learning, and the expectation-maximization (EM) algorithm was employed to determine the conditional probability distribution of each node. The optimal prediction window statistics for each of the three etiological types were embedded into the network as prior knowledge.

[0077] (2) Offline pre-training of the PINN digital twin base model:

[0078] Pre-train a physical information neural network on a historical dataset. For example... Figure 4 As shown, the neural network adopts a dual-branch architecture: the data-driven branch consists of a 4-layer LSTM encoder (with 128, 64, 64, and 32 hidden units in each layer) plus a 2-layer fully connected decoder (with 64 and 1 neurons in each layer). The input is a multi-parameter time series matrix, and the output is the predicted future blood pressure value. The physical constraint branch embeds a residual calculation module for four physiological sub-equations, using automatic differentiation techniques to calculate the residuals of the differential equations. The constraint weights of the four sub-equations in the joint loss function are initialized equally (each to 0.25) during the pre-training phase, providing a basis for online etiological adaptive adjustment. The Adam optimizer is used with a learning rate of 1×10⁻⁶. -3 Train for 200 epochs.

[0079] (3) Historical statistics and threshold learning of phase space reconstruction parameters:

[0080] Phase space reconstruction was performed on dialysis records for each patient in the historical dataset. Statistical analysis was conducted to examine the differences in Lyapunov index distribution and trajectory curvature abrupt change characteristics between normal dialysis and the prodromal period of hypotension. A quantile regression method was used, with patient characteristic combinations (age, diabetes status, cardiac function classification, frequency of previous hypotension episodes, etc.) as independent variables and the Lyapunov index warning threshold as the dependent variable, to learn the threshold distribution corresponding to different patient characteristic combinations. On the validation set, this individualized threshold improved the F1 score by 12% to 18% compared to a uniform fixed threshold.

[0081] Online operation phase:

[0082] like Figure 2 and Figure 5 As shown, the online operation phase is executed in the order of steps S1 to S6, as follows:

[0083] Online Implementation of Step S1: After dialysis begins, the multi-source data acquisition module receives real-time data streams of blood pressure, blood volume, ultrafiltration parameters, and dialysate parameters. The raw sampling interval for blood pressure data is 4 to 5 minutes (intermittent cuff measurement), blood volume data is sampled every 10 seconds (continuous monitoring by an optical sensor), ultrafiltration parameters are updated every 30 seconds, and dialysate parameters are either fixed values ​​or updated every 5 minutes. An adaptive resampling engine unifies the data streams with different sampling rates to a time resolution of 30 seconds, using cubic spline interpolation for low-sampling-rate data and window mean downsampling for high-sampling-rate data. A motion artifact detector based on Db4 wavelets performs 3-level wavelet decomposition on the blood pressure signal. When the energy of high-frequency components exceeds the historical statistical threshold for each patient, it is identified as a motion artifact and marked. Adaptive Z-Score normalization uses the mean and standard deviation of the corresponding parameters in the patient's historical dialysis records as the normalization benchmark. Patient baseline features are also loaded.

[0084] Online implementation of step S2: The real-time acquired blood pressure time series is reconstructed into an m-dimensional phase space trajectory using Takens embedding, with the blood volume change rate as an additional embedding dimension. In this embodiment, the sliding window length is 5 minutes (containing 10 data points). The maximum Lyapunov exponent is calculated in real time within the window. (Using Rosenstein algorithm), trajectory curvature abrupt change rate and the rate of change of attractor dimension .when Exceeding the individualization threshold or When a sudden change occurs, a dynamic state transition signal DSTS is generated, and the system enters a three-layer linkage activation state. Before the three-layer linkage activation, the system is in a low-power monitoring mode, where only the perception layer performs real-time dynamic calculations, while the cognition and simulation layers are in standby mode to conserve computing resources.

[0085] Online implementation of step S3: The time series of kinetic indicators, current blood pressure trends, relative changes in blood volume, cumulative ultrafiltration load, dialysate parameter deviations, and patient baseline characteristics are input into a causal Bayesian network as observational evidence. A variational inference algorithm is used to calculate the posterior probabilities of the three etiological types in real time. The most probable etiological type and its corresponding optimal prediction window are then determined. and physical model configuration instructions For example, when the calculated values ​​are P(blood volume depletion type) = 0.72, P(vasodilation type) = 0.18, and P(neuroregulation disorder type) = 0.10, the cause is determined to be blood volume depletion type. minute, The weights for the Frank-Starling equation, capillary refill equation, vascular compliance equation, and pressure-reflex equation are 0.35, 0.1 ...15, 0.15,

[0086] Online implementation of step S4: Receiving physical model configuration instructions The constraint weights of the four physiological sub-equations in the PINN joint loss function are dynamically adjusted. Individual patient physiological parameters are rapidly calibrated online using data from the first 10 minutes after dialysis begins. , , Then, in the optimal prediction window... A forward simulation is performed to generate a prediction of the future blood pressure trajectory and its confidence interval. During the simulation, the calibrated physiological parameters are simultaneously fed back to step S2 via a forward feedback loop to update their individualized parameters.

[0087] Online implementation of step S5: The dual-loop feedback linkage controller performs a complete dual-loop feedback iteration every 60 seconds. The forward feedback loop transmits the physiological parameters calibrated in the simulation layer to the perception layer, based on... Value adjustment embedding dimension m ( (Increase the embedding dimension by 1 to 2 dimensions when the concentration is below 0.5 mL / mmHg), according to Value adjustment for Lyapunov exponent individualization threshold ( (The threshold is reduced by 15% to 25% when it is below 0.3 bpm / mmHg). The backward calibration loop transmits the updated kinetic parameters to the causal network to update the causal probability, and then transmits the updated causal configuration to the digital twin to adjust the physical constraints.

[0088] Online implementation of step S6: such as Figure 6 As shown, the system outputs multi-level early warnings based on the combined prediction results from the three layers. For example, when the system detects... When the individualized threshold is exceeded and the causal network outputs a 78% probability of blood volume depletion, and the digital twin predicts that blood pressure will drop from the current 108 mmHg to below 85 mmHg within the next 20 minutes, the system outputs a high-risk emergency alarm and generates the following interpretable report: The kinetic description is "The patient's hemodynamic stability has significantly decreased," the causal analysis is "The primary cause is determined to be blood volume depletion, as the accumulated ultrafiltration volume has reached 3.2% of body weight and blood volume continues to decline," the simulation prediction is "Blood pressure is predicted to drop below 85 mmHg within the next 20 minutes," and the intervention recommendation is "It is recommended to immediately reduce the ultrafiltration rate from the current 800 mL / h to 500 mL / h." This intervention recommendation was validated for dialysis adequacy before being output, confirming that reducing the ultrafiltration rate would still achieve more than 90% of the preset dehydration target within the remaining dialysis time.

[0089] Example 2: System Example

[0090] This embodiment provides a multi-parameter intelligent early warning system for the risk of hypotension during hemodialysis, such as... Figure 2 As shown, the system includes the following modules:

[0091] Multi-source data acquisition and preprocessing module: Connects to the data interface of the dialysis equipment to acquire multi-source monitoring data in real time. This module internally includes an adaptive resampling engine, a motion artifact detector based on wavelet decomposition, and an adaptive Z-Score normalizer based on historical distribution. After performing unified time axis alignment and quality control on the multi-source heterogeneous data, it forms a multi-parameter time-series matrix. For dialysis equipment that does not support direct data access, protocol conversion and format standardization can be performed through a data middleware layer.

[0092] Phase space dynamics sensing layer: Receives blood pressure time series and blood volume change rate data from a multi-parameter time series matrix, and reconstructs the phase space using the delayed embedding theorem. The sensing layer internally includes an automatic embedding parameter determiner, a dynamic index calculator, and a state transition detector.

[0093] Causal Inference and Etiology Cognition Layer: Receives dynamic indicators and real-time monitoring features as observational evidence, and calculates the posterior probabilities of three hypotension etiologies in real time using a pre-constructed causal Bayesian network. Outputs the most probable etiology type, the corresponding optimal prediction window, and physical model configuration instructions.

[0094] The physical information digital twin simulation layer receives physical model configuration instructions and internally includes a two-branch physical information neural network and an online calibration module for patient physiological parameters. It dynamically adjusts the constraint weights of the four hemodynamic sub-equations according to the configuration instructions, and performs forward simulation within the optimal prediction window after online calibration of individual patient physiological parameters.

[0095] Dual-loop feedback linkage controller: Connecting the simulation layer and the perception layer to form a forward feedback loop, and connecting the perception layer, the cognition layer, and the simulation layer to form a backward calibration loop. A complete dual-loop feedback iteration is performed every 60 seconds.

[0096] The clinical decision output module integrates the prediction results from the three layers and generates multi-level early warning information through a risk level assessor. Internally, it includes a triple interpretable evidence report generator, an intervention suggestion generator based on etiological subtyping, and a dialysis adequacy and safety checker. The system also includes an anomaly handling mechanism: when the data quality confidence level falls below a preset threshold, the system automatically downgrades to a simplified early warning mode based solely on the perception layer, avoiding misjudgments caused by low-quality data.

[0097] Example 3: Verification Experiment

[0098] To verify the technical effects of the present invention, a retrospective verification experiment was conducted at the hemodialysis center of a tertiary hospital.

[0099] Data source: 2400 dialysis records from 200 maintenance hemodialysis patients were collected, including 532 confirmed intradialysis hypotension events. According to clinical expert annotation, of the 532 hypotension events, 308 (58%) were volume depletion-related, 128 (24%) were vasodilatory, and 96 (18%) were neuromodulation-related. The data were divided into training and testing sets in a 7:3 ratio. Patients ranged in age from 32 to 82 years (mean age 61.3 ± 12.7 years), including 96 patients (48%) with diabetes and 64 patients (32%) with heart failure (NYHA class II or higher).

[0100] The comparison methods include: Method A is an LSTM prediction model with a fixed 15-minute window; Method B is an LSTM prediction model with a fixed 30-minute window; Method C is a causal-digital twin two-layer model without a phase space perception layer; Method D is a phase space-digital twin two-layer model without a causal cognition layer; and Method E is the three-layer linkage dual-loop feedback method of the present invention.

[0101] Experimental results are as follows Figure 7As shown, Method E achieved a composite AUC of 0.93, representing improvements of 14.8% and 17.7% compared to Method A (AUC=0.81) and Method B (AUC=0.79), respectively. Regarding etiological classification accuracy, Method E achieved classification accuracies of 87%, 82%, and 79% for the three etiological types, respectively. In terms of prediction lead time, Method E had an average prediction lead time of 26.3 minutes for volume depletion hypotension, 12.8 minutes for vasodilatory hypotension, and 17.1 minutes for neuromodulation hypotension, all within their respective optimal windows. Method E had a composite sensitivity of 89.2%, a specificity of 85.6%, and an F1 score of 0.87.

[0102] Ablation experiments showed that removing the phase-space perception layer (Method C) resulted in a decrease in AUC to 0.87 and a reduction in the mean precursor signal detection lead time of 8.5 minutes. Removing the causal cognition layer (Method D) resulted in a decrease in AUC to 0.85 and an inability to adaptively select the prediction window based on etiological type, with a particularly significant decrease in accuracy for vasodilatory hypotension, where the F1 score dropped from 0.83 to 0.67. Removing the double-loop feedback mechanism resulted in a decrease in AUC to 0.90 and a significant increase in the false alarm rate for patients with low vascular compliance, validating the necessity of physiological parameter feedback to adjust individualized thresholds.

[0103] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention. For example, the causal inference network is not limited to the causal Bayesian network, and other causal inference methods such as causal forests and structural causal models can also be used; the physiological constraint equations are not limited to the four specific equations in this embodiment, and other hemodynamic differential equations can also be used; the kinetic indices are not limited to the three specific indices in this embodiment, and other nonlinear kinetic indices can also be used.

Claims

1. A multi-parameter intelligent early warning method for the risk of hypotension during hemodialysis, characterized in that, Specifically, the following steps are included: S1. Real-time acquisition of multi-source monitoring data during dialysis, preprocessing of multi-source monitoring data to form multi-parameter time-series data with a unified time axis; S2. Perform phase space reconstruction on the hemodynamic time series in the multi-parameter time series data, calculate the dynamic index in real time in the reconstructed phase space, and generate a dynamic state transition signal when the dynamic index meets the state transition trigger condition. S3. Input the dynamic indicators and real-time monitoring features as observational evidence into the causal inference network, calculate the posterior probability of various hypotension etiologies, determine the most likely etiology and its corresponding optimal prediction window, and generate physical model configuration instructions. S4. Dynamically adjust the weight configuration of the physiological constraint equations embedded in the physical information neural network according to the physical model configuration instructions; use dialysis process monitoring data to calibrate individual patient physiological parameters online; perform forward simulation within the optimal prediction window to generate future blood pressure trajectory prediction and the probability of hypotension. S5. The individual physiological parameters calibrated online in step S4 are transmitted to step S2 through the forward feedback loop to dynamically adjust the embedding parameters of phase space reconstruction and the individualized detection threshold of kinetic indicators; the kinetic indicators updated in step S2 are transmitted to step S4 through the backward calibration loop after updating the etiology classification probability in step S3 to adjust the physical constraint weights. S6 integrates dynamic state transition signals, etiological classification results, and blood pressure trajectory prediction to output multi-level early warning information and targeted intervention suggestions based on etiological classification.

2. The multi-parameter intelligent early warning method for hypotension risk in hemodialysis according to claim 1, characterized in that, In step S1, the multi-source monitoring data includes continuous blood pressure monitoring data, online blood volume monitoring data, ultrafiltration rate and cumulative volume data, and dialysate parameter data; the preprocessing includes: using an adaptive resampling engine to unify data streams with different sampling rates to a preset time resolution, using a motion artifact detector based on wavelet decomposition to identify and mark abnormal data segments, using an adaptive Z-Score normalization method based on historical distribution to eliminate dimensional differences, forming a multi-parameter time series matrix and outputting a data quality confidence vector.

3. The intelligent early warning method for the risk of hypotension during hemodialysis according to claim 1, characterized in that, In step S2, the phase space reconstruction is based on Takens' delayed embedding theorem, and the blood volume change rate is included as an additional embedding dimension. The embedding dimension is automatically determined by the mutual information method, and the delay time is automatically optimized by the spurious nearest neighbor method. The dynamic indicators include the maximum Lyapunov exponent, the phase space trajectory curvature mutation rate, and the attractor dimension change rate. The state transition triggering condition is: the maximum Lyapunov exponent exceeds the individualization threshold, or the trajectory curvature mutation rate exceeds the preset mutation detection threshold.

4. The multi-parameter intelligent early warning method for hypotension risk in hemodialysis according to claim 1, characterized in that, In step S3, the causal inference network is a pre-constructed causal Bayesian network based on clinical pathophysiology knowledge. Its directed acyclic graph contains three core causal links: the first causal link is a blood volume depletion type link, where ultrafiltration rate points to decreased blood volume, decreased blood volume points to decreased cardiac output, and decreased cardiac output points to decreased blood pressure; the second causal link is a vasodilatory type link, where dialysate temperature deviation points to decreased peripheral vascular resistance, and decreased peripheral vascular resistance points to decreased blood pressure; the third causal link is a neuromodulation type link, where autonomic dysfunction points to weakened baroreflex, weakened baroreflex points to compensatory dysfunction, and compensatory dysfunction points to decreased blood pressure. The kinetic indicators are inserted into the directed acyclic graph as enhanced observational evidence nodes.

5. The intelligent early warning method for the risk of hypotension during hemodialysis according to claim 1, characterized in that, In step S3, the variational inference algorithm is used to calculate the posterior probability of each etiological type in real time; the optimal prediction window is determined according to the etiological type: a long window of 25 to 30 minutes corresponds to the blood volume depletion type, a short window of 10 to 15 minutes corresponds to the vasodilation type, and a medium window of 15 to 20 minutes corresponds to the neuromodulation disorder type; when the etiological classification probability changes dynamically during dialysis, the optimal prediction window is adjusted accordingly.

6. The intelligent early warning method for the risk of hypotension during hemodialysis according to claim 1, characterized in that, In step S4, the physiological constraint equations are a set of hemodynamic differential equations, including four sub-equations: the Frank-Starling cardiac output equation, the vascular compliance equation, the capillary refill equation, and the baroreflex regulation equation. The physical information neural network adopts a two-branch architecture. The first branch is a data-driven branch used to fit the measured monitoring data, and the second branch is a physical constraint branch used to calculate the differential equation residuals of the four sub-equations. The constraint weights of the four sub-equations in the joint loss function are determined by the weighted sum of the etiological classification probability vectors.

7. The intelligent early warning method for the risk of hypotension during hemodialysis according to claim 1, characterized in that, In step S4, the patient's individual physiological parameters calibrated online include vascular compliance coefficient, baroreflex gain, and capillary refill coefficient. In step S5, the adjustment rules for the forward feedback loop are as follows: the embedding dimension of the phase space reconstruction is adjusted according to the online calibrated vascular compliance coefficient, and the embedding dimension is increased when the vascular compliance coefficient is lower than a preset lower limit; the individualized threshold of the maximum Lyapunov index is adjusted according to the online calibrated baroreflex gain, and the individualized threshold is decreased when the baroreflex gain is lower than a preset lower limit; the iteration period of the dual-loop feedback is 30 seconds to 120 seconds.

8. The intelligent early warning method for the risk of hypotension during hemodialysis according to claim 1, characterized in that, In step S6, the multi-level early warning information includes three levels: low-risk reminder, medium-risk warning, and high-risk emergency alarm; each level of early warning outputs three interpretable pieces of evidence, which include a description of system stability changes at the dynamic level, an etiological attribution analysis at the causal level, and a blood pressure trajectory prediction at the simulation level. Targeted intervention recommendations are generated based on etiological classification results and patient contraindications, and undergo safety verification for dialysis adequacy targets before output.

9. A multi-parameter intelligent early warning system for the risk of hypotension during hemodialysis, characterized in that, A method for implementing a multi-parameter intelligent early warning system for hypotension risk in hemodialysis as described in any one of claims 1-8 includes: The multi-source data acquisition and preprocessing module is used to acquire multi-source monitoring data in real time during dialysis and preprocess the data to form multi-parameter time-series data with a unified time axis. The phase space dynamics sensing layer is used to reconstruct the phase space of hemodynamic time series, calculate dynamic indicators in real time, and generate dynamic state transition signals when the dynamic indicators meet the state transition triggering conditions. The causal inference etiology cognition layer is used to input dynamic indicators and real-time monitoring features as observational evidence into the causal inference network, calculate the posterior probability of various hypotension etiology types, and output the most likely etiology type, the corresponding optimal prediction window, and physical model configuration instructions. The physical information digital twin simulation layer is used to dynamically adjust the weight configuration of the physiological constraint equations in the physical information neural network according to the physical model configuration instructions. After online calibration of individual patient physiological parameters, it performs forward simulation within the optimal prediction window to generate blood pressure trajectory prediction. The dual-loop feedback linkage controller is used to transmit the individual physiological parameters calibrated online in the simulation layer to the perception layer through the forward feedback loop to adjust the embedded parameters and dynamic thresholds, and to transmit the updated dynamic indicators to the simulation layer through the cognitive layer via the backward calibration loop to adjust the physical constraint weights. The clinical decision output module integrates dynamic state transition signals, etiological classification results, and blood pressure trajectory prediction to output multi-level early warning information and targeted intervention suggestions based on etiological classification.

10. A multi-parameter intelligent early warning system for the risk of hypotension during hemodialysis according to claim 9, characterized in that, The causal inference etiology cognition layer contains a causal Bayesian network pre-constructed based on clinical pathophysiology knowledge. Its directed acyclic graph contains three causal links: blood volume depletion, vasodilation, and neural regulation disorder. Dynamic indicators are inserted into the directed acyclic graph as enhanced observational evidence nodes. The physiological constraint equations in the physical information digital twin simulation layer are a set of hemodynamic differential equations, including the Frank-Starling cardiac output equation, vascular compliance equation, capillary refill equation, and baroreflex regulation equation. The constraint weights of each sub-equation are determined by weighting the etiology classification probability vector. The iteration cycle of the dual-loop feedback linkage controller is 30 to 120 seconds.

Images