Hierarchical warning system based on multi-parameter vital sign detection

Abstract

The application relates to the technical field of medical early warning, and discloses a hierarchical early warning system based on multi-parameter vital sign detection. The system comprises the following steps: collecting real-time physiological parameters such as heart rate, blood pressure, blood oxygen saturation and body temperature of a patient through a vital sign monitoring device; inputting the parameters into a feature extraction network to generate a multi-dimensional physiological feature vector, and constructing a dynamic risk assessment matrix containing physiological state change trends of different time windows according to the physiological feature vector. The risk level interval is divided according to a preset hierarchical early warning threshold, an adaptive weight distribution strategy is adopted to adjust the interval and generate a comprehensive risk score; when the score exceeds a preset early warning trigger line, the corresponding early warning response mechanism is activated. The system can realize comprehensive dynamic assessment of the physiological state of the patient, is suitable for emergency, intensive care and chronic disease care scenes, and meets the clinical health risk monitoring and early warning requirements.

Description

Technical Field

[0001] This invention relates to the field of medical early warning technology, specifically a graded early warning system based on multi-parameter vital sign detection. Background Technology

[0002] In modern healthcare systems, real-time monitoring of patients' vital signs and risk warning are crucial components of clinical diagnosis and treatment. This is especially true in emergency care, intensive care, and home care for chronic diseases, where timely and accurate vital sign assessment directly impacts the timing of treatment and the protection of patients' health and safety. Currently, most vital sign monitoring methods used in clinical practice rely on the independent monitoring of single or a few physiological parameters. For example, heart rate data is obtained through a standard heart rate monitor, or blood pressure is measured periodically using an electronic blood pressure monitor. These methods often only capture the state of a single physiological indicator at a specific moment and cannot comprehensively reflect the overall trend of changes in the patient's physiological functions.

[0003] Existing early warning systems often rely on fixed thresholds for risk assessment, triggering an alert when a physiological parameter exceeds a preset value. This fixed-threshold approach has significant limitations. Due to individual differences in age, underlying diseases, and physical condition, the same physiological parameter value may represent entirely different levels of health risk for different patients. For example, a heart rate of around 90 beats per minute may be within the normal range for a healthy adult, but for an elderly patient with severe heart disease, the same heart rate may indicate a high health risk. Furthermore, existing early warning systems lack effective weighting strategies when comprehensively analyzing multiple parameters, failing to dynamically adjust based on the varying importance of different physiological parameters under specific conditions, resulting in insufficient accuracy and specificity in risk assessment results.

[0004] In clinical practice, this traditional monitoring and early warning method often leads to two situations: for some potential, slowly developing health risks, the system cannot capture risk signals in time because the change in a single parameter does not reach the fixed early warning threshold, thus delaying the best intervention opportunity; for some physiological parameter fluctuations caused by individual differences, the system may misjudge them as risk events, triggering unnecessary early warning responses, which not only increases the workload of medical staff but may also cause unnecessary panic among patients and their families. With the continuous development of medical technology and the increasing demands for the quality of medical services, the existing vital sign monitoring and early warning methods based on single parameters and fixed thresholds are no longer sufficient to meet the clinical needs for accurate, dynamic, and comprehensive assessment of patients' health risks. A new system that can integrate multi-parameter information and achieve dynamic risk assessment and adaptive early warning is needed. Summary of the Invention

[0005] The purpose of this invention is to provide a graded early warning system based on multi-parameter vital sign detection to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a graded early warning system based on multi-parameter vital sign detection, the system comprising:

[0007] The patient's real-time physiological parameters are collected through vital sign monitoring equipment, including heart rate, blood pressure, blood oxygen saturation, and body temperature.

[0008] The real-time physiological parameters are input into a feature extraction network to generate a multi-dimensional physiological feature vector.

[0009] A dynamic risk assessment matrix is ​​constructed based on the multi-dimensional physiological feature vectors, and the dynamic risk assessment matrix contains the trend of physiological state changes under different time windows.

[0010] The dynamic risk assessment matrix is ​​divided into regions based on preset graded early warning thresholds to obtain multiple risk level intervals;

[0011] An adaptive weighting strategy is used to dynamically adjust the risk level intervals to generate a comprehensive risk score for the current patient.

[0012] When the comprehensive risk score exceeds the preset warning trigger line, the corresponding warning response mechanism is activated.

[0013] Preferably, the feature extraction network includes:

[0014] The real-time physiological parameters are subjected to time series standardization to eliminate noise interference during the acquisition process;

[0015] Local temporal features of the real-time physiological parameters are extracted using a convolutional neural network;

[0016] The local temporal features are fused with the global statistical features to generate the multi-dimensional physiological feature vector.

[0017] Preferably, the process of constructing the dynamic risk assessment matrix includes:

[0018] The deviation of physiological state is calculated based on the rate of change of the multi-dimensional physiological feature vectors.

[0019] By combining historical physiological data, the trend of physiological state changes under the different time windows is generated;

[0020] The deviation of the physiological state is correlated with the trend of physiological state change, and then the dynamic risk assessment matrix is ​​filled.

[0021] Preferably, the process of setting the graded early warning threshold includes:

[0022] The normal fluctuation range of different physiological parameters is determined based on clinical data statistics;

[0023] Based on the normal fluctuation range, the risk levels are divided into three categories: low risk, medium risk, and high risk.

[0024] A dynamic boundary adjustment coefficient is set for each level interval to accommodate individual differences.

[0025] Preferably, the adaptive weight allocation strategy includes:

[0026] The contribution weight of each physiological parameter is calculated based on the variance distribution of the multidimensional physiological feature vector.

[0027] The contribution weights were adjusted based on the patient's medical history data;

[0028] The modified contribution weights are smoothed using a moving average algorithm to generate the comprehensive risk score.

[0029] Preferably, the activation process of the early warning response mechanism includes:

[0030] When the comprehensive risk score enters the high-risk range, a Level 1 early warning response is triggered.

[0031] When the comprehensive risk score continues to exceed the warning trigger line, the warning response will be upgraded to Level II.

[0032] The corresponding clinical intervention plan is automatically matched according to the early warning response level.

[0033] Preferably, the matching process for the clinical intervention plan includes:

[0034] Retrieve treatment plans corresponding to the current risk level from the medical knowledge base;

[0035] Exclusion criteria related to contraindications were screened based on the patient's electronic medical record;

[0036] Generate personalized intervention instruction sets and push them to terminal devices.

[0037] Preferably, the system further includes:

[0038] Establish cross-modal correlations between the real-time physiological parameters and medical imaging data;

[0039] When the comprehensive risk score is abnormal, the relevant imaging examination results are automatically retrieved for auxiliary verification.

[0040] Preferably, the system further includes:

[0041] The version iteration records of the dynamic risk assessment matrix are stored through a blockchain network;

[0042] Digital signature technology is used to preserve the evidence of early warning trigger events in an unalterable manner.

[0043] Preferably, the system further includes:

[0044] Deploy edge computing nodes to perform local preprocessing of the real-time physiological parameters;

[0045] When network latency exceeds a threshold, a local risk assessment model is activated to generate a temporary early warning decision.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] By integrating real-time data collection of multiple key physiological parameters such as heart rate, blood pressure, blood oxygen saturation, and body temperature, this method overcomes the limitations of traditional single-parameter monitoring. It allows for the acquisition of patients' physiological status information from multiple dimensions, enabling medical staff to gain a more comprehensive understanding of the overall physiological function. Compared to traditional monitoring models that rely on only one or two parameters, the coordinated acquisition of multiple parameters avoids biases in judging patients' health status caused by incomplete information from a single parameter. This results in a more three-dimensional and comprehensive perception of the patient's physiological state, helping to detect potential health risks earlier.

[0048] In the feature extraction and risk assessment stages, the system employs a feature extraction network to generate multi-dimensional physiological feature vectors. Based on these vectors, a dynamic risk assessment matrix is ​​constructed that includes the trends of physiological state changes across different time windows. This design breaks away from the limitations of traditional early warning systems that focus solely on parameter values ​​at a single point in time, instead focusing on the dynamic changes of physiological parameters over a period of time. By analyzing data from different time windows, the evolution of the patient's physiological state can be clearly presented, providing richer information support for risk assessment. This allows risk assessment to move beyond static numerical comparisons and incorporate dynamic features over time, enhancing the foresight of health risk assessments.

[0049] The system divides the dynamic risk assessment matrix into regions based on preset graded early warning thresholds, resulting in multiple risk level intervals. An adaptive weight allocation strategy is then used to dynamically adjust these intervals, generating a comprehensive risk score. This process effectively addresses the shortcomings of traditional fixed-threshold early warning methods. The adaptive weight allocation strategy fully considers individual differences among patients. For example, it flexibly adjusts the weights of parameters such as heart rate, blood pressure, and blood oxygen saturation in risk assessment based on the specific circumstances of elderly and young patients, or patients with underlying diseases and healthy patients. This ensures that the importance of different physiological parameters matches the patient's actual condition. This dynamic adjustment method makes the risk assessment results more targeted, avoiding the inaccurate risk judgments caused by fixed thresholds and weights. Whether considering individual variations in physiological parameter fluctuations or changes in parameter importance under different conditions, it provides a more realistic assessment, reducing the possibility of misjudgments and omissions.

[0050] When the comprehensive risk score exceeds a preset warning trigger line, the corresponding warning response mechanism is activated. This warning method based on comprehensive risk scores ensures the timeliness and accuracy of the warning response. Because the comprehensive risk score is derived from dynamic changes in multiple parameters and adaptive weight adjustments, it more accurately reflects the patient's current health risk level. When the risk reaches a level requiring intervention, the system can quickly trigger a warning, allowing medical staff to obtain risk information and take appropriate intervention measures in a timely manner. Simultaneously, the warning response mechanisms corresponding to different risk level ranges allow medical staff to adopt different response strategies based on the warning level. For example, for low-risk warnings, only increased monitoring frequency may be necessary, while for high-risk warnings, immediate emergency treatment is required. This tiered response method helps optimize the allocation of medical resources, enabling medical staff to handle risk events of varying degrees more efficiently, improving the overall efficiency of medical services while ensuring patient health and safety.

[0051] The system's design philosophy aligns perfectly with the actual needs of clinical diagnosis and treatment, seamlessly integrating into existing healthcare service processes. Whether in hospital intensive care units, general wards, community health centers, or home care settings, the system operates stably and effectively. For healthcare professionals, the system requires no complex training; its intuitive risk assessment results and early warning information help them make faster treatment decisions. For patients, real-time vital sign monitoring and precise risk warnings provide more personalized and targeted health management services, enhancing their sense of control over their health. Overall, through the innovative integration of multi-parameter vital sign monitoring and dynamic risk warning technologies, this system offers a new solution for improving the quality of healthcare services and ensuring patient health and safety, possessing broad clinical application value and practical significance. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the working principle of the graded early warning system based on multi-parameter vital sign detection described in this invention.

[0053] Figure 2 A flowchart for constructing a dynamic risk assessment matrix;

[0054] Figure 3 A flowchart for the adaptive weight allocation strategy. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1 This invention provides a graded early warning system based on multi-parameter vital sign detection. The system includes: a vital sign monitoring device, a feature extraction network, a dynamic risk assessment matrix construction module, a graded early warning threshold setting module, an adaptive weight allocation module, and an early warning response mechanism activation module, which realizes real-time monitoring and intelligent early warning of the patient's physiological state.

[0057] Vital signs monitoring equipment continuously collects real-time physiological parameters such as heart rate, blood pressure, blood oxygen saturation, and body temperature from patients. These parameters are transmitted to a feature extraction network for preprocessing and feature abstraction, generating a multi-dimensional physiological feature vector that comprehensively reflects the patient's physiological state. Based on this feature vector, the system constructs a dynamic risk assessment matrix. This matrix not only contains the current physiological parameter values ​​but, more importantly, records the trends in physiological state changes across different time windows, thus capturing the dynamic evolution of the physiological state. The system divides the dynamic risk assessment matrix into regions based on pre-set tiered warning thresholds from clinical experience, defining multiple risk level intervals such as low, medium, and high risk. Employing an adaptive weight allocation strategy, the system dynamically adjusts the contribution of each physiological parameter to the risk assessment based on the variance distribution of each parameter at the current moment and the patient's historical data, thereby calculating a comprehensive risk score. When the comprehensive risk score exceeds a pre-set warning trigger line, the system automatically activates the corresponding warning response mechanism, initiating the appropriate clinical intervention process, achieving closed-loop management from data collection to warning response.

[0058] Example 1: The implementation of the feature extraction network involves a series of refined processing steps on the raw physiological signals, aiming to extract feature representations that can characterize the essence of physiological state from noisy time-series data. Parameters such as heart rate, blood pressure, blood oxygen saturation, and body temperature continuously collected by vital sign monitoring equipment are initially presented as raw reading sequences with timestamps. These sequences are easily affected by various interference factors during acquisition and transmission, such as signal interruptions caused by poor sensor-skin contact, artifacts generated by patient limb movements, or the superposition of environmental electromagnetic noise. To eliminate these noise interferences and prepare a clean data foundation for subsequent deep feature extraction, time-series standardization is strictly performed as the first step. This processing typically includes two main stages: noise filtering and data normalization. In the noise filtering stage, the system uses a sliding window-based digital filter, such as Savitz. The Ky-Golay filter, while smoothing the signal, effectively preserves key morphological features of the original waveform, such as the R-wave peak of the heartbeat or the dicrotic notch of the blood pressure curve. It estimates the optimal value of each data point through local polynomial fitting, thereby effectively suppressing random noise without significantly distorting the physiological meaning of the signal. In the data normalization stage, to overcome the analytical difficulties caused by the differences in the dimensions of different physiological parameters, the system applies the Z-score normalization method. That is, the mean and standard deviation of the numerical sequence of each parameter within its time window are calculated, and then the mean is subtracted from each original data point and divided by the standard deviation. The transformed data conforms to a standard normal distribution with a mean of zero and a standard deviation of one. This process not only eliminates the scale difference between parameters, allowing different parameters to be compared and calculated in the same space, but also indirectly weakens the impact of the difference in baseline levels between individuals on the model's generalization ability.

[0059] The standardized time-series data is then fed into a convolutional neural network for deep feature learning. This network is specifically designed to process one-dimensional time-series signals, and its structure typically includes alternating stacked convolutional and pooling layers. Each convolutional layer is equipped with multiple one-dimensional convolutional kernels that slide along the time axis with a set stride. Through convolution operations, local temporal patterns in the input sequence are automatically extracted. For example, a narrow convolutional kernel might focus on identifying the interval variation between adjacent R waves in a heart rate sequence, i.e., the high-frequency component of heart rate variability, while a wider convolutional kernel might capture the complete systolic and diastolic morphology in a blood pressure waveform, or blood oxygen saturation... The saturation curve contains a long plateau period or a sudden falling edge. Each convolutional kernel is essentially a feature detector, whose parameters are learned during training through backpropagation. It can produce a high response to local patterns. The output of the convolutional layer is passed through a non-linear activation function, such as the ReLU function, to introduce the non-linear expressive power of the model. Subsequently, the pooling layer downsamples the convolutional feature map, retaining the most significant feature response while reducing the data dimensionality and enhancing the invariance of features to small time drifts. Through multiple layers of convolution and pooling operations, the network can gradually combine shallow local features into deeper and more abstract features.

[0060] The features extracted by convolutional neural networks mainly focus on patterns at local time scales. To obtain a more comprehensive description of physiological states, the system calculates global statistical features for each physiological parameter in parallel. These features are calculated from the overall data distribution across the entire analysis time window and reflect the macroscopic behavioral characteristics of the parameters over a longer period. Commonly used global statistical features include the arithmetic mean, which indicates the central tendency of the parameter within the window; the standard deviation, which quantifies the degree of fluctuation of the parameter; the maximum and minimum values, which define the range of parameter variation; skewness, which describes the asymmetry of the data distribution, such as whether the heart rate distribution is biased towards the high-speed side; and kurtosis, which reflects the sharpness or flatness of the distribution shape. Compared with the normal distribution, these statistics supplement the global perspective missing from local time-series features from different angles. For example, even if local convolutional features show occasional premature beats in the heart rate sequence, if the global heart rate mean remains within the normal range and the standard deviation is small, the overall risk may still be manageable.

[0061] The fusion of local temporal features and global statistical features is a key step in generating vectors with strong representational power. The local features output by a convolutional neural network are typically high-dimensional feature maps, which are flattened into a feature vector. Global statistical features, on the other hand, are low-dimensional vectors. The system designs a fully connected layer network structure to process these two feature vectors from different sources. First, the flattened local feature vector is concatenated with the global statistical feature vector to form a longer joint feature vector. Then, this joint vector is input into one or more fully connected layers. Each neuron in the fully connected layer is connected to all neurons in the previous layer. Through linear transformations of the weight matrix and bias vector, coupled with a nonlinear activation function, deep interaction and information integration of features are achieved. In this fusion process, the network automatically learns how to weigh the importance of local details and global statistical information. For example, it may learn to rely more on capturing local features of abnormal fluctuations when judging the risk of acute events, while giving higher weight to global statistical features when assessing long-term trends. Finally, the output of the fully connected layer is normalized into a fixed-dimensional multi-dimensional physiological feature vector, which encapsulates the most discriminative information extracted from the original signal, providing a solid and rich input for subsequent dynamic risk assessment.

[0062] In practical engineering implementation, the parameters of the feature extraction network need to be trained on a large amount of labeled physiological data to optimize its performance. The training process uses historically collected vital sign data and their corresponding clinical status labels. Through supervised learning, such as using gradient descent algorithm and its variants, the loss function between the network output and the real label is minimized, thereby adjusting parameters such as convolutional kernel weights and fully connected layer weights in the network. This enables the network to learn to extract features that are highly correlated with clinical risk from the raw data. In addition, in order to adapt to different monitoring scenarios and patient groups, the network structure may need to be appropriately adjusted. For example, for electrocardiogram signals with high sampling frequency, deeper network layers or more convolutional kernels may be needed to capture subtle changes, while for body temperature data with slower sampling, the network structure may be simplified. The entire feature extraction module is usually deployed on edge devices or central servers with sufficient computing resources to achieve real-time or near real-time processing capabilities to meet the timeliness requirements of clinical monitoring.

[0063] Example 2: See Figure 2The construction of a dynamic risk assessment matrix is ​​a process of transforming discrete time-point data into a continuous dynamic risk profile. Its core lies in quantifying the interaction between the instantaneous abnormality of physiological states and their evolutionary trends over time. The system receives multi-dimensional physiological feature vectors from a feature extraction network. Although these vectors have been refined and condensed, their dynamic characteristics over time still need further analysis to effectively assess risk. The construction process begins with a temporal analysis of the feature vector sequence. The system maintains a sliding time window, continuously tracking the feature vector sequence over a recent period. For the feature vector at the current moment, the system calculates its difference from the feature vectors at several previous key time points in the abstract feature space. This difference can be achieved through mathematical methods that measure the distance between vectors, such as Euclidean distance or Manhattan distance. Euclidean distance measures the straight-line distance between vectors in space, comprehensively reflecting the overall changes in all feature dimensions. Manhattan distance, on the other hand, takes the sum of the absolute values ​​of the differences in each dimension and may be more sensitive to drastic changes in individual dimensions. The choice of distance metric depends on the specific clinical application scenario and the focus on changes in different parameters.

[0064] After calculating the feature vector distance between the current moment and historical moments, the system further analyzes the rate of change of this distance over time, i.e., the deviation of physiological state. This deviation is estimated by calculating the first difference or instantaneous derivative of the feature vector distance between consecutive time points. A rapidly increasing distance derivative means that the patient's physiological state is rapidly deviating from its recent baseline. This rapid dynamic change is often more clinically significant than a static absolute value exceeding the standard. For example, a steady increase in heart rate from 70 beats / min to 85 beats / min is different from a sudden spike from 70 beats / min to 100 beats / min in a very short time. The higher rate of change in the latter corresponds to a higher physiological state deviation score. The system calculates the deviation of each physiological parameter or feature dimension and may form a comprehensive deviation index. Meanwhile, the system performs another crucial calculation in parallel: generating trends in physiological state changes across different time windows. This task relies on historical physiological data over a longer period. The system extracts recent physiological feature vector sequences or raw values ​​of key parameters from the patient database and applies trend fitting algorithms to these time series data. Linear regression is a fundamental method that attempts to find a straight line that best represents the overall trend of the data points. The slope of this line intuitively indicates whether the parameter is rising, falling, or remaining stable within the observation window. For data exhibiting non-linear changes, polynomial fitting or exponential smoothing may be used. Exponential smoothing is particularly suitable for situations where more weight is given to recent data, as it can better reflect the latest changes. Through trend analysis, the system can determine whether a physiological parameter is in a downward trend of continuous deterioration, a steady recovery process, or fluctuating within the normal range.

[0065] The system needs to correlate and map the calculated deviation of physiological state with the trend of physiological state change to fill the dynamic risk assessment matrix. This matrix can be conceived as a multi-dimensional table or space, whose dimensions include, but are not limited to, deviation level and trend direction and intensity. Each cell of the matrix corresponds to a "deviation-trend" combination and a predefined risk level or risk score. The correlation and mapping process is to find the corresponding position in the matrix based on the currently calculated specific deviation value and trend value, and use the predefined risk value at that position as the preliminary risk assessment result at the current moment. For example, a high deviation of physiological state combined with a long-term trend of continuous deterioration will be mapped to the area with the highest risk level in the matrix, while a low deviation combined with an improving or stable trend will be mapped to a low-risk area. This mapping relationship is usually predefined based on clinical knowledge and can be optimized and learned from historical case data through machine learning methods.

[0066] The dynamic risk assessment matrix itself is also an entity that evolves over time. The system updates the matrix content at a fixed frequency. The updates not only include loading the latest deviation and trend data, but may also involve fine-tuning the matrix structure. For example, it may adaptively adjust the thresholds of different risk intervals based on the patient's current overall condition. The maintenance of the matrix requires efficient data structures and algorithms to ensure that calculations and updates can be completed quickly in real-time monitoring scenarios. The resulting dynamic risk assessment matrix provides a richer and more forward-looking risk view than a single-point judgment. It can not only tell the patient what state they are "now" in, but also indicate how this state developed and in which direction it "might" develop, providing valuable dynamic contextual information for clinical decision-making. To achieve the above process, the system needs to reliably store and manage time-series physiological data. A time-series database is typically used to efficiently process time-stamped feature vector streams. Algorithms for calculating deviation and trends need optimization to achieve low latency and meet real-time monitoring requirements. For the selection of time window length involved in trend analysis, a balance needs to be struck between sensitivity and specificity. Too short a window may be overly sensitive to noise, generating false positive alarms; too long a window may lead to a slow but definite deterioration trend. Furthermore, the dimensions of the dynamic risk assessment matrix can be expanded according to clinical needs, for example, by adding a third dimension to represent the synergistic change patterns between different physiological parameters, thereby further improving the depth and accuracy of risk assessment. The entire construction process emphasizes the capture and interpretation of the dynamic characteristics of physiological signals, elevating static parameter monitoring to the level of dynamic risk trajectory assessment.

[0067] Example 3: See Figure 3The setting of graded warning thresholds is the cornerstone of risk assessment. Its primary basis is the normal fluctuation range of physiological parameters based on the statistical analysis of a large number of clinical data. These ranges come from authoritative medical guidelines, textbooks, and consensus knowledge obtained by analyzing long-term monitoring data of healthy people and people with specific diseases. For example, for the resting heart rate of adults, the low-risk range is usually set between 60 and 100 beats per minute, the low-risk range of systolic blood pressure may be set between 90 and 140 mmHg, and the lower limit of the low-risk range of blood oxygen saturation is generally set at 95%. These numerical ranges constitute the initial, universal risk classification benchmark. However, strict fixed thresholds cannot accommodate individual physiological differences. A professional athlete who has been training for a long time may have a resting heart rate of around 50 beats per minute. If a lower limit of 60 beats per minute is mechanically applied to him, it will cause the system to continuously generate false alarms. Therefore, the system introduces the concept of dynamic boundary adjustment coefficient. This coefficient allows for personalized offsetting of the universal threshold based on the patient's own steady-state baseline. The system selects data from a period of relatively stable physiological state during the patient's admission or early monitoring period, calculates the average value of each parameter as the individual baseline value, and calculates the dynamic boundary adjustment coefficient based on the difference between this individual baseline value and the standard population mean. This allows for adaptive adjustment of the boundaries of low, medium, and high risk zones, making the warning threshold more in line with the individual's physiological characteristics.

[0068] After establishing a tiered early warning threshold adapted to individual baselines, the system needs to integrate multiple physiological parameters to generate a single comprehensive risk score. An adaptive weight allocation strategy plays a crucial role in this step, aiming to dynamically determine the relative importance of each physiological parameter in the overall risk assessment. The initial weight allocation is based on the variance distribution of each parameter within a recent time window. Variance is a statistic that measures the degree to which a data point deviates from its mean. For a physiological parameter sequence, the variance is calculated using the following formula:

[0069] ;

[0070] in: This represents the variance of a certain physiological parameter p within the time window of the most recent N sampling points. This represents the specific value of the parameter at the i-th sampling point. This represents the arithmetic mean of the parameter within this window, where N is the total number of sampling points within the window. A high variance value for a parameter indicates that its value has fluctuated drastically in the recent period and is less stable. This unstable state may indicate potential physiological disorders or decompensation in response to changes in the internal and external environment. Therefore, the contribution of this parameter to the current overall risk should be given a higher weight. Conversely, a parameter with an extremely low variance value indicates that its state is very stable, and its weight in risk assessment should be reduced accordingly. In this way, the system can automatically focus on parameters that have undergone abnormal changes, rather than treating all parameters equally.

[0071] Weighting based solely on variance is a purely data-driven approach that may overlook important prior clinical knowledge. Therefore, the system further refines the initial weights by incorporating patient medical history data. This refinement process relies on a predefined clinical rule base, which encodes clinical knowledge about the importance of various physiological parameters under specific disease conditions. For example, for a patient with a clear history of chronic heart failure, the base weights of blood pressure and heart rate will be appropriately increased based on the variance calculation, as these two parameters are particularly crucial for assessing decompensation in heart failure patients. For a patient with a history of chronic obstructive pulmonary disease, oxygen saturation and respiratory rate will receive greater attention. The refinement rule may manifest as a weight scaling factor, which queries and matches key diagnostic information extracted from the patient's electronic medical record. The initial weights are multiplied by the scaling factor, thus integrating clinical experience into the weighting process.

[0072] Even after adjusting for medical history, the weights may still jump due to sudden and drastic fluctuations in parameters. To smooth the temporal changes in weights and avoid excessive oscillations in the comprehensive risk score, the system uses a moving average algorithm to process the weight sequence. Moving average is a commonly used time series smoothing technique. It calculates the arithmetic mean of the weights at the current time and several previous times as the final output weight at the current time. For example, a moving average with a window of length of 5 will take the weight values ​​at the current time and the previous 4 times to calculate the average. This method can effectively filter out short-term noise in the weight sequence, so that the weights assigned to each parameter can reflect their recent importance trend rather than instantaneous fluctuations, thereby generating more stable and reliable final weights. The comprehensive risk score is calculated through weighted summation. The system first maps the current value of each physiological parameter to a standardized risk score based on preset graded warning thresholds. For example, the low-risk range is mapped to 1 point, the medium-risk range to 2 points, and the high-risk range to 3 points. Then, the risk score of each parameter is multiplied by its smoothed adaptive weight, and the weighted risk scores of all parameters are summed to obtain a continuous comprehensive risk score. This score integrates multi-parameter information, the dynamic fluctuation characteristics of parameters, and the patient's individualized clinical background, realizing the transformation from multi-dimensional physiological characteristics to a single quantitative risk indicator, providing a clear and operable basis for early warning decisions.

[0073] Example 4: The system's early warning response mechanism adopts a tiered upgrade mode. Its triggering depends entirely on the comparison between the comprehensive risk score calculated in real time and the preset threshold. When the score first exceeds the set high-risk threshold and enters the high-risk level range, the first-level early warning response is triggered. At this time, the system will display the patient's bed information in a bright yellow on the central monitoring screen of the nurse's workstation responsible for monitoring the patient, accompanied by a continuous, medium-frequency prompt sound. At the same time, an alarm message containing the patient's basic information, a summary of abnormal physiological parameters, and the current risk score will be pushed to the nurse's mobile handheld terminal. This level of response aims to attract the attention of medical staff and remind them that it is necessary to strengthen the observation and evaluation of the patient. If the patient's physiological condition does not improve effectively, and the comprehensive risk score remains above the high-risk threshold for three consecutive monitoring cycles, or if the score shows a rapid upward trend in a short period of time, the system will automatically upgrade the warning level to a Level 2 warning response. A Level 2 warning means that the condition may be deteriorating or there is a high risk of acute events, requiring more urgent medical intervention. At this time, the system will activate a stronger alarm mode. The patient's information box on the central monitoring screen will turn red and flash, the alarm tone will become high-frequency and rapid, and the alarm information will be directly pushed to the attending physician's mobile terminal and even the department's emergency team's call system through the hospital information system. The information content will clearly mark "Level 2 Warning" and briefly point out the most prominent abnormal indicator combination.

[0074] Once an alert is triggered, the associated clinical intervention plan matching process is immediately initiated. The core of this process is to retrieve and select the most suitable treatment plan for the current risk status and individual patient condition from a structured medical knowledge base. The system first analyzes the specific abnormal combination of physiological parameters when the alert is triggered, such as "low blood oxygen saturation accompanied by tachycardia and decreased blood pressure". Using this as a key search term, the system retrieves a set of relevant standard treatment plans from a medical knowledge base that integrates clinical pathways, critical value management procedures, drug manuals, etc. The search results may include multiple alternative plans such as "suspected hypoxemia management procedure" and "early identification and intervention guidelines for shock", see Table 1.

[0075] Table 1: Screening for Contraindications in Patients

[0076]

[0077] The retrieved standard protocols undergo rigorous personalized filtering. The system retrieves the patient's electronic medical record in real time, focusing on screening for contraindications, allergy history, and vital organ function. As shown in the table above, for patient A-123, the system retrieved three potentially relevant interventions. However, based on the patient's history of "sulfonamide allergy" in their electronic medical record, Protocol A, which includes the use of this type of antibiotic, was directly excluded. For Protocol B (fluid resuscitation), the system recognized the patient's history of "renal insufficiency." Therefore, while retaining the protocol, a warning was added to the generated instructions: "Caution is required; strictly control the infusion rate and closely monitor urine output and renal function indicators." Only measures without explicit contraindications, such as Protocol C (adjusting oxygen therapy), are adopted without reservation. After screening, the system integrates the remaining feasible treatment steps to generate a structured set of personalized intervention instructions. This set of instructions is usually organized according to the reasonable order of clinical operations. For example, for a low blood oxygen saturation warning, the instruction set may include the following in sequence: "1. Immediately assess the patient's level of consciousness and breathing pattern; 2. Check whether the nasal cannula or mask is unobstructed and correctly positioned; 3. Increase the oxygen concentration or flow rate to [the recommended value calculated by the system based on the current saturation]; 4. Instruct the patient to cough and expectorate effectively; 5. Prepare equipment for arterial blood gas analysis; 6. Notify the attending physician for further evaluation." This set of instructions will include a summary of the physiological data that triggered the warning and will be pushed to the nurse's mobile terminal or the smart display screen at the bedside in the ward through an interface to guide medical staff to perform rapid, standardized, and individualized interventions. The system will also record the time of instruction push, the recipient, and the subsequent execution status entered by medical staff, forming a complete warning-response closed-loop management record.

[0078] Example 5: Cross-modal correlation analysis aims to break down information silos between vital sign data and medical imaging data. When the system continuously monitors and calculates data to detect an abnormality in a patient's comprehensive risk score, such as a persistent and inexplicable decline in blood oxygen saturation over half an hour, simple numerical alarms may not reveal the root cause. In this case, the system automatically initiates a correlation retrieval process. It uses the patient's unique identifier, such as their hospital number, to query the hospital's medical imaging archive and communication system, retrieving recently completed imaging reports and raw image data that may be related to the current abnormal vital signs. For example, for respiratory-related risks, the system... The system prioritizes retrieving the most recent chest X-ray or CT scan images and diagnostic reports. After these images are retrieved, they are not automatically diagnosed by the system, but are pushed to the workstation interface of the responsible physician and displayed side by side with the abnormal vital sign trend curve. The physician can immediately observe whether the chest X-ray shows new lung infiltrates, whether the amount of pleural effusion has increased, or whether there are signs of pneumothorax while the blood oxygen saturation is decreasing. This mutual corroboration between the temporal changes of vital signs and the spatial morphological changes of imaging greatly assists clinicians in making rapid judgments on the cause of the disease. It is a key step in shifting the early warning from "what happened" to "why it happened".

[0079] The data security and traceability mechanism relies on blockchain technology to build a trusted audit trail chain. The system treats each generated or updated dynamic risk assessment matrix as a significant data version. When the matrix is ​​finalized or undergoes major changes, the system calculates the cryptographic hash value of the matrix data. This hash value acts like a unique digital fingerprint for the data; any minor alteration to the data will result in a significant change in the hash value. This hash value, along with a timestamp, operator identification, and the hash value of the previous block, constitutes a new block and is broadcast to a permissioned blockchain network within the hospital or consortium for notarization. After the nodes in the chain verify the validity of the block through the consensus mechanism, they add it to the chain to form an immutable and continuous record. Similarly, each warning trigger event, including key information such as the trigger time, risk score, trigger threshold level, system-suggested intervention instruction set, and the ID of the medical staff receiving the alarm, is also hashed, generated a digital signature, and recorded on the chain. This method of evidence storage ensures that the evolution history of all risk assessments and key warning nodes cannot be unilaterally altered afterward, providing highly legally valid electronic evidence for medical quality traceability, verification of the authenticity of clinical research data, and even potential medical disputes.

[0080] The deployment strategy of edge computing nodes aims to address potential network instability issues in the clinical environment and ensure the continuity of monitoring services. The system deploys edge servers or gateway devices with certain computing capabilities within the ward. Vital signs monitoring devices first send the collected raw physiological parameter data to the local edge nodes via short-range communication methods such as Bluetooth, Zigbee, or LAN. The edge nodes are pre-installed with optimized lightweight feature extraction and risk assessment algorithms, which can independently perform real-time preprocessing of data, basic feature calculations, and even simple risk assessments, generating a local, temporary comprehensive risk score. Under good network connectivity, the edge nodes will synchronize these pre-processed data packets and local score results to the central server in real time, where the central server will perform more complex model calculations and global data fusion. Once an edge node detects that the network communication latency with the central server exceeds a set safety threshold (e.g., 500 milliseconds) or the network connection is completely interrupted, it immediately switches to autonomous operation mode. At this time, the edge node no longer attempts to send data to the cloud, but relies entirely on local computing resources to continue generating risk assessment results and early warning decisions based on the latest data stream. It also triggers audible and visual alarms locally in the wards and highlights abnormal patient information on the local area display screen at the nurses' station. This design ensures that even in the extreme case of a hospital-wide network outage, each ward can still maintain the most basic ability to monitor and alarm for abnormal patient vital signs, adding an important local defense line for patient safety. When the network connection is restored, the edge node will upload the local data and event records stored during the network interruption to the central server in batches for data archiving and model calibration, ensuring the integrity of the data chain.

[0081] The combination of these three functions transforms the system from an isolated data analysis tool into a resilient system capable of deeply integrating multi-source clinical information, ensuring traceability and reliability of the operational process, and maintaining core functionality even in harsh network environments. For example, if a postoperative patient's vital signs show an increased heart rate and a slight decrease in blood pressure, the system triggers a medium-level alert and automatically retrieves the patient's postoperative chest CT images, presenting this information to the nurses. All details of this alert event are instantly recorded on the blockchain. Even if a brief fluctuation occurs in the hospital's internal network, thanks to the local processing capabilities of the edge nodes, the alarm information is promptly sent within the ward without delay due to network issues. This entire process demonstrates the practical value of the expanded functionality in Implementation Example 5 when facing complex real-world scenarios.

[0082] 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.

[0083] 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. A graded early warning system based on multi-parameter vital sign detection, characterized in that, include: The patient's real-time physiological parameters are collected through vital sign monitoring equipment, including heart rate, blood pressure, blood oxygen saturation, and body temperature. The real-time physiological parameters are input into a feature extraction network to generate a multi-dimensional physiological feature vector. A dynamic risk assessment matrix is ​​constructed based on the multi-dimensional physiological feature vectors, and the dynamic risk assessment matrix contains the trend of physiological state changes under different time windows. The dynamic risk assessment matrix is ​​divided into regions based on preset graded early warning thresholds to obtain multiple risk level intervals; An adaptive weighting strategy is used to dynamically adjust the risk level intervals to generate a comprehensive risk score for the current patient. When the comprehensive risk score exceeds the preset warning trigger line, the corresponding warning response mechanism is activated.

2. The graded early warning system based on multi-parameter vital sign detection according to claim 1, characterized in that, The feature extraction network includes: The real-time physiological parameters are subjected to time series standardization to eliminate noise interference during the acquisition process; Local temporal features of the real-time physiological parameters are extracted using a convolutional neural network; The local temporal features are fused with the global statistical features to generate the multi-dimensional physiological feature vector.

3. The graded early warning system based on multi-parameter vital sign detection according to claim 1, characterized in that, The process of constructing the dynamic risk assessment matrix includes: The deviation of physiological state is calculated based on the rate of change of the multi-dimensional physiological feature vectors. By combining historical physiological data, the trend of physiological state changes under the different time windows is generated; The deviation of the physiological state is correlated with the trend of physiological state change, and then the dynamic risk assessment matrix is ​​filled.

4. The graded early warning system based on multi-parameter vital sign detection according to claim 1, characterized in that, The process of setting the tiered early warning threshold includes: The normal fluctuation range of different physiological parameters is determined based on clinical data statistics; Based on the normal fluctuation range, the risk levels are divided into three categories: low risk, medium risk, and high risk. A dynamic boundary adjustment coefficient is set for each level interval to accommodate individual differences.

5. The graded early warning system based on multi-parameter vital sign detection according to claim 1, characterized in that, The adaptive weight allocation strategy includes: The contribution weight of each physiological parameter is calculated based on the variance distribution of the multidimensional physiological feature vector. The contribution weights were adjusted based on the patient's medical history data; The modified contribution weights are smoothed using a moving average algorithm to generate the comprehensive risk score.

6. The graded early warning system based on multi-parameter vital sign detection according to claim 1, characterized in that, The activation process of the early warning response mechanism includes: When the comprehensive risk score enters the high-risk range, a Level 1 early warning response is triggered. When the comprehensive risk score continues to exceed the warning trigger line, the warning response will be upgraded to Level II. The corresponding clinical intervention plan is automatically matched according to the early warning response level.

7. The graded early warning system based on multi-parameter vital sign detection according to claim 6, characterized in that, The matching process for the clinical intervention plan includes: Retrieve treatment plans corresponding to the current risk level from the medical knowledge base; Exclusion criteria related to contraindications were screened based on the patient's electronic medical record; Generate personalized intervention instruction sets and push them to terminal devices.

8. The graded early warning system based on multi-parameter vital sign detection according to claim 1, characterized in that, Also includes: Establish cross-modal correlations between the real-time physiological parameters and medical imaging data; When the comprehensive risk score is abnormal, the relevant imaging examination results are automatically retrieved for auxiliary verification.

9. The graded early warning system based on multi-parameter vital sign detection according to claim 1, characterized in that, Also includes: The version iteration records of the dynamic risk assessment matrix are stored through a blockchain network; Digital signature technology is used to preserve the evidence of early warning trigger events in an unalterable manner.

10. The graded early warning system based on multi-parameter vital sign detection according to claim 1, characterized in that, Also includes: Deploy edge computing nodes to perform local preprocessing of the real-time physiological parameters; When network latency exceeds a threshold, a local risk assessment model is activated to generate a temporary early warning decision.

Images