Intelligent linkage system for icu bedside dialysis vital signs adapted to nasal ventilation support

Abstract

The application belongs to the technical field of medical equipment and biosensor, and specifically discloses an ICU bedside dialysis vital sign intelligent linkage system suitable for nasal cavity ventilation support. The system comprises a multi-source data fusion acquisition module, a ventilation-dialysis coupling state evaluation module, a multi-modal physiological situation awareness module, a risk evolution prediction module and a hierarchical early warning and intervention suggestion generation module. The interactive state of ventilation and dialysis is quantitatively evaluated, and dynamic physiological feature fusion and time sequence risk prediction are carried out based on the interactive state, so that hierarchical early warning and targeted intervention suggestions are finally generated, and early and accurate early warning and clinical decision support for dialysis-related hypotension and hypoxemia are realized.

Description

Technical Field

[0001] This invention belongs to the field of medical equipment and biosensor technology, specifically relating to an intelligent linkage system for bedside dialysis vital signs in ICU patients that is adapted to support nasal ventilation. Background Technology

[0002] In the field of intensive care medicine, continuous renal replacement therapy is one of the life support technologies for critically ill patients with acute and chronic renal failure, severe infections, and multiple organ dysfunction syndrome. The implementation of bedside dialysis is closely related to the real-time monitoring of patients' vital signs and the timely warning of complications, constituting a complex clinical management task within the ICU.

[0003] For critically ill patients receiving bedside dialysis, especially those with respiratory failure requiring mechanical ventilation, real-time monitoring of vital signs and intelligent early warning of complications are crucial for ensuring treatment safety. This technological approach aims to achieve early identification and intervention of treatment-related complications through real-time collection and comprehensive analysis of multiple physiological parameters.

[0004] Current technologies typically rely on independent monitoring devices to monitor vital signs such as heart rate, blood pressure, and blood oxygen saturation, triggering alarms when any single indicator exceeds a preset threshold. During bedside dialysis, complications such as hypotension and hypoxemia are often multifactorial, gradual, and interconnected physiological processes. Existing early warning models suffer from the following problems: single-threshold alarms cannot capture the dynamic correlations and early trends of changes between vital sign parameters, resulting in low specificity and a high false alarm rate; this delayed alarm often triggers only when complications are already significant, missing the optimal intervention window. For patients receiving both dialysis and ventilation support, the interaction between hemodynamics and respiratory status is even more complex. Traditional, decentralized, and isolated monitoring and alarm methods struggle to integrate nasal ventilation support status with dialysis treatment parameters, making it impossible to construct an intelligent early warning model capable of predicting the risk of hypoxia during dialysis in ventilated patients. How to achieve intelligent linkage analysis and early warning of dialysis parameters, ventilation support status, and multi-dimensional vital signs has become an urgent technical challenge to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent linkage system for vital signs of bedside dialysis in ICU that is adapted to nasal ventilation support, so as to solve the technical contradictions in the prior art for critically ill patients receiving bedside dialysis and ventilation support. The vital signs monitoring and early warning modes are scattered and isolated, unable to capture the dynamic correlation of multiple parameters and early change trends, resulting in delayed warning of complications and low specificity.

[0006] To achieve the above objectives, the present invention provides an intelligent linkage system for vital signs in bedside dialysis in the ICU, adapted for nasal ventilation support, comprising: The multi-source data fusion acquisition module is used to acquire raw data streams from bedside dialysis equipment, mechanical ventilation equipment and multi-parameter vital signs monitoring equipment in real time and in parallel, and to perform millisecond-level time synchronization and data cleaning on all acquired heterogeneous data streams to form a time-aligned multidimensional physiological data matrix. The ventilation-dialysis coupling status assessment module is connected to the multi-source data fusion acquisition module and is used to quantitatively assess the interaction status between mechanical ventilation support and bedside dialysis treatment based on the multi-dimensional physiological data matrix. The multimodal physiological situation awareness module is connected to the multi-source data fusion acquisition module and the ventilation-dialysis coupling state assessment module. It is used to perform deep feature extraction and fusion analysis on vital sign data based on the current coupling state mode encoding to generate a high-dimensional physiological situation vector. The risk evolution prediction module is connected to the multimodal physiological situation awareness module and is used to predict the risk probability and evolution trajectory of dialysis-related hypotension or hypoxemia in ventilation patients within a future time window based on a continuous physiological situation vector sequence. The graded early warning and intervention suggestion generation module is connected to the risk evolution prediction module and the ventilation-dialysis coupling status assessment module. It is used to generate graded early warning signals and targeted intervention suggestions based on the predicted risk probability value, risk evolution trajectory and current coupling status pattern.

[0007] Preferably, the ventilation-dialysis coupling state assessment module constructs a quantitative model of the effect of ventilation on hemodynamics and a dynamic model of the effect of dialysis on blood gases and electrolytes, and performs coupling state calculation to generate a two-dimensional coupling state coordinate consisting of a first dimension characterizing the degree of inhibition or enhancement of circulation by ventilation and a second dimension characterizing the direction of metabolic environment changes caused by dialysis. The two-dimensional coupling state coordinate is then mapped to a preset coupling state matrix divided into 9 regions, and the corresponding coupling state pattern code is output.

[0008] Preferably, the multimodal physiological situation awareness module includes a waveform feature extraction submodule, a trend derivation calculation submodule, and a modality fusion submodule; The waveform feature extraction submodule performs real-time analysis on the electrocardiogram waveform and blood pressure waveform, and extracts feature parameters including time-domain and frequency-domain indices of heart rate variability, the systolic rise slope of the blood pressure waveform and the diastolic fall time constant. The trend-derived calculation submodule calculates the linear regression slope, standard deviation of the moving average, and approximate entropy over the past 5 minutes based on time-series data of blood oxygen saturation, blood pressure, and central venous pressure. The modal fusion submodule receives waveform features, trend-derived data, and current coupling state mode encoding, and performs nonlinear fusion through a 3-layer fully connected neural network. The weight parameters of the 3-layer fully connected neural network are dynamically switched according to different coupling state modes, and finally outputs a 128-dimensional physiological state vector.

[0009] Preferably, the risk evolution prediction module has a built-in dual-channel time series prediction network. The first channel is a causal convolutional network, which is responsible for capturing the short-term and local dependencies and rapid change patterns in the physiological state vector. The second channel is a gated recurrent unit network, which is responsible for modeling the long-term evolution trend and periodic characteristics of the physiological state. The outputs of the two channels are spliced ​​together in the time dimension and input into the attention mechanism layer, which dynamically assigns the importance weight of the physiological state at different historical moments to the current prediction. Finally, through a fully connected output layer, the risk evolution trajectory consisting of the risk probability values ​​for severe hypotension within the next 15 minutes, the risk probability values ​​for hypoxemia within the next 30 minutes, and the risk probability curves at multiple consecutive time points in the future is simultaneously output.

[0010] Preferably, the graded early warning and intervention suggestion generation module presets three risk probability thresholds, which correspond to three early warning levels: attention, warning, and crisis, respectively. When the predicted risk probability value is greater than the attention level threshold but less than the warning level threshold, a level 1 warning signal is generated and visual prompts and text pushes are provided. When the predicted risk probability value is greater than the warning level threshold, a level 2 warning signal is generated, triggering an audible and visual alarm. Based on the risk type and the current coupling state mode, a preliminary list of intervention suggestions is generated by matching from the pre-set intervention strategy knowledge base. When the predicted risk probability value is greater than the critical level threshold or the risk evolution trajectory shows that the probability rises sharply in a short period of time, a level 3 warning signal is generated, triggering the highest level audible and visual alarm, and the highest priority intervention recommendations are pushed to medical staff. At the same time, the warning information and relevant data snapshots are recorded.

[0011] Preferably, the ventilation-dialysis coupling status assessment module constructs a quantitative model of the effect of ventilation on hemodynamics, and the specific process is as follows: based on the collected tidal volume and positive end-expiratory pressure, combined with the preset patient chest and lung compliance empirical value, the mean intrathoracic pressure is estimated. The reduction factor for right atrial reflux pressure and the enhancement factor for left ventricular afterload are calculated based on the change in mean intrathoracic pressure relative to baseline. The reduction factor and the enhancement factor are weighted and summed, with the weights dynamically adjusted according to the current blood pressure and heart rate status. The net effect index is then calculated, where a positive net effect number indicates a net inhibitory effect and a negative net enhancement effect.

[0012] Preferably, the dynamic weight switching mechanism of the three-layer fully connected neural network in the multimodal physiological situation awareness module is as follows: the system pre-stores nine different network weight parameter files, which correspond to the nine coupling state regions divided by the ventilation-dialysis coupling state assessment module. When the ventilation-dialysis coupling state assessment module outputs the current state region code, the modality fusion submodule immediately loads the weight parameter file corresponding to the current state region code and replaces the existing weights in the neural network.

[0013] Preferably, the training process of the dual-channel time-series prediction network in the risk evolution prediction module adopts a multi-task learning framework; The network's training objective is to simultaneously minimize the prediction errors for both hypotension and hypoxemia risks. The training data comes from historically desensitized ICU multimodal monitoring data, with each training sample containing a 60-minute sequence of physiological state vectors and subsequent labels for actual hypotension and hypoxemia events. The network's parameters are optimized using backpropagation and adaptive moment estimation algorithms.

[0014] Preferably, the intervention strategy knowledge base in the graded early warning and intervention suggestion generation module is constructed based on clinical guidelines, expert consensus, and machine learning data mining; each intervention item is associated with multiple feature tags, including the type of risk targeted, the applicable coupling state pattern, the urgency level of the intervention, and the expected physiological parameter response target; the graded early warning and intervention suggestion generation module achieves the matching and ranking of intervention measures by calculating the cosine similarity between the feature vector of the current risk scenario and the feature tags of the items in the knowledge base.

[0015] Preferably, the system further includes a feedback learning optimization module, which connects the hierarchical early warning and intervention suggestion generation module and the risk evolution prediction module; The feedback learning optimization module is used to collect the actual intervention operation records of medical staff after each warning and the patient's subsequent physiological response data, and calculate the performance index of the prediction model by comparing the predicted risk with the actual outcome. If a persistent deviation is found in the prediction of the coupled state mode, an incremental learning process is triggered for the corresponding sub-model in the risk evolution prediction module, and the model parameters are fine-tuned using new feedback data.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by constructing a ventilation-dialysis coupling state assessment module, is the first to quantify and pattern-identify the interaction effects of mechanical ventilation and bedside dialysis in real time. The two-dimensional coupling state coordinates and 9-region matrix calculated by this module provide the system with precise context for understanding the current complex treatment background. This allows subsequent physiological state analysis and risk prediction to move beyond judgments based on isolated parameters, instead building upon a higher-dimensional, more clinically relevant information foundation of the "treatment interaction state," enhancing the system's understanding of the context of complication occurrence and providing a crucial prior framework for early warning. 2. The multimodal physiological state perception module designed in this invention employs a dynamically weighted neural network architecture, achieving adaptive matching between the analysis model and the current coupling state. This design enables the system to flexibly capture the specific correlations and changing patterns between vital sign parameters under different treatment interaction modes. The context-aware feature fusion method significantly enhances the information density and clinical relevance of the generated 128-dimensional physiological state vector, laying a solid data foundation for prediction. 3. The risk evolution prediction module of this invention employs a dual-channel temporal network combined with an attention mechanism, enabling simultaneous modeling of short-term abrupt changes and long-term evolution of physiological states, focusing on the most critical historical moments. This design allows the system to not only output the static probability of future hypotension or hypoxemia but also depict the dynamic trajectory of risk evolution, thereby identifying acceleration points of risk. 4. The graded early warning and intervention suggestion generation module of this invention achieves a closed loop from risk prediction to clinical decision support. The graded early warning mechanism dynamically adjusts the alarm intensity based on risk probability and evolution speed, reducing interfering false alarms caused by fluctuations in single parameters. The function of generating intervention suggestions based on intelligent matching of coupled state patterns and risk types directly transforms early warning information into preliminary, targeted action clues, shortening the decision-making time from risk discovery to initial intervention, and improving the efficiency and quality of clinical response. 5. The system's built-in feedback learning optimization module constructs a continuous improvement cycle. By using data from actual clinical interventions and patient feedback to incrementally learn the prediction model, the system can continuously adapt to the specificity of individual patients and new knowledge in clinical practice, making the early warning model more accurate and reliable over time, possessing long-term evolutionary capabilities. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the ventilation-dialysis coupling status assessment module in this invention; Figure 3 This is a logical flowchart of the multimodal physiological situational awareness module in this invention; Figure 4This is a schematic diagram of the multi-level interaction relationship and data flow of the risk evolution prediction module in this invention; Figure 5 This is a schematic diagram of the core principle framework of the graded early warning and intervention suggestion generation module in this invention. Detailed Implementation

[0018] Example 1: The overall technical architecture of the intelligent linkage system for bedside dialysis vital signs in the ICU, adapted to nasal ventilation support, proposed in this invention, is shown in the attached figure. Figure 1 As shown in the figure. Based on multi-source heterogeneous data fusion, the system constructs four functional modules: ventilation-dialysis coupled status assessment, multimodal physiological situational awareness, risk evolution prediction, and generation of graded early warning and intervention suggestions. This enables real-time modeling, dynamic risk prediction, and closed-loop decision support for the high-dimensional physiological status of critically ill patients receiving dual support of mechanical ventilation and bedside continuous renal replacement therapy (CRRT). The following will combine the attached... Figure 1 To be continued Figure 5 The specific implementation methods of each component of the system are described.

[0019] The system includes a multi-source data fusion acquisition module. This module serves as the data entry point for the entire system, acquiring raw physiological and therapeutic parameter streams in parallel and in real-time from three key medical devices. This module connects to the bedside dialysis equipment, mechanical ventilation equipment, and multi-parameter vital signs monitoring equipment via standard medical device communication interfaces. From the bedside dialysis equipment, the module continuously collects data on dialysate flow rate, ultrafiltration rate, transmembrane pressure, sodium ion concentration, potassium ion concentration, calcium ion concentration in the replacement fluid, and the cumulative time since the start of treatment, with a sampling frequency of at least once per second.

[0020] From the mechanical ventilation equipment, the multi-source data fusion acquisition module simultaneously acquires the current ventilation mode, tidal volume, respiratory rate, inhaled oxygen concentration, positive end-expiratory pressure, and real-time leakage of the nasal mask. The leakage is calculated by the built-in flow sensor of the ventilation equipment, in liters per minute, and the sampling frequency is also once per second.

[0021] The multi-source data fusion acquisition module acquires high-resolution electrocardiogram waveforms, invasive or non-invasive arterial blood pressure waveforms, blood oxygen saturation waveforms, central venous pressure values, and body temperature values ​​from the multi-parameter vital signs monitoring device.

[0022] Upon entering the system, all the aforementioned data streams are immediately fed into a unified timestamp alignment unit. This unit uses the system's internal high-precision real-time clock as a reference, appending millisecond-level timestamps to data packets from different devices. It also compensates for missing or misaligned data points using linear interpolation or spline interpolation methods, ensuring that all parameters are strictly aligned on the time axis. Subsequently, the data cleaning subunit performs outlier detection, employing a sliding window-based 3x standard deviation criterion or an isolated forest algorithm to identify and remove noise points that significantly deviate from the physiologically reasonable range. The final output is a structured, time-aligned multidimensional physiological data matrix, with each row representing a time point and each column representing a collected parameter, providing high-quality input for subsequent modules.

[0023] Please refer to the attached document. Figure 2 The system further includes a ventilation-dialysis coupling status assessment module. This module directly receives a time-aligned multidimensional physiological data matrix from the multi-source data fusion acquisition module and focuses on quantifying the interaction between mechanical ventilation and bedside dialysis. The module first performs quantitative modeling of the hemodynamic impact of ventilation. During this process, the system estimates the mean intrathoracic pressure using an empirical formula based on the currently acquired tidal volume and positive end-expiratory pressure (PEEP), combined with the patient's individualized chest and lung compliance empirical values. The estimation logic is as follows: the mean intrathoracic pressure is estimated by adding the ratio of tidal volume to chest and lung compliance multiplied by 0.5 to the PEEP.

[0024] This estimated change in intrathoracic pressure relative to the patient's baseline is used to calculate two key coefficients: the right atrial reflux pressure reduction coefficient and the left ventricular afterload enhancement coefficient. The former reflects the degree to which increased intrathoracic pressure impedes venous return, while the latter reflects the degree to which it increases resistance to left ventricular ejection. These two coefficients are obtained through a pre-defined nonlinear function mapping. The system dynamically adjusts the weighting of the two coefficients based on the current systolic blood pressure and heart rate. If the current systolic blood pressure is less than 90 mmHg and the heart rate is greater than 120 beats / minute, the enhancement coefficient is given a higher weight; otherwise, the reduction coefficient is given a higher weight. The net effect index is calculated by multiplying each coefficient by its corresponding weight, summing the results, and then subtracting 1. A positive net effect index indicates a net inhibitory effect of ventilation on circulation, while a negative value indicates a net enhancing effect.

[0025] Simultaneously, the ventilation-dialysis coupling status assessment module performs dynamic modeling of the effects of dialysis on blood gases and electrolytes. Inputs include the current ultrafiltration rate, the concentrations of major ions (Na⁺, K⁺, Ca²⁺) in the replacement fluid, and the cumulative treatment time since the start of dialysis. The model simulates the dynamic changes in plasma crystalloid osmolality using a first-order differential equation; the rate of change is directly proportional to the ultrafiltration rate and inversely proportional to the difference between the replacement fluid osmolality and the initial plasma osmolality. Based on the buffer base concentration of the replacement fluid and the patient's current partial pressure of carbon dioxide, the model estimates the trend of acid-base imbalance and outputs a standardized acid-base shift index, with positive values ​​indicating a tendency towards alkalosis and negative values ​​indicating a tendency towards acidosis.

[0026] After completing the above calculations, the coupling state is resolved. The aforementioned net effect index is used as the first dimension, and the acid-base shift index as the second dimension, forming a two-dimensional coupling state coordinate system. This two-dimensional coupling state coordinate system is mapped to a preset 3×3 grid-like coupling state matrix. This 3×3 grid-like coupling state matrix is ​​divided into low, medium, and high levels in both dimensions, forming a total of nine regions. Each region is assigned a unique code and corresponds to a ventilation-dialysis interaction dominant mode. This coupling state code serves as context information and is transmitted to the downstream module in real time.

[0027] Please refer to the attached document. Figure 3 The system includes a multimodal physiological situation awareness module. This module receives raw waveforms and time-series data from a multi-source data fusion acquisition module, as well as the current coupling state code from a ventilation-dialysis coupling state assessment module, aiming to generate a high-dimensional vector that comprehensively and precisely characterizes the patient's current physiological homeostasis. This module contains three sequentially connected sub-modules. The waveform feature extraction sub-module performs real-time signal processing on high-sampling-rate ECG and arterial blood pressure waveforms. For ECG waveforms, it calculates time-domain and frequency-domain indices. For arterial blood pressure waveforms, it calculates the systolic rise slope, diastolic fall time constant, and pulse pressure variability. This sub-module outputs no fewer than 15 feature parameters. The trend derivation calculation sub-module focuses on the time-series data of blood oxygen saturation, arterial blood pressure, and central venous pressure. It uses a 5-minute sliding window to calculate the linear regression slope, standard deviation of the moving average, and approximate entropy for each parameter. These derived indices collectively characterize the dynamic stability of the physiological parameters.

[0028] The modal fusion submodule receives all the aforementioned features and the current coupling state encoding. This submodule comprises a three-layer fully connected neural network, with the number of input layer nodes equal to the total number of features, 64 hidden layers, and a fixed 128-dimensional output layer. The weight parameters of this three-layer fully connected neural network are not fixed but dynamically switched according to the current coupling state encoding. During system initialization, nine different weight parameter files are pre-stored through offline training, each corresponding to a region in the coupling state matrix. When a new coupling state encoding is received, the multimodal physiological situation awareness module immediately loads the corresponding weight file from memory, replacing the existing parameters in the network. This mechanism ensures that the feature fusion process can adaptively match the specific physiological association patterns in the current therapeutic interaction context. The multimodal physiological situation awareness module outputs a 128-dimensional physiological situation vector, which is cached in a circular buffer for subsequent temporal analysis.

[0029] Please refer to the attached document. Figure 4 The system includes a risk evolution prediction module. This module reads a continuous historical physiological situation vector sequence from the circular buffer of the multimodal physiological situation awareness module, typically spanning the past 30 minutes and corresponding to 1800 time points. Its goal is to predict the risk of two key complications occurring within future time windows: dialysis-related hypoxia and hypoxemia in ventilation patients. To achieve this, the risk evolution prediction module incorporates a dual-channel temporal prediction network. The first channel is a causal convolutional network, consisting of multiple stacked dilated causal convolutional layers with dilation factors increasing exponentially by powers of 2 to expand the receptive field. This first channel excels at capturing local abrupt changes in physiological situation vectors over short timescales, such as sudden changes in blood pressure waveform characteristics. The second channel is a gated recurrent unit network, containing two stacked layers of gated recurrent units, each with 128 units. This second channel excels at modeling periodic or gradual evolutionary trends over long timescales, such as the cumulative effect of insufficient capacity due to slow ultrafiltration. The two channels process the input sequence independently and output a feature vector at each time step.

[0030] These two feature vectors are concatenated along the time dimension to form a composite feature sequence. This composite feature sequence is then fed into a multi-head attention mechanism layer. This layer learns the query, key, and value matrix to dynamically calculate the importance weight of each historical moment for the current prediction. For example, when predicting hypotension, if a rapid drop in central venous pressure has recently occurred, the weight of that time period will be increased. The output of the attention layer then passes through a fully connected layer, ultimately outputting two scalar values ​​in parallel: the probability of developing severe hypotension within the next 15 minutes and the probability of developing hypoxemia within the next 30 minutes. The risk evolution prediction module also generates a continuous risk evolution trajectory curve by repeating the prediction process at multiple future time points, visually demonstrating the trend of risk changes over time.

[0031] Please refer to the attached document. Figure 5 The system includes a tiered early warning and intervention suggestion generation module. This module receives two risk probability values ​​and a risk evolution trajectory from the risk evolution prediction module, and the current coupling state code from the ventilation-dialysis coupling state assessment module. The module internally pre-sets two independent threshold systems, corresponding to hypotension and hypoxemia respectively. Each system contains three thresholds: a concern level threshold, a warning level threshold, and a critical level threshold. The module determines whether any risk probability is greater than its corresponding concern level threshold. If it is greater than the concern level but not the warning level, a level one warning is triggered. At this time, the system marks the patient's bed with a yellow highlighted border on the graphical user interface of the medical workstation and pushes a text prompt in the message center, containing the current main abnormal parameters and a description of the coupling state. If any risk probability is greater than the warning level threshold, a level two warning is triggered. The system activates the audible and visual alarm of the bedside device and pops up a semi-transparent warning window on the workstation interface. The tiered early warning and intervention suggestion generation module starts the intervention suggestion generation engine. The engine first determines the dominant risk type based on which risk probability is higher. It accesses a structured intervention strategy knowledge base. Each intervention measure in this knowledge base is stored as a structured entry, containing fields such as: measure text, applicable risk type (hypotension / hypoxemia), applicable coupling state pattern list, urgency level, and expected response goal. The engine encodes the current scenario (risk type + coupling state encoding) into a query vector and calculates the cosine similarity between the vector and the feature label vectors of all entries in the knowledge base. It selects the 3 to 5 measures with the highest similarity, sorts them by urgency and similarity, and generates a preliminary list of intervention recommendations, displayed in the warning window. If any risk probability exceeds the critical level threshold, or if the risk evolution trajectory shows a probability increase of more than 0.4 within 5 minutes, a level 3 warning is triggered. At this time, the audible and visual alarm is upgraded to the highest intensity, a full-screen alarm is forcibly displayed on the workstation interface, and the 1 to 2 highest priority intervention recommendations are displayed at the top in bold red font. Simultaneously, the system automatically packages this warning event, relevant physiological data snapshots, and the generated intervention recommendations, encrypts them, and writes them to the security event log for post-event review and auditing.

[0032] As a further optimization of the invention, the system also includes a feedback learning optimization module. This module runs continuously in the background, monitoring the medical order execution records and vital sign database in the electronic medical record system. Whenever the graded warning and intervention suggestion generation module triggers a level two or three warning, the feedback learning optimization module tracks two key data points within the following 30 minutes: first, whether medical staff have implemented the intervention measures recommended by the system; and second, the patient's actual physiological outcome. By comparing the predicted risk with the actual outcome, the feedback learning optimization module calculates the performance indicators of the prediction model for the current coupled state mode, such as precision, recall, and calibration.

[0033] If a consistently high false positive or false negative rate is detected for hypotension prediction within a specific coupling state region, the module automatically triggers an incremental learning process. It adds recently collected samples with true labels to the training set and fine-tunes only the sub-model corresponding to that coupling state region. The fine-tuning process employs an adaptive moment estimation algorithm with a small learning rate and strictly limits the number of iterations to prevent catastrophic forgetting. The updated model parameters are then safely deployed to the production environment, thus achieving a closed-loop and continuous evolution of the system's early warning capabilities.

[0034] In summary, this embodiment, through the close collaboration of the five modules mentioned above, constructs a complete technical system that deeply understands the treatment interaction context, accurately perceives subtle physiological disturbances, proactively predicts risk evolution trajectories, and intelligently generates intervention cues. This system not only solves the problem of isolated monitoring data in existing technologies but also, by introducing the high-dimensional context of coupled states, achieves a fundamental leap from passive alarms to proactive early warnings, and from general suggestions to contextualized decision support, providing unprecedented safety assurance for critically ill patients receiving dual life support in the ICU.

[0035] Example 2: Building upon Example 1, this example further expands the waveform feature extraction submodule within the multimodal physiological situational awareness module to enhance its ability to analyze cardiopulmonary interaction effects. The submodule not only processes ECG and arterial blood pressure waveforms but also incorporates synchronous analysis of respiratory flow waveforms and establishes cross-modal phase relationship features among the three.

[0036] The system additionally acquires high-resolution respiratory flow waveforms from the mechanical ventilation equipment, with a sampling rate of at least 100 Hz. The waveform feature extraction submodule preprocesses the respiratory flow waveform, removing baseline drift and high-frequency noise through bandpass filtering. It then performs respiratory cycle segmentation, accurately identifying the start and end times of each inspiration and expiration using flow zero-crossing detection. Based on this, the submodule calculates multiple fine features within the respiratory cycle, including peak inspiratory flow, peak expiratory flow, inspiratory time, expiratory time, the proportion of inspiratory time to the respiratory cycle, and the area of ​​the flow-volume loop. The submodule also performs heart-respiratory synchronicity analysis. It performs cross-correlation analysis between the ECG R-wave peak time sequence and the respiratory cycle start time sequence to calculate the amplitude of heart rate changes with the respiratory cycle, i.e., the respiratory sinus arrhythmia index. Simultaneously, the submodule performs beat-by-beat analysis of the arterial blood pressure waveform, calculating systolic blood pressure, diastolic blood pressure, and pulse pressure for each cardiac cycle, and classifies these blood pressure parameters according to their respiratory phase (inspiratory or expiratory). The waveform feature extraction submodule calculates pulse pressure variability, i.e., (mean pulse pressure during inspiration - mean pulse pressure during expiration) / mean pulse pressure during expiration. This indicator is a classic parameter for assessing a patient's volume responsiveness and has significant clinical value in the context of mechanical ventilation. The waveform feature extraction submodule also calculates the respiratory phase dependence of heart rate variability, i.e., the ratio of the spectral power of heart rate variability during inspiration to that during expiration.

[0037] These newly added respiratory-related features, along with the existing ECG and blood pressure features, are fed into the modality fusion submodule. Since these features directly reflect the immediate modulatory effect of ventilation on circulation, they have extremely high discriminative value in coupled states such as "high positive end-expiratory pressure" or "high air leakage." When the ventilation-dialysis coupling state assessment module determines that such a state is currently in place, the dynamic weight switching mechanism activates the corresponding weight file. This file has been specifically enhanced during the training phase to be sensitive to features such as pulse pressure variability and the respiratory sinus arrhythmia index. The final output 128-dimensional physiological state vector can capture the precursors of circulatory instability caused by improper ventilation settings or insufficient volume earlier and more accurately, thus providing a more forward-looking input to the risk evolution prediction module and further improving the system's timeliness and specificity in warning of dialysis-related hypotension. This expansion scheme, without changing the system's main architecture, achieves a refined leap in perception dimensions through in-depth mining of existing data sources.

[0038] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent linkage system for vital signs in ICU bedside dialysis, adapted for nasal ventilation support, characterized in that: include: The multi-source data fusion acquisition module is used to acquire raw data streams from bedside dialysis equipment, mechanical ventilation equipment and multi-parameter vital signs monitoring equipment in real time and in parallel, and to perform millisecond-level time synchronization and data cleaning on all acquired heterogeneous data streams to form a time-aligned multidimensional physiological data matrix. The ventilation-dialysis coupling status assessment module is connected to the multi-source data fusion acquisition module and is used to quantitatively assess the interaction status between mechanical ventilation support and bedside dialysis treatment based on the multi-dimensional physiological data matrix. The multimodal physiological situation awareness module is connected to the multi-source data fusion acquisition module and the ventilation-dialysis coupling state assessment module. It is used to perform deep feature extraction and fusion analysis on vital sign data based on the current coupling state mode encoding to generate a high-dimensional physiological situation vector. The risk evolution prediction module is connected to the multimodal physiological situation awareness module and is used to predict the risk probability and evolution trajectory of dialysis-related hypotension or hypoxemia in ventilation patients within a future time window based on a continuous physiological situation vector sequence. The graded early warning and intervention suggestion generation module is connected to the risk evolution prediction module and the ventilation-dialysis coupling status assessment module. It is used to generate graded early warning signals and targeted intervention suggestions based on the predicted risk probability value, risk evolution trajectory and current coupling status pattern.

2. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support as described in claim 1, characterized in that, The ventilation-dialysis coupling state assessment module constructs a quantitative model of the effect of ventilation on hemodynamics and a dynamic model of the effect of dialysis on blood gases and electrolytes. It then performs coupling state calculations to generate a two-dimensional coupling state coordinate system consisting of a first dimension characterizing the degree of inhibition or enhancement of circulation by ventilation and a second dimension characterizing the direction of changes in the metabolic environment caused by dialysis. This two-dimensional coupling state coordinate system is then mapped to a preset coupling state matrix divided into 9 regions, and the corresponding coupling state pattern code is output.

3. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support as described in claim 2, characterized in that, The multimodal physiological situation awareness module includes a waveform feature extraction submodule, a trend derivation calculation submodule, and a modality fusion submodule; The waveform feature extraction submodule performs real-time analysis on the electrocardiogram waveform and blood pressure waveform, and extracts feature parameters including time-domain and frequency-domain indices of heart rate variability, the systolic rise slope of the blood pressure waveform and the diastolic fall time constant. The trend-derived calculation submodule calculates the linear regression slope, standard deviation of the moving average, and approximate entropy over the past 5 minutes based on time-series data of blood oxygen saturation, blood pressure, and central venous pressure. The modal fusion submodule receives waveform features, trend-derived data, and current coupling state mode encoding, and performs nonlinear fusion through a 3-layer fully connected neural network. The weight parameters of the 3-layer fully connected neural network are dynamically switched according to different coupling state modes, and finally outputs a 128-dimensional physiological state vector.

4. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support as described in claim 3, characterized in that, The risk evolution prediction module has a built-in dual-channel time series prediction network. The first channel is a causal convolutional network, which is responsible for capturing the short-term and local dependencies and rapid change patterns in the physiological state vector. The second channel is a gated recurrent unit network, which is responsible for modeling the long-term evolution trend and periodic characteristics of the physiological state. The outputs of the two channels are spliced ​​together in the time dimension and input into the attention mechanism layer, which dynamically assigns the importance weight of the physiological state at different historical moments to the current prediction. Finally, through a fully connected output layer, the risk evolution trajectory consisting of the risk probability values ​​for severe hypotension within the next 15 minutes, the risk probability values ​​for hypoxemia within the next 30 minutes, and the risk probability curves at multiple consecutive time points in the future is simultaneously output.

5. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support as described in claim 4, characterized in that, The graded early warning and intervention suggestion generation module presets three risk probability thresholds, which correspond to three early warning levels: attention, warning, and critical. When the predicted risk probability value is greater than the attention level threshold but less than the warning level threshold, a level 1 warning signal is generated and visual prompts and text pushes are provided. When the predicted risk probability value is greater than the warning level threshold, a level 2 warning signal is generated, triggering an audible and visual alarm. Based on the risk type and the current coupling state mode, a preliminary list of intervention suggestions is generated by matching from the pre-set intervention strategy knowledge base. When the predicted risk probability value is greater than the critical level threshold or the risk evolution trajectory shows that the probability rises sharply in a short period of time, a level 3 warning signal is generated, triggering the highest level audible and visual alarm, and the highest priority intervention recommendations are pushed to medical staff. At the same time, the warning information and relevant data snapshots are recorded.

6. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support according to claim 5, characterized in that, The ventilation-dialysis coupling status assessment module constructs a quantitative model of the impact of ventilation on hemodynamics. The specific process is as follows: based on the collected tidal volume and positive end-expiratory pressure, combined with the preset patient chest and lung compliance empirical value, the mean intrathoracic pressure is estimated. The reduction factor for right atrial reflux pressure and the enhancement factor for left ventricular afterload are calculated based on the change in mean intrathoracic pressure relative to baseline. The reduction factor and the enhancement factor are weighted and summed, with the weights dynamically adjusted according to the current blood pressure and heart rate status. The net effect index is then calculated. A positive net effect index indicates a net inhibitory effect, while a negative net effect index indicates a net enhancing effect.

7. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support as described in claim 6, characterized in that, The dynamic weight switching mechanism of the three-layer fully connected neural network in the multimodal physiological situation awareness module is as follows: the system pre-stores nine different network weight parameter files, which correspond to the nine coupling state regions divided by the ventilation-dialysis coupling state assessment module. When the ventilation-dialysis coupling state assessment module outputs the current state region code, the modality fusion submodule immediately loads the weight parameter file corresponding to the current state region code and replaces the existing weights in the neural network.

8. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support as described in claim 7, characterized in that, The training process of the dual-channel time-series prediction network in the risk evolution prediction module adopts a multi-task learning framework. The training objective of the network is to simultaneously minimize the prediction error of hypotension risk and the prediction error of hypoxemia risk; The training data comes from historical anonymized ICU multimodal monitoring data. Each training sample contains a 60-minute sequence of physiological state vectors and labels of subsequent actual hypotension and hypoxemia events. The network is optimized for parameters using backpropagation and adaptive moment estimation algorithms.

9. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support as described in claim 8, characterized in that, The intervention strategy knowledge base in the graded early warning and intervention suggestion generation module is constructed based on clinical guidelines, expert consensus, and machine learning data mining. Each intervention item is associated with multiple feature tags, including the type of risk targeted, the applicable coupling state pattern, the urgency level of the intervention, and the expected physiological parameter response target. The graded early warning and intervention suggestion generation module achieves the matching and ranking of intervention measures by calculating the cosine similarity between the feature vector of the current risk scenario and the feature tags of the items in the knowledge base.

10. The intelligent linkage system for ICU bedside dialysis vital signs adapted to nasal ventilation support according to claim 9, characterized in that, The system also includes a feedback learning optimization module, which connects the hierarchical early warning and intervention suggestion generation module with the risk evolution prediction module; The feedback learning optimization module is used to collect the actual intervention operation records of medical staff after each warning and the patient's subsequent physiological response data, and calculate the performance index of the prediction model by comparing the predicted risk with the actual outcome. If a persistent deviation is found in the prediction of the coupled state mode, an incremental learning process is triggered for the corresponding sub-model in the risk evolution prediction module, and the model parameters are fine-tuned using new feedback data.

Images