A VTE real-time monitoring and intelligent prevention system

Abstract

The application discloses a VTE real-time monitoring and intelligent prevention and treatment system and belongs to the technical field of intelligent prevention and treatment, comprising: a baseline construction module, which is used for collecting multi-dimensional VTE parameters, calculating the normal fluctuation range of each parameter to form an initial individual baseline, and constructing an adaptive individual baseline through threshold calibration; a trend identification module, which is used for dynamically setting the length of a sliding window, calculating the trend slope and VTE accumulation bias of VTE, constructing trend constraints and accumulation bias constraints, and determining that there is a continuous abnormal deviation when both constraints are not met at the same time and the length of continuous deviation exceeds a certain time; a time sequence risk prediction module, which is used for calculating the deviation value of each parameter of a target patient, constructing a space-time fusion feature matrix, inputting a time sequence risk prediction model, and outputting a VTE risk probability; and an early warning intervention module, which is used for double determination intervention, setting three-level early warning and intervention measures, calculating an improvement rate to verify the prevention and control effect in real time, and realizing early identification and intelligent prevention and control of VTE.

Description

Technical Field

[0001] This invention relates to a VTE real-time monitoring and intelligent prevention system, belonging to the field of smart prevention technology. Background Technology

[0002] Venous thromboembolism (VTE) is a serious vascular disease, often called a "silent killer," and frequently leads to severe complications, even death. Traditional prevention and treatment rely on manual assessment, resulting in high rates of missed diagnoses and delayed responses. Existing monitoring equipment is mostly limited to collecting single physiological parameters and lacks multimodal data fusion and intelligent early warning capabilities, making it difficult to achieve early, dynamic, and precise intervention. There is an urgent need to improve the timeliness and effectiveness of VTE prevention and control.

[0003] Chinese patent CN111968747A discloses a VTE intelligent prevention and management system. The system includes a word segmentation module that segments preprocessed historical text data to identify associated influencing factors; a calculation module that calculates the support for VTE when any associated influencing factor appears in all historical text data, and the confidence level for VTE in all historical text data containing that associated influencing factor; a judgment module that determines whether the support and confidence levels are greater than corresponding set values, using the associated influencing factor as a key influencing factor; a dimensionality reduction module that reduces the dimensionality of the key influencing factors to obtain the optimal influencing factor; a training module that uses a portion of the historical text data to train a random forest model to obtain a trained risk assessment model; a testing module that uses the optimal influencing factor from the remaining historical text data to test the risk assessment model and obtain the probability value of a patient having VTE; and a grading module that uses a normal distribution 3. Based on the principle of categorizing patients into risk levels, a corresponding risk level label is obtained for each test patient.

[0004] Although existing VTE intelligent prevention and management systems can automatically classify and alert patients to VTE risk by segmenting historical text data, filtering related influencing factors, and constructing a random forest risk assessment model based on support and confidence analysis, they lack an individualized physiological baseline and cannot identify abnormal deviations in the target patient's own parameters, resulting in inaccurate assessments of high-risk groups. In addition, VTE is often accompanied by subtle but persistent deviations in physiological parameters, such as a slow decrease in blood oxygen and a slight increase in heart rate at night. Current technology lacks temporal sensitivity to subtle abnormal patterns, leading to missed early warning windows. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a VTE real-time monitoring and intelligent prevention system. By constructing an adaptive individualized baseline, trend constraints, and product bias constraints, it achieves early identification of continuous abnormal deviation patterns. It integrates spatiotemporal features to construct a time-series risk prediction model for VTE risk prediction and establishes a hierarchical early warning and closed-loop intervention mechanism based on dual judgment. Combined with the improvement rate, it evaluates the prevention and control effect in real time, effectively solving the problems of delayed early warning, high false alarm rate, and lack of personalized prevention and control in existing technologies, and realizing early and accurate identification and dynamic intelligent prevention and control of VTE.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A VTE real-time monitoring and intelligent prevention system includes: a baseline construction module, a trend recognition module, a time-series risk prediction module, and an early warning and intervention module;

[0008] The baseline construction module is used to monitor and collect multidimensional VTE parameters in real time, calculate the normal fluctuation range of each parameter to form an initial individualized baseline, and construct an adaptive individualized baseline through threshold calibration. This includes: acquiring the static characteristics of the target patient, quantifying and generating a static feature vector, and extracting key statistical features to construct an association matrix.

[0009] The static feature vector is concatenated with the key statistical features to generate an individual feature vector, and the threshold adjustment range of each parameter is calculated.

[0010] Calculate the risk probability of the target patient's underlying disease;

[0011] When the probability of underlying disease risk exceeds a preset probability threshold, the patient is identified as a high-risk patient. The threshold adjustment range is calibrated by setting a risk weighting coefficient according to the risk level, and the calibrated threshold adjustment range is obtained.

[0012] The adjustment range of the calibration threshold exceeding the preset safety boundary is truncated to the boundary value to obtain the initial individualized baseline threshold;

[0013] Based on the current physiological state of the target patient, the initial individualized baseline threshold is dynamically adjusted to obtain a dynamic threshold baseline;

[0014] Differentiate physiological adaptations, calculate the trend change rate, and adaptively adjust the dynamic threshold baseline;

[0015] Identify pathological trends, recalculate the threshold adjustment range, and iteratively update the dynamic threshold baseline;

[0016] Integrate threshold results from all states and trends to construct an adaptive individualized baseline;

[0017] The trend recognition module is used to identify abnormal VTE trends in key parameters and dynamically set the duration of the sliding window. Calculate the VTE trend slope and VTE product deviation of the parameters within the sliding window, and construct trend constraints and product deviation constraints. When the parameters do not satisfy both trend constraints and product deviation constraints and the continuous offset time exceeds [a certain value], [further action is taken]. When this occurs, it is determined to be a continuous abnormal offset;

[0018] The time-series risk prediction module is used to calculate the deviation values ​​of various VTE parameters of the target patient, and constructs a spatiotemporal fusion feature matrix by combining time-series features and static features as input to build a time-series risk prediction model and output the VTE risk probability.

[0019] The project constructs a time-series risk prediction model, which includes an input layer, an LSTM feature extraction layer, an attention weighting layer, and an output layer.

[0020] The input layer receives the spatiotemporal fusion feature matrix, normalizes it, and adapts it to the length of the bias time series;

[0021] The LSTM feature extraction layer introduces a temporal attention mechanism to improve the two-layer bidirectional LSTM structure. The first layer extracts the first feature of the target patient, and the second layer extracts the second feature of the target patient. Residual connections are used to obtain residual fusion features, and linear transformation is used to generate LSTM feature sequences.

[0022] The attention weighting layer calculates the temporal attention weight and the feature attention weight, and combines them with the LSTM feature sequence to obtain the key temporal feature vector through weighted fusion;

[0023] The output layer maps the key temporal feature vectors to risk probability values;

[0024] The early warning and intervention module is used to make dual judgments and interventions based on parameter trend changes and VTE risk probability, set three levels of early warning and intervention measures, monitor VTE parameter changes in real time after intervention, and calculate the improvement rate to verify the prevention and control effect.

[0025] Specifically, the steps for establishing an initial individualized baseline include:

[0026] The VTE parameters are preprocessed to obtain a VTE parameter set, which is divided into high-frequency dynamic parameters and low-frequency static parameters.

[0027] The missing time point data of low-frequency static parameters are supplemented by the timestamps of high-frequency dynamic parameters to obtain the classification parameter sequence;

[0028] The mean values ​​of each parameter are calculated based on the classification parameter sequence, and the correction coefficient and the basic parameter values ​​are combined to generate a correction parameter baseline value sequence.

[0029] A differentiated window segmentation strategy is adopted to calculate the moving mean and standard deviation of the parameters within each sliding window, and the initial fluctuation range is determined with the benchmark values ​​of each correction parameter as the center.

[0030] Calculate the correlation degree between high-frequency dynamic parameters, filter strongly correlated parameter pairs according to a preset correlation threshold, and obtain the high-frequency correction interval by correcting the initial fluctuation interval based on the correlation direction.

[0031] If the low-frequency static parameter value is lower than the preset deviation threshold, the associated vital sign interval is contracted upward; otherwise, the high-frequency correction interval is maintained to obtain the correction fluctuation interval.

[0032] The revised fluctuation range is updated to form an initial individualized baseline.

[0033] Specifically, the trend recognition module includes a trend analysis unit and an anomaly recognition unit;

[0034] The trend analysis unit is used for real-time dynamic monitoring and feature extraction of key parameters. Based on the adaptive individualized baseline, a trend-sensitive detection algorithm is introduced to identify VTE abnormal trends, and the sliding window duration is dynamically set. Within each sliding window, calculate the VTE trend slope and VTE product bias of the parameters;

[0035] The anomaly identification unit is used to receive the VTE trend slope and the VTE product skewness, construct trend constraints and product skewness constraints, and when the parameter changes within the sliding window do not simultaneously satisfy the trend constraints and product skewness constraints, and the continuous offset time exceeds [a certain value], [the anomaly identification unit will detect the anomaly]. If the deviation is abnormal, it is determined to be a continuous abnormal deviation, and the temporal characteristics are recorded; otherwise, it is determined to be a normal fluctuation within the physiological range, and monitoring continues.

[0036] Specifically, the steps for identifying abnormal VTE trends include:

[0037] Set the base duration of the sliding window for key parameters in real-time monitoring;

[0038] statistics Calculate the baseline fluctuation coefficient using adaptive individualized baseline fluctuation data from the past few days. ;

[0039] Set maximum fluctuation limit With minimum fluctuation limit ;

[0040] like If the fluctuation is severe, the duration of the sliding window is determined based on the base duration and a preset first weight. ;

[0041] like If the fluctuation is stable, the sliding window duration is determined based on the base duration and a preset second weight. ;

[0042] Using the timestamp within the sliding window as the independent variable and the parameter value as the dependent variable, and setting the time decay weight, the VTE trend slope is obtained by fitting using the weighted least squares method.

[0043] The difference between the parameter value that exceeds the normal range of the adaptive individualized baseline and the value in the normal range is calculated, and then weighted and summed in combination with the time decay weight to obtain the VTE product bias.

[0044] Specifically, the steps for identifying abnormal VTE trends also include:

[0045] Set slope limits;

[0046] A trend constraint is constructed. If the slope of the VTE trend is less than the slope limit, the trend is determined to be normal; otherwise, it is determined to be a suspected abnormal trend.

[0047] Set the product bias constraint value;

[0048] Construct an integral bias constraint. If the VTE integral bias is less than the integral bias constraint value, the offset is determined to be normal; otherwise, it is determined to be a suspected offset anomaly.

[0049] If both the trend constraint and the product deviation constraint are satisfied simultaneously, then the fluctuation is determined to be normal.

[0050] If only one constraint is satisfied, it is marked as to be observed, and high-density monitoring mode is started;

[0051] If neither the trend constraint nor the product deviation constraint is satisfied, then a consistency verification is performed. If the continuous offset duration exceeds [a certain value], [the verification will be performed]. If the previous sliding window is the one to be observed, it is determined to be a continuous abnormal offset, and the temporal characteristics are recorded.

[0052] Otherwise, it is judged as an abnormal instantaneous fluctuation, and the monitoring continues in the next sliding window.

[0053] Specifically, the steps for outputting the VTE risk probability include:

[0054] For each parameter value within a sliding window, calculate the absolute deviation between the parameter value and the baseline mean, and calculate the relative deviation percentage by combining the adaptive individualized baseline normal range.

[0055] When the relative deviation percentage exceeds the upper limit of the normal range of the adaptive individualized baseline or falls below the lower limit, an abnormal deviation is marked.

[0056] By integrating the parameter values ​​of all marked abnormal deviations, a parameter deviation sequence is obtained;

[0057] For a single parameter, the duration of a single missing event is less than the time. For missing data, the weighted average of adjacent valid data is used to fill in the missing data;

[0058] For single parameter consecutive missing duration exceeding time The data segment is marked with a missing data identifier;

[0059] The final bias time series is obtained;

[0060] The deviation time series is segmented according to a preset physiological state cycle, and key dynamic fluctuation features of each time period are extracted.

[0061] By aligning the key dynamic fluctuation features, the temporal features, and the static features in the time dimension, a spatiotemporal fusion feature matrix is ​​constructed.

[0062] Specifically, the steps for outputting the VTE risk probability also include:

[0063] Historical VTE labeled data was used as the training set, weighted Focal Loss was used as the loss function, and the AdamW optimizer was used to train the time series risk prediction model.

[0064] Input the spatiotemporal fusion feature matrix of the target patient into the trained temporal risk prediction model, and output the future... Hourly VTE initial risk probability;

[0065] After dynamically calibrating by calling the historical clinical rule base, output the future... Hourly VTE risk probability.

[0066] Specifically, the steps for conducting dual-judgment intervention include:

[0067] Calculate the deviation score of each parameter trend indicator from the preset normal trend range, and weight and merge the deviation scores of all parameter trend indicators to obtain the trend risk score;

[0068] Set trend score weights and VTE risk probability weights, and calculate a dual risk score by weighted summation based on the VTE risk probability. ;

[0069] Set the maximum threshold With minimum threshold ;

[0070] Construct dual risk assessment conditions, if If so, it is classified as a Level 1 risk; if If so, it is judged as a level two risk; if If so, it is classified as a level three risk.

[0071] Specifically, the steps for conducting dual-judgment intervention also include:

[0072] After the intervention measures are implemented at the nursing end, the implementation time, the content of the measures, and the implementer are recorded in real time.

[0073] Healthcare workers can confirm the implementation status on their mobile devices and provide additional suggestions for adjustments.

[0074] By comparing before and after the implementation of intervention measures The improvement rate is calculated based on the change in the dual risk score within one hour;

[0075] If the improvement rate is greater than or equal to the preset improvement threshold, the prevention and control effect is deemed effective, and the intervention plan is recorded as an effective measure.

[0076] Otherwise, if the risk is deemed not to have been effectively controlled, the intervention measures will be adjusted and the risk level will be reassessed. If the improvement rate still does not meet the standard after adjustment, a multidisciplinary assessment process will be triggered.

[0077] The beneficial effects of this invention are:

[0078] 1. This invention utilizes multidimensional parameter acquisition and individualized baseline construction, combined with patient static characteristics and dynamic calibration thresholds based on underlying disease risks. The trend recognition module employs a dynamic sliding window, employing both slope and product bias constraints, along with coherence verification, to effectively distinguish between physiological fluctuations and pathological abnormalities, capturing the early hidden trends of VTE in target patients. The LSTM feature extraction layer of the time-series risk prediction model introduces a time-series attention mechanism, improving the dual-layer bidirectional LSTM structure. The first layer extracts the first feature of the target patient, and the second layer extracts the second feature. Residual fusion features are obtained using residual connections, and LSTM feature sequences are generated through linear transformation. These sequences are then trained and calibrated using clinical data to output the VTE risk probability. Compared to traditional fixed threshold monitoring, this invention adapts to individual differences among different target patients, identifying potential risks in advance, thereby reducing the probability of missed or false positives.

[0079] 2. This invention uses a weighted dual-risk score based on trend risk score and VTE risk probability to construct dual-risk judgment conditions and implement dual-judgment intervention, corresponding to three-level early warning and targeted intervention measures. After intervention, parameter changes are monitored in real time, and the improvement rate is calculated to evaluate the prevention and control effect. If effective, the intervention measures are recorded; if ineffective, a multidisciplinary assessment is triggered, forming a closed-loop management system with real-time dynamic monitoring throughout the process. Combined with adaptive baseline and model iterative updates, the system can promptly follow up on changes in the target patient's status, avoid intervention lag or inappropriate measures, and improve the VTE prevention and control effect. Attached Figure Description

[0080] Figure 1 This is a structural diagram of a VTE real-time monitoring and intelligent prevention system;

[0081] Figure 2This is a flowchart of constructing an adaptive individualized baseline in this invention;

[0082] Figure 3 This is a flowchart of the VTE abnormal trend identification process in this invention;

[0083] Figure 4 This is a flowchart for outputting the VTE risk probability in this invention;

[0084] Figure 5 This is a flowchart illustrating the dual-determination intervention in this invention. Detailed Implementation

[0085] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0086] Example

[0087] refer to Figures 1 to 5 As shown in the figure, this embodiment introduces a VTE real-time monitoring and intelligent prevention system, including a baseline construction module, a trend recognition module, a time-series risk prediction module, and an early warning and intervention module;

[0088] The baseline construction module utilizes medical-grade wearable sensor patches and bedside monitors to collect multidimensional VTE parameters in real time, including vital signs, coagulation parameters, and activity data. After preprocessing, a VTE parameter set is obtained, and the initial fluctuation range of each parameter is calculated using moving average and standard deviation analysis to form an initial individualized baseline. Based on the target patient's age and the static characteristics of underlying diseases, the initial individualized baseline threshold is calibrated using a random forest algorithm to construct an adaptive individualized baseline that fits the individual physiological patterns of the target patient, thereby improving the accuracy of assessment for high-risk groups and reducing the rate of missed diagnoses. Vital signs include, but are not limited to, blood oxygen saturation, heart rate, and blood pressure; coagulation parameters include, but are not limited to, platelet count and prothrombin time; activity data includes, but is not limited to, activity intensity and steps; and underlying diseases include, but are not limited to, diabetes and cardiovascular disease.

[0089] The trend recognition module is used to identify abnormal trends in VTE (Vacuum-Activated Tissue Excitation) based on an adaptive individualized baseline and a trend-sensitive detection algorithm for key parameters that require real-time monitoring, such as blood oxygen saturation, heart rate, and blood pressure. The sliding window duration is dynamically set. The VTE trend slope and VTE product skewness of each parameter within each sliding window are calculated in real time. Trend constraints and product skewness constraints are constructed based on the normal fluctuation range of the baseline. When a parameter does not satisfy both the trend constraints and product skewness constraints and the continuous offset time exceeds [a certain value], [further action is taken]. If the time frame is not met, it is determined to be a continuous abnormal offset and the time series characteristics are recorded; otherwise, it is determined to be a normal fluctuation and monitoring continues. Among them, the continuous abnormal offset is used to capture the early small offset of the parameter. The time series characteristics include VTE trend slope, duration, VTE product bias, and corresponding parameter deviation data.

[0090] The temporal risk prediction module is used to construct a temporal risk prediction model based on long short-term memory networks. It calculates the deviation values ​​between the historical sequences of various parameters of the target patient and the adaptive individualized baseline. Using the time series, temporal features, and static features constructed from the deviation values ​​as input, it learns multi-parameter co-evolution patterns, captures the hidden evolutionary patterns before VTE occurs, and outputs future... Hourly VTE risk probability;

[0091] The early warning and intervention module is used to build a risk intervention linkage mechanism. It makes a dual judgment and intervention based on parameter trend changes and VTE risk probability, and sets three levels of early warning, including mild deviation warning, moderate trend alarm, and high risk warning. The early warning information is pushed to the mobile terminal of the responsible medical staff in real time, and linked to the nursing terminal. Corresponding intervention measures are taken for different warning levels. After the intervention, parameter changes are monitored in real time to verify the prevention and control effect and avoid missing the early VTE prevention and control window.

[0092] Specifically, the steps for establishing an initial individualized baseline include:

[0093] The VTE parameter set is divided into high-frequency dynamic parameters and low-frequency static parameters by classifying the parameters by frequency characteristics, so as to avoid interference analysis from data of different frequencies. Among them, high-frequency dynamic parameters include vital signs and activity data, while low-frequency static parameters include coagulation indicators.

[0094] Using the timestamps of high-frequency dynamic parameters as a benchmark, linear interpolation is used to supplement the missing time point data of low-frequency static parameters. For example, in the time interval between two adjacent collections of coagulation indicators, the values ​​are smoothly filled according to the proportion of the middle time point to the interval between two collections, ensuring that all parameters form a continuous sequence on the same time axis. Finally, a time-aligned classification parameter sequence is obtained, which includes high-frequency dynamic parameters and low-frequency static parameters.

[0095] Based on the classification parameter sequence, the mean of each parameter is calculated during the initial stabilization period as a temporary baseline value for each parameter. For example, the temporary baseline value of heart rate is equal to the mean heart rate without interference within a fixed time period. In this embodiment, the fixed time period is set to 24 hours, and the initial stabilization period is the data segment without interference in the first 24 hours.

[0096] A linear regression model trained on clinical data is used to correct temporary baseline values. The temporary baseline values ​​are taken as input, and the output is a correction coefficient used to adjust the temporary baseline values. The product of the correction coefficient and the baseline parameter value is calculated to obtain the correction magnitude at the parameter level. This correction magnitude is applied to the temporary baseline values, such as through addition or multiplication, to obtain the corrected parameter baseline value for each time point. The corrected parameter baseline values ​​for all time points are arranged in chronological order to obtain the final sequence of corrected parameter baseline values. In this embodiment, the baseline parameter value is the baseline magnitude for the correction operation, and the baseline parameter value is set to 1.

[0097] Differentiated window segmentation strategies are adopted for different types of parameters. For high-frequency dynamic parameters, a sliding window method is used, in which a 5-minute window slides continuously along the time axis to ensure that the covered parameter values ​​reflect short-term fluctuations. For low-frequency static parameters, a fixed window method is used, in which the monitoring period is divided into multiple non-overlapping windows of equal length to reduce the impact of data sparsity caused by low sampling frequency on statistical results. In this embodiment, the sliding window length is set to 5 minutes and the fixed window length is set to 24 hours. However, those skilled in the art can make adaptive adjustments according to the monitoring equipment, clinical scenario, or parameter type. This invention does not limit these adjustments.

[0098] For each parameter within a sliding window, the moving mean and standard deviation are calculated. Using the baseline values ​​of each corrected parameter in the corrected parameter baseline value sequence as the center, the initial fluctuation range of each corrected parameter baseline value is determined. For example, using the heart rate corrected parameter baseline value as the center value, combined with the heart rate moving mean and standard deviation calculated within the sliding window, the lower limit 'a' of the initial heart rate fluctuation range is calculated as the heart rate corrected parameter baseline value minus q times the standard deviation, and the upper limit 'b' is the heart rate corrected parameter baseline value plus q times the standard deviation; where q is a coefficient set based on clinical experience or statistical significance requirements.

[0099] The correlation between high-frequency dynamic parameters is calculated using the Pearson correlation coefficient. A correlation threshold is set based on the characteristics of clinical physiological parameters, the accuracy requirements of equipment monitoring, and the historical correlation distribution characteristics. For example, the correlation threshold is set to 0.7. If the correlation exceeds the correlation threshold, it is determined to be a strong correlation, and strongly correlated parameter pairs are selected; otherwise, it is determined to be a weak correlation, and weakly correlated parameter pairs are removed.

[0100] For strongly correlated parameter pairs, the initial fluctuation range is corrected based on the correlation direction to obtain a high-frequency correction range. For example, when the actual value of the activity intensity parameter exceeds the upper limit of its own initial range, the upper limit of the initial range of heart rate associated with the activity intensity parameter is automatically adjusted upward to avoid misjudging the physiological increase in heart rate after activity as abnormal. In this embodiment, the correlation direction is used to represent the physiological driving relationship between strongly correlated parameter pairs, including positive correlation driving and negative correlation driving, reflecting the objective physiological law that a change in one parameter causes a change in the same direction or direction of another parameter.

[0101] For low-frequency static parameters, such as platelet count (a coagulation indicator), a deviation threshold is set based on historical abnormal fluctuation characteristics and industry expert experience. For example, the deviation threshold is set to 80% of the lower limit of the normal reference range. If the low-frequency static parameter value is lower than the deviation threshold, the associated vital sign range is contracted upward to improve the sensitivity to pathological states; otherwise, the high-frequency correction range is maintained, ultimately obtaining a multi-parameter associated correction fluctuation range. In this embodiment, the normal reference range is determined to be the 95% medical reference range of the corresponding coagulation indicator in healthy individuals based on clinical medical guidelines and population health statistics.

[0102] An improved quartile method is used to mark values ​​in the VTE parameter set that exceed the corrected fluctuation range as outliers, remove outliers, calculate the standard deviation of the remaining effective parameters, and update the corrected fluctuation range to form an initial individualized baseline containing the corrected fluctuation range of each parameter.

[0103] For example, taking platelet count as an example, suppose the target patient's historical platelet count parameter set is [180,190,170,160,150,80,175,185]× The first quartile is calculated to be 160, the third quartile to be 185, and the interquartile range to be 25. The interquartile range is multiplied by 1.5 to obtain the abnormal judgment range as (160-37.5, 185+37.5), i.e. (122.5, 222.5). 80 of the historical platelet count parameters are lower than 122.5 and are marked as outliers and removed. The mean is calculated to be 175 and the standard deviation is 12 using the remaining effective parameters. The mean is updated and the fluctuation range is corrected by adding or subtracting twice the standard deviation to be (151, 199), which is used as the individualized baseline range for platelet count. This is then integrated with the corrected fluctuation ranges of other parameters to form the initial individualized baseline for the target patient.

[0104] Specifically, the steps for constructing an adaptive individualized baseline include:

[0105] Static characteristics of the target patient, such as age, type of surgery, and history of underlying diseases, are obtained from the electronic medical record system. These static characteristics are then quantified to obtain a quantified static feature vector. Key statistical features are extracted from the initial individualized baseline, including the upper and lower limits of the corrected fluctuation range, the range width, the moving mean, and the standard deviation of each parameter. This forms a correlation matrix between the parameters and the statistical features, reflecting the stability and fluctuation characteristics of the individual's current physiological parameters.

[0106] Based on the correlation matrix, the static feature vector is concatenated with the key statistical features of the initial individualized baseline to generate the individual feature vector of the target patient;

[0107] Based on individual feature vectors and combined with the random forest algorithm, the threshold adjustment range of various parameters is obtained, such as increasing the upper limit of platelet count. ;

[0108] For example, the target patient is a 65-year-old woman with a history of diabetes and mild renal dysfunction. Her individual feature vector includes age, gender, history of diabetes, baseline values ​​of renal function indicators, and historical medication history. This individual feature vector is input into a trained random forest model. By learning the correlation between individual features and the normal fluctuation range of various physiological parameters in historical patient data, the model outputs the threshold adjustment range for various parameters specific to the target patient. For instance, if the random forest model determines that the target patient's platelet count is affected by renal function and its normal fluctuation range is theoretically higher than that of the general population, then the upper limit of the output platelet count will be increased. 2×10 9 / L; meanwhile, the adjustment range of the upper limit of D-dimer output by the random forest model was increased by 0.1 mg / L, and the adjustment range of the lower limit of heart rate was decreased by 5 beats / minute;

[0109] Based on clinically recognized risk factors associated with underlying diseases, and combined with static feature vectors, clinical experts determine risk assignments according to the degree of influence of these risk factors on thrombosis. For example, a history of VTE is assigned 2 points, hypertension is assigned 1 point, and no relevant risk factors are assigned 0 points. The various risk factors included in the static feature vectors of target patients are statistically analyzed, and the total risk score is obtained by summing them. Based on the correlation between the total risk score and the probability of thrombosis based on large-sample clinical data, the total risk score of the target patient is mapped to the probability of underlying disease risk. For example, a total risk score of 0-2 corresponds to a probability of 0-10% for underlying diseases, 3-5 corresponds to a probability of 11%-30% for underlying diseases, and 6 or above corresponds to a probability of 31% or above for underlying diseases. Based on clinical risk... The risk grading standards and industry expert experience are used to set probability thresholds. If the probability of underlying disease risk exceeds the threshold, the patient is classified as high-risk. Risk weighting coefficients are set according to the underlying disease risk probability grading, including: 1.2 when the underlying disease risk probability is 11%-30%; 1.5 when the underlying disease risk probability is 31%-50%; and 1.8 when the underlying disease risk probability exceeds 50%. The product of the threshold adjustment range for each parameter and the risk weighting coefficient is calculated to obtain the calibration threshold adjustment range for high-risk patients. Otherwise, the patient is classified as a medium-risk patient, and the threshold adjustment ranges for each parameter are maintained. Underlying disease risk-related factors include, but are not limited to, hypertension, diabetes, and a history of thrombosis.

[0110] The calibration threshold adjustment range is compared with the preset safety boundary. If it exceeds the preset safety boundary, it is truncated to the boundary value to obtain the initial individualized baseline threshold constrained by the boundary, ensuring that the threshold values ​​of each parameter conform to common clinical physiology and avoiding interference from extreme values; otherwise, the calibration threshold adjustment range is maintained. Specifically, the statistical distribution of each physiological parameter is calculated based on large sample clinical data to determine the (95%, 99%) confidence interval as the initial safety boundary. The initial safety boundary is then corrected by combining clinical guidelines and expert knowledge to remove statistical outliers and non-clinically reasonable intervals. The initial safety boundary is then calibrated a second time according to the monitoring range and error characteristics of the equipment to obtain the preset safety boundary.

[0111] Based on the target patient's current physiological state, such as exercise, sleep, and recovery period, the initial individualized baseline threshold is dynamically adjusted to obtain a dynamic threshold baseline with multiple states, such as increasing the upper limit of heart rate during exercise and decreasing the lower limit of blood pressure during sleep.

[0112] Physiological adaptations, such as a decrease in resting heart rate due to long-term rehabilitation training, can be distinguished by sliding window mean and trend test methods; pathological trends, such as a continuous decrease in platelets after chemotherapy, can be identified by abnormal fluctuation detection and persistence analysis.

[0113] For physiological adaptation, the slope is obtained by linearly fitting the time-series parameter values ​​within the sliding window. The slope is divided by the window duration to obtain the change per unit time. The trend change rate is calculated by comparing the change per unit time with the mean value of the parameters within the sliding window. The dynamic threshold baseline is then adaptively adjusted according to the trend change rate and the principle of proportional adaptation. For example, if blood pressure decreases due to reasonable weight control and the change rate is negative, it is adjusted by 40% of the change rate to avoid over-adjustment leading to misjudgment.

[0114] For pathological trends, the key dynamic fluctuation features, trend change rate, and static feature vector within the current sliding window are input into the already trained randomized training model. The model learns the mapping relationship between pathological trends and threshold adjustment magnitudes using historical patient data, and outputs corresponding suggested threshold adjustment values. The dynamic threshold baseline is iteratively updated based on these suggested values. For example, when the randomized training model determines that platelet counts show a continuously decreasing pathological trend, a positive suggested threshold adjustment value is output, expanding the monitoring range of the dynamic threshold baseline. When the randomized training model determines that inflammatory indicators show an improving trend, a negative suggested threshold adjustment value is output, narrowing the monitoring range of the dynamic threshold baseline. This allows the threshold to be adjusted accordingly as the patient's physiological state deteriorates or improves, improving sensitivity and adaptability to pathological changes. Finally, the threshold results of all states and trends are integrated to construct an adaptive individualized baseline. The already trained randomized training model includes an input layer, a feature selection layer, a decision tree ensemble layer, and an output layer.

[0115] The input layer is used to receive and normalize the key dynamic fluctuation features, trend change rate, and static feature vectors within the sliding window;

[0116] The feature selection layer is used to calculate the information entropy of the target variable and the conditional entropy of the target variable under the condition of feature values, with pathological trend as the target variable. The information gain of each candidate feature is obtained by subtracting the conditional entropy from the information entropy, which measures the information contribution of the feature to the target variable. According to the statistical results of historical VTE monitoring data, when the information gain of high-contribution features is greater than 0.1 and the information gain of low-contribution features is less than 0.1, the gain threshold is set to 0.1, and low-contribution features with information gain lower than the gain threshold are eliminated. The Gini coefficient of the remaining features is obtained by subtracting the sum of squares of the proportion of each category of samples from the value 1. According to the sample purity requirements, when the accuracy of distinguishing pathological trend samples from normal trend samples exceeds 85% and the feature discrimination accuracy exceeds 80%, the Gini coefficient threshold is set to 0.4, retaining features with Gini coefficient lower than the Gini coefficient threshold, eliminating redundant features and noisy features, and generating a subset of high-contribution features.

[0117] The decision tree ensemble layer is used to construct multiple independent decision trees. Each decision tree is trained in parallel based on a subset of high-contribution features, learning the nonlinear mapping relationship between pathological trends and threshold adjustment magnitudes. After training, a weighted voting strategy is used to fuse the outputs of each decision tree, reducing the risk of overfitting of a single decision tree and improving the robustness and generalization ability of the model. The weighted voting strategy assigns corresponding weights based on the prediction accuracy of a single decision tree on the validation subset. The higher the prediction accuracy, the larger the weight. The final threshold adjustment magnitude is the average of the weighted sum of the outputs of each decision tree.

[0118] The output layer is used to output a suggested threshold adjustment range and a confidence level that match the current pathological trend based on the fusion results of the decision tree ensemble layer. The confidence level threshold is set differently according to different VTE risk levels. When the confidence level of the suggested value is lower than the confidence level, a secondary feature input verification is triggered to ensure the accuracy of the output results. Specifically, the confidence level threshold is set to 0.7 for low-risk scenarios, 0.8 for medium-risk scenarios, and 0.9 for high-risk scenarios.

[0119] The training set is composed of labeled patient historical physiological index time series data and corresponding threshold adjustment records, which are divided into training subset and validation subset in an 8:2 ratio. The mean squared error loss function is used as the loss function, and the random training model is trained in combination with the SGD optimizer. After training, the key dynamic fluctuation features, trend change rate, and static feature vectors of the current sliding window after input layer preprocessing are input, and the suggested value and confidence level of the threshold adjustment range under the corresponding pathological trend are output.

[0120] For example, taking the pathological trend analysis of platelet markers in a high-risk VTE scenario for target hospitalized patients as an example, the input layer receives and normalizes the key dynamic fluctuation features of platelets in the target patient over 24 hours, the trend change rate, and the static feature vector; wherein, the key dynamic fluctuation features of platelets include platelet count fluctuation values. / L, fluctuation frequency 3 times / 6h; trend change rate including platelet count decrease per hour. / L; Static feature vector includes age 65 years, postoperative day 3, and body mass index. ;

[0121] The feature selection layer uses whether a sustained decrease in platelets indicates a pathological trend in VTE as the target variable. The information entropy of the target variable is calculated to be 0.98, and the conditional entropy of the candidate feature of platelet trend change rate is calculated to be 0.05. The resulting information gain is 0.98 - 0.05 = 0.93, which is greater than the gain threshold of 0.1. The information gain of the body mass index feature is calculated to be 0.08, and features below the gain threshold of 0.1 are discarded. The Gini coefficient is calculated for the remaining features, including platelet fluctuation, platelet trend change rate, and postoperative days. The Gini coefficient for the platelet trend change rate is calculated as: 1 - (proportion of samples with pathological trends). +Percentage of samples exhibiting normal trends The value is calculated as 1 - (0.49 + 0.09) = 0.42, which is greater than the Gini coefficient threshold of 0.4, so it is excluded. The Gini coefficient for platelet fluctuation is calculated as 1 - ( + =1-(0.64+0.04)=0.32, which is less than the Gini coefficient threshold of 0.4, so it is retained, and finally a high-contribution feature subset containing platelet fluctuation value, postoperative days and age is generated;

[0122] The decision tree ensemble layer constructs 10 independent decision trees, each trained in parallel based on a subset of high-contribution features. The first decision tree has a prediction accuracy of 92% on the validation subset, a weight of 0.12, and a suggested threshold adjustment value of +0.8. The second decision tree has a prediction accuracy of 88%, a weight of 0.11, and a suggested threshold adjustment value of +0.7. The suggested threshold adjustment values ​​of the remaining 8 decision trees range from +0.6 to +0.9. These values ​​are then weighted and summed (0.12×0.8 + 0.11×0.7 + … + 0.09×0.9) to obtain a final suggested threshold adjustment value of +0.75 with a confidence level of 0.92. The suggested threshold adjustment value of +0.8 indicates a positive adjustment, expanding the baseline of platelet monitoring thresholds.

[0123] Since the target patients are in a high-risk VTE scenario, the output layer suggests setting the confidence threshold to 0.9. If 0.92 ≥ 0.9, then the secondary verification will not be triggered, and the suggested threshold adjustment range + 0.75 and the confidence of the suggested value of 0.92 will be directly output.

[0124] The training set for the randomized training model consisted of 10,000 labeled historical data of VTE patients, including 8,000 training subsets and 2,000 validation subsets. Training was completed by combining the mean squared error loss function and the SGD optimizer. The mean squared error loss function loss value converged from the initial 1.2 to 0.08, and the learning rate of the SGD optimizer was set to 0.01.

[0125] After inputting the preprocessed feature vector, the final output is a suggested threshold adjustment value of +0.75, which adjusts the target patient's platelet monitoring threshold baseline from... / L adjusted to / L, enabling sensitive monitoring of a continuous decline in platelet count.

[0126] Specifically, the trend recognition module includes a trend analysis unit and an anomaly recognition unit;

[0127] The trend analysis unit is used for real-time dynamic monitoring and feature extraction of key parameters. Using an adaptive individualized baseline as a reference, it introduces a trend-sensitive detection algorithm to identify abnormal VTE trends and dynamically sets the sliding window duration. The window slides in real time over time. Within each sliding window, the unit calculates the VTE trend slope of the parameter through a linear regression model, reflecting the rate of change of the parameter within the window period, and calculates the VTE product bias of the parameter deviating from the normal fluctuation range of the adaptive individualized baseline, quantifying the total degree of deviation of the parameter from the adaptive individualized baseline.

[0128] The anomaly detection unit is used to determine the slope and offset based on the VTE trend slope and VTE product skewness of the parameters output by the trend unit, and constructs trend constraints and product skewness constraints based on the normal fluctuation range of the adaptive individualized baseline. When the parameter changes within the sliding window do not simultaneously satisfy the trend constraints and product skewness constraints, and the continuous offset duration exceeds [a certain value], the anomaly detection unit will detect the anomaly. If the parameter shows a persistent abnormal shift, it indicates an early signal of potential pathological changes, and the time-series characteristics are recorded; otherwise, it indicates that the parameter change is within the individual's physiological regulatory capacity and is judged as a normal fluctuation within the physiological range, and monitoring continues.

[0129] Specifically, the steps for identifying abnormal VTE trends include:

[0130] For key parameters monitored in real time, the base duration of the sliding window is set according to the physiological fluctuation frequency of the parameter; for example, for blood oxygen saturation, which is normally stable but suddenly becomes abnormal, the base duration is set to... Minutes; Heart rate is sensitive to short-term fluctuations, so the base duration is set to [minutes]. Minutes; blood pressure fluctuates mainly in the medium term, with the baseline duration set to [missing information]. minutes; among them, ;

[0131] Combine adaptive individualized baseline fluctuation characteristics to adjust the sliding window duration, and statistically... Using adaptive individualized baseline fluctuation data for each day, such as the baseline mean of each parameter and the upper and lower limits of the normal fluctuation range, the ratio of the standard deviation to the mean is calculated to obtain the baseline fluctuation coefficient. Based on the requirements for stability of clinical physiological parameters, the physiological rationality and sensitivity of vital sign monitoring, the historical baseline fluctuation distribution characteristics, the parameter fluctuation range distribution pattern formed by long-term monitoring of similar patient groups, and the monitoring accuracy of the equipment, the following considerations are taken into account. At that time, set the maximum fluctuation limit. With minimum fluctuation limit ,like If the fluctuation is severe, it is determined by calculating the preset first weight. The sliding window duration is determined by multiplying the base duration by the base duration. ,like Suitable for scenarios with drastic short-term heart rate fluctuations; if If the fluctuation is stable, it is determined by calculating the preset second weight. The sliding window duration is determined by multiplying the base duration by the base duration. ,like This is suitable for scenarios where blood pressure fluctuates steadily in the middle range; specifically, this embodiment sets... Preset first weight Based on the degree of parameter fluctuation Exceeding The magnitude and short-term change response speed are determined by the clinical requirement that abnormal events be identified in the shortest possible time; a second weight is preset. Based on the stability of parameter fluctuations Below The magnitude and medium-term trend of change characterize the demand to accurately reflect the slow change trend of parameters over three consecutive hours.

[0132] Using timestamps within a sliding window as independent variables and parameter values ​​as dependent variables, a weighted linear regression algorithm is introduced. A time decay weight is assigned to each parameter value, and the VTE trend slope is obtained by fitting the data using weighted least squares. This includes mapping each timestamp within the sliding window to a continuous time series independent variable. The corresponding parameter values ​​are used as dependent variables. Set time decay weights based on the interval between the data acquisition time and the current time. Time decay weight Substitute into the weighted least squares method to construct the loss function. The intercept is solved by minimizing the loss function. With slope slope Let VTE be the trend slope, where, , This represents the total number of valid parameter data points contained within the sliding window.

[0133] For each parameter value within the sliding window, the normal range of the adaptive individualized baseline corresponding to the matched timestamp is calculated. If the parameter value exceeds the normal range, the difference between the parameter value and the midpoint of the adaptive individualized baseline normal range is calculated, and then combined with the time decay weight. We calculate the effective deviation by weighting; otherwise, the effective deviation is zero. We sum all the effective deviations to obtain the VTE integral bias, which is used to quantify the effective cumulative degree of VTE parameter deviation from the adaptive individualized baseline, and avoid misjudging fluctuations within the normal range as VTE integral bias.

[0134] Based on the historical characteristics of the target patients, and according to the range of fluctuations in the patients' historical physiological parameters, the target population is considered as healthy individuals. Reference interval The clinical diagnosis and treatment stages are the postoperative recovery period and the long-term bed rest period. The evolution of VTE risk shows that parameters in the pre-thrombotic phase exhibit a slow upward trend, while parameters in the acute thrombotic phase exhibit a rapid upward trend. A slope limit is set at [value missing]. Units per minute are used to construct trend constraints, including: if the VTE trend slope is less than the slope limit, the trend is considered normal; otherwise, it is considered a suspected abnormal trend.

[0135] Based on the historical deviation characteristics of the target patients, when the maximum deviation of the historical parameters from the baseline is 1.5 times the normal range and the parameter deviation is positively correlated with the probability of VTE diagnosis, the product deviation constraint value is set to... Units per minute, construct the product bias constraint, including: if the VTE product bias is less than the product bias constraint value, it is determined that the offset is normal; otherwise, it is determined that the offset is suspected to be abnormal.

[0136] If both the trend constraint and the product constraint are satisfied, the fluctuation is considered normal. If only one constraint is satisfied, it is marked as to be observed, and a high-density monitoring mode is activated. The sliding window duration is halved, the parameter sampling frequency is increased, and the frequency of anomaly detection is increased, thereby enhancing the ability to capture subtle changes in parameters in real time.

[0137] If neither the trend constraint nor the product bias constraint is satisfied, then a consistency verification is performed. If the duration of continuous offset exceeds [a certain value], then [the verification is performed]. If the previous sliding window is to be observed, it indicates an early signal of potential pathological changes, is judged as a continuous abnormal shift, and its temporal characteristics are recorded; otherwise, it is judged as a transient fluctuation abnormality, and the next sliding window is monitored.

[0138] Specifically, the steps for outputting the VTE risk probability include:

[0139] For each parameter value within a sliding window, the absolute deviation between the parameter value and the baseline mean is calculated, and the absolute deviation is divided by the width of the normal range of the adaptive individualized baseline to obtain the relative deviation percentage, which is used to measure the relative degree to which the parameter deviates from the adaptive individualized baseline.

[0140] If the relative deviation percentage exceeds the upper limit of the normal range of the adaptive individualized baseline or falls below the lower limit, the parameter value is determined to be outside the normal range of the adaptive individualized baseline, an abnormal deviation is marked, and the duration of the deviation is recorded; otherwise, the parameter value is determined to be within the normal range of the adaptive individualized baseline; all parameter values ​​marked with abnormal deviations are integrated to obtain the parameter deviation sequence.

[0141] Align the initial parameter deviation sequence according to the timestamp, and for single parameters, the duration of a single missing parameter is less than the time. For missing data, the weighted average of adjacent valid data is used for filling. The total adjacent interval is obtained by calculating the sum of the time intervals between the two valid data points before and after the missing point. Taking the midpoint of the missing time period as the reference, the weight of the valid data point before the missing point is the inverse ratio of the time interval between the missing point and the previous valid data point to the total adjacent interval, and the weight of the valid data point after the missing point is the inverse ratio of the time interval between the missing point and the next valid data point to the total adjacent interval. For single-parameter consecutive missing durations exceeding the specified time... The data segments are marked with missing data indicators to avoid errors caused by a single filling method; finally, the deviation time series is obtained; where, the single missing duration is the duration of only one uninterrupted missing segment of the same parameter, and the single parameter continuous missing duration is the total duration of multiple missing segments spliced ​​together or a single missing segment continuing uninterruptedly for the same parameter.

[0142] For the deviation time series, the system segments the data according to a preset physiological state cycle and extracts key dynamic fluctuation features for each time period, including but not limited to the deviation mean, variance, percentage of abnormal events, and longest consecutive abnormal duration. A feature splicing layer aligns the key dynamic fluctuation features, temporal features, and static features in the time dimension to construct a spatiotemporal fusion feature matrix. The core clinical stage is determined based on the patient's treatment scenario, matching the clinical fluctuation patterns of monitored indicators, such as D-dimer peaking 24-48 hours post-surgery and gradually declining after 72 hours. The cycle length is calibrated in conjunction with high-risk periods for VTE, such as using a fine-grained 2-hour sub-cycle in the acute phase and a coarse-grained 4-hour sub-cycle in the recovery phase. Finally, the preset physiological state cycle, adapted to the risk evolution characteristics and dynamic changes of indicators, is determined through validation and dynamic fine-tuning using historical patient data.

[0143] A temporal risk prediction model is constructed based on a long short-term memory network, including an input layer, an LSTM feature extraction layer, an attention weighting layer, and an output layer.

[0144] The input layer receives the spatiotemporal fusion feature matrix, normalizes the key dynamic fluctuation features, time series features and static features of the input, maps the features to a unified distribution interval, and adapts them by combining the length of the deviation time series to obtain a normalized time series sequence.

[0145] The LSTM feature extraction layer introduces a temporal attention mechanism to improve the two-layer bidirectional LSTM structure. The first layer extracts the first feature of the target patient to capture the short-term fluctuation pattern of parameters, and the second layer extracts the second feature of the target patient to capture the hidden evolution pattern before VTE occurs. The output of the first layer and the input of the second layer are directly concatenated using residual connections to obtain residual fusion features. LSTM feature sequences are generated through linear transformation to assess the VTE risk of the target patient, thereby improving the accuracy of VTE risk identification and early warning capability.

[0146] The attention weighting layer introduces a dual attention mechanism of time and features. It calculates the temporal attention weight for each time step through a fully connected layer and softmax, and calculates the feature attention weight for each feature dimension through a convolutional layer and softmax. The temporal attention weight, feature attention weight and LSTM feature sequence are weighted and fused to obtain the key temporal feature vector.

[0147] The output layer maps key temporal feature vectors to risk scores, and the sigmoid activation layer converts the risk scores into risk probability values.

[0148] Historical VTE labeled data is obtained and set as the training set. The training set, validation set, and test set are divided according to time order and preset ratio. Weighted Focal Loss is set as the loss function, and AdamW optimizer is used to train the time series risk prediction model. In this embodiment, the preset ratio is set to 7:2:1.

[0149] The spatiotemporal fusion feature matrix of the target patient is input into the trained time-series risk prediction model. After input layer normalization, LSTM feature extraction layer, attention weighting layer, and output layer calculation, the future risk prediction model is output. Hourly VTE initial risk probability;

[0150] By accessing the historical clinical rule base, including risk enhancement and risk reduction rules, the initial risk probability of VTE is dynamically calibrated to obtain the calibrated future risk probability. Hourly VTE risk probability; where risk enhancement rules include, but are not limited to, if the surgical recovery period is less than If the relative deviation of activity is less than the preset relative deviation threshold, the initial risk probability of VTE is increased; if there is a history of VTE and no anticoagulants are used, the initial risk probability of VTE is increased; the risk weakening rules include, but are not limited to, if the patient regularly uses therapeutic doses of anticoagulants, the initial risk probability of VTE is decreased; if the patient is a non-surgical patient and has no underlying disease, the initial risk probability of VTE is decreased. The specific size can be set by those skilled in the art according to actual needs, and this embodiment does not limit it.

[0151] For example, taking the VTE risk prediction scenario after orthopedic surgery as an example, the input layer of the time-series risk prediction model receives a spatiotemporal fusion feature matrix, where the static features include the patient's age of 65 years, history of major orthopedic surgery, no history of VTE, and irregular use of anticoagulants, and the dynamic fluctuation features include the mean D-dimer deviation of 82.86%, the abnormal frequency ratio of 75%, and the longest continuous abnormal duration of 3 hours, the mean lower limb circumference difference deviation of 170%, the abnormal frequency ratio of 100%, and the longest continuous abnormal duration of 4 hours.

[0152] The input layer normalizes the spatiotemporal fusion feature matrix, mapping the features to the [0,1] interval to obtain a normalized temporal sequence. The LSTM feature extraction layer uses a two-layer bidirectional LSTM structure, combined with residual connections, to capture the hidden evolution patterns of short-term fluctuations in D-dimers and lower limb circumference differences, generating an LSTM feature sequence. The attention weighting layer introduces a dual attention mechanism of time and features, calculating the time attention weights [0.1, 0.15, 0.2, 0.25, 0.3] and the feature attention weights [0.3, 0.2, 0.15, 0.15, 0.2], and outputs a key temporal feature vector after weighted fusion. The output layer maps the key temporal feature vector to an initial risk probability of 0.82 using a Sigmoid activation function.

[0153] Annotated data from 1000 post-orthopedic surgery patients were divided into 700 training sets, 200 validation sets, and 100 test sets according to a pre-defined ratio of 7:2:1. A temporal risk prediction model was trained using the AdamW optimizer with an initial learning rate of 0.001 combined with a weighted Focal Loss function. The training output predicted the initial risk probability as follows: given the spatiotemporal fusion feature matrix of patient A, the initial risk probability was 0.82. Validation on the validation set yielded an accuracy of 91%, a recall of 89%, and an F1 score of 0.90. These results met industry standards of ≥85% accuracy and ≥85% recall. A historical clinical rule base was then used to match patient A's characteristics: 2 days post-surgery, a relative deviation of 80% in activity level, and irregular use of anticoagulants. This triggered a risk enhancement rule, increasing the initial risk probability. The final calibration yielded a VTE risk probability of 0.88 for the next 24 hours. The weighted Focal Loss function... The category balance weight for Loss was set to 0.75, the difficulty sample adjustment weight was set to 2, the surgical recovery period threshold for the risk enhancement rule was set to 3 days, and the activity relative deviation threshold was set to 50%.

[0154] Specifically, the steps for conducting dual-judgment intervention include:

[0155] By combining the real-time vital signs of target patients with static VTE risk factors, a risk intervention linkage mechanism is constructed. For each trend indicator, including VTE trend slope and VTE skewness, normal ranges are set for each trend indicator based on clinical guidelines and historical data. For example, the normal range for VTE trend slope is (…). The normal range for VTE product bias is ( The current trend indicator value is compared with the normal range. If the current trend indicator value is within the normal range, the deviation score is 0; otherwise, the absolute difference between the current trend indicator value and the boundary of the normal range is calculated. The absolute difference is divided by the width of the normal range to obtain the normalized deviation ratio. The normalized deviation ratio is mapped to a preset scoring interval to obtain the deviation score. The deviation scores of all parameter trend indicators are weighted and merged to obtain the trend risk score, which reflects the risk level of parameter evolution. Among them, VTE static risk factors include, but are not limited to, BMI, previous history of thrombosis, surgical type and duration, varicose veins and hereditary thrombosis tendency, with a preset scoring interval of 0 to 10.

[0156] Trend score weights and VTE risk probability weights are set according to clinical importance. A dual risk score is calculated by weighted summation based on the VTE risk probability. ;

[0157] The highest threshold is set based on clinical expert experience. With minimum threshold Construct dual risk assessment conditions, if If so, it is classified as a Level 1 risk; if If so, it is judged as a level two risk; if If so, it is classified as a level three risk;

[0158] A tiered early warning and intervention system is implemented. For Level 1 risk, a high-risk alert is immediately issued, triggering an emergency assessment process, such as multidisciplinary consultation and requesting an ultrasound examination; anticoagulation interventions are implemented, such as initiating intravenous anticoagulation. For Level 2 risk, a moderate trend alert is issued, generating recommendations for adjusting anticoagulation medication, such as upgrading from a preventative dose to a therapeutic dose, and calculating the dose based on renal function; physical interventions are implemented, such as extending the use of intermittent pneumatic compression devices. For Level 3 risk, a mild deviation alert is triggered, and individualized activity guidance plans are pushed out, such as setting the number of ankle pump exercises per hour and the duration of bedside activities based on the target patient's physical condition; dietary interventions are implemented, such as increasing fluid intake.

[0159] After the nursing staff implements the intervention, they record the implementation time, the content of the measures, and the implementer in real time; the medical staff confirm the implementation status on the mobile device and add adjustment opinions, such as adjusting the anticoagulation dose due to the bleeding risk of the target patient.

[0160] Comparison before and after the implementation of intervention measures The changes in dual risk scores, trend indicators, and the decrease in VTE risk probability within an hour were analyzed. The difference between the pre-intervention dual risk scores and the post-intervention dual risk scores was calculated, and the ratio of this difference to the pre-intervention dual risk scores was calculated to obtain the improvement rate. For example, [the following is an example]. The specific size can be set by those skilled in the art according to actual needs, and this embodiment does not limit it.

[0161] Based on the experience of clinical experts and clinical prevention and control standards, an improvement threshold is set. If the improvement rate is greater than or equal to the improvement threshold, the prevention and control effect is deemed effective, and the intervention plan is recorded as an effective measure. Otherwise, the risk is deemed not to have been effectively controlled, and the intervention measures need to be adjusted, such as changing anticoagulants, strengthening physical prevention, and reassessing the risk level. If the improvement rate still does not meet the standard after adjustment, a multidisciplinary assessment process needs to be triggered, such as consultation with vascular surgery and hematology departments.

[0162] In summary, this invention proposes a real-time monitoring and intelligent prevention system for VTE. By constructing a four-dimensional closed-loop architecture encompassing baseline, identification, prediction, and intervention, it achieves intelligent management from individualized physiological modeling to early warning and prevention. Multidimensional VTE parameters of target patients, including vital signs, coagulation indicators, and activity data, are collected via medical-grade wearable sensor patches and bedside monitors. These parameters are preprocessed to form a VTE parameter set. Based on the frequency characteristics of different parameters, the system employs a classification and alignment strategy. For high-frequency dynamic parameters, including vital signs and activity data, a sliding window method is used; for low-frequency static parameters, linear interpolation is used to supplement missing values, ensuring temporal consistency. Combining moving mean and standard deviation analysis, the initial fluctuation range of each parameter is calculated, and Pearson correlation coefficient is introduced to identify strongly correlated parameter pairs, such as activity intensity and heart rate, dynamically correcting fluctuation boundaries to avoid errors due to physiological fluctuations. The system integrates static features such as the target patient's age and underlying diseases, uses a random forest algorithm to calibrate the initial baseline, and constructs an adaptive individualized baseline, effectively improving the accuracy of assessment for high-risk groups. The trend recognition module, based on the adaptive individualized baseline, employs a trend-sensitive detection algorithm, dynamically sets the sliding window duration, and calculates the VTE trend slope and VTE product skewness in real time. Through a dual judgment mechanism of trend constraint and product skewness constraint, it identifies continuous abnormal shifts and records temporal characteristics. The temporal risk prediction module constructs a temporal risk prediction model based on LSTM, inputs the deviation time series and temporal characteristics, learns the multi-parameter co-evolution law, and outputs the future... The system calculates the hourly VTE risk probability and dynamically calibrates it using a clinical rule base to enhance predictive interpretability. The early warning and intervention module establishes a risk linkage mechanism, using a dual assessment of trend risk scores and VTE risk probabilities. It sets three response levels: mild deviation warning, moderate trend alarm, and high-risk alert. Warning information is pushed to healthcare workers' mobile devices in real time, linking with nursing staff to implement individualized interventions. After intervention, the system continuously monitors parameter changes and calculates the improvement rate by comparing changes in dual risk scores before and after intervention, thus evaluating the prevention and control effect. This forms a complete closed loop of monitoring, early warning, intervention, and feedback, effectively avoiding the problems of reliance on manual assessment and delayed response in traditional VTE prevention and control, and significantly improving the early VTE risk identification rate and intervention timeliness.

[0163] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A VTE real-time monitoring and intelligent prevention system, characterized in that, include: Baseline construction module, trend identification module, time-series risk prediction module, and early warning and intervention module; The baseline construction module is used to monitor and collect multidimensional VTE parameters in real time, calculate the normal fluctuation range of each parameter to form an initial individualized baseline, and construct an adaptive individualized baseline through threshold calibration. This includes: acquiring the static characteristics of the target patient, quantifying and generating a static feature vector, and extracting key statistical features to construct an association matrix. The static feature vector is concatenated with the key statistical features to generate an individual feature vector, and the threshold adjustment range of each parameter is calculated. Calculate the risk probability of the target patient's underlying disease; When the probability of underlying disease risk exceeds a preset probability threshold, the patient is identified as a high-risk patient. The threshold adjustment range is calibrated by setting a risk weighting coefficient according to the risk level, and the calibrated threshold adjustment range is obtained. For calibration threshold adjustments that exceed the preset safety boundary, the initial individualized baseline threshold is obtained by truncating to the boundary value. Based on the current physiological state of the target patient, the initial individualized baseline threshold is dynamically adjusted to obtain a dynamic threshold baseline; Differentiate physiological adaptations, calculate the trend change rate, and adaptively adjust the dynamic threshold baseline; Identify pathological trends, recalculate the threshold adjustment range, and iteratively update the dynamic threshold baseline; Integrate threshold results from all states and trends to construct an adaptive individualized baseline; The trend recognition module is used to identify abnormal VTE trends in key parameters and dynamically set the duration of the sliding window. Calculate the VTE trend slope and VTE product deviation of the parameters within the sliding window, and construct trend constraints and product deviation constraints. When the parameters do not satisfy both trend constraints and product deviation constraints and the continuous offset time exceeds [a certain value], [further action is taken]. When this occurs, it is determined to be a continuous abnormal offset; The time-series risk prediction module is used to calculate the deviation values ​​of various VTE parameters of the target patient, and constructs a spatiotemporal fusion feature matrix by combining time-series features and static features as input to build a time-series risk prediction model and output the VTE risk probability. The project constructs a time-series risk prediction model, which includes an input layer, an LSTM feature extraction layer, an attention weighting layer, and an output layer. The input layer receives the spatiotemporal fusion feature matrix, normalizes it, and adapts it to the length of the bias time series; The LSTM feature extraction layer introduces a temporal attention mechanism to improve the two-layer bidirectional LSTM structure. The first layer extracts the first feature of the target patient, and the second layer extracts the second feature of the target patient. Residual connections are used to obtain residual fusion features, and linear transformation is used to generate LSTM feature sequences. The attention weighting layer calculates the temporal attention weight and the feature attention weight, and combines them with the LSTM feature sequence to obtain the key temporal feature vector through weighted fusion; The output layer maps the key temporal feature vectors to risk probability values; The early warning and intervention module is used to make dual judgments and interventions based on parameter trend changes and VTE risk probability, set three levels of early warning and intervention measures, monitor VTE parameter changes in real time after intervention, and calculate the improvement rate to verify the prevention and control effect.

2. The VTE real-time monitoring and intelligent prevention system according to claim 1, characterized in that, The specific steps for establishing an initial individualized baseline include: The VTE parameters are preprocessed to obtain a VTE parameter set, which is divided into high-frequency dynamic parameters and low-frequency static parameters. The missing time point data of low-frequency static parameters are supplemented by the timestamps of high-frequency dynamic parameters to obtain the classification parameter sequence; The mean values ​​of each parameter are calculated based on the classification parameter sequence, and the correction coefficient and the basic parameter values ​​are combined to generate a correction parameter baseline value sequence. A differentiated window segmentation strategy is adopted to calculate the moving mean and standard deviation of the parameters within each sliding window, and the initial fluctuation range is determined with the benchmark values ​​of each correction parameter as the center. Calculate the correlation degree between high-frequency dynamic parameters, filter strongly correlated parameter pairs according to a preset correlation threshold, and obtain the high-frequency correction interval by correcting the initial fluctuation interval based on the correlation direction. If the low-frequency static parameter value is lower than the preset deviation threshold, the associated vital sign interval is contracted upward; otherwise, the high-frequency correction interval is maintained to obtain the correction fluctuation interval. The revised fluctuation range is updated to form an initial individualized baseline.

3. The VTE real-time monitoring and intelligent prevention system according to claim 2, characterized in that, The trend recognition module includes a trend analysis unit and an anomaly recognition unit; The trend analysis unit is used for real-time dynamic monitoring and feature extraction of key parameters. Based on the adaptive individualized baseline, a trend-sensitive detection algorithm is introduced to identify VTE abnormal trends, and the sliding window duration is dynamically set. Within each sliding window, calculate the VTE trend slope and VTE product bias of the parameters; The anomaly identification unit is used to receive the VTE trend slope and the VTE product skewness, construct trend constraints and product skewness constraints, and when the parameter changes within the sliding window do not simultaneously satisfy the trend constraints and product skewness constraints, and the continuous offset time exceeds [a certain value], [the anomaly identification unit will detect the anomaly]. When this occurs, it is determined to be a continuous abnormal offset, and its temporal characteristics are recorded; Otherwise, it is considered a normal fluctuation within the physiological range, and monitoring continues.

4. The VTE real-time monitoring and intelligent prevention system according to claim 3, characterized in that, The specific steps for identifying abnormal VTE trends include: Set the base duration of the sliding window for key parameters in real-time monitoring; statistics Calculate the baseline fluctuation coefficient using adaptive individualized baseline fluctuation data from the past few days. ; Set maximum fluctuation limit With minimum fluctuation limit ; like If the fluctuation is severe, the duration of the sliding window is determined based on the base duration and a preset first weight. ; like If the fluctuation is stable, the sliding window duration is determined based on the base duration and a preset second weight. ; Using the timestamp within the sliding window as the independent variable and the parameter value as the dependent variable, and setting the time decay weight, the VTE trend slope is obtained by fitting using the weighted least squares method. The difference between the parameter value that exceeds the normal range of the adaptive individualized baseline and the value in the normal range is calculated, and then weighted and summed in combination with the time decay weight to obtain the VTE product bias.

5. The VTE real-time monitoring and intelligent prevention system according to claim 4, characterized in that, The specific steps for identifying abnormal VTE trends also include: Set slope limits; A trend constraint is constructed. If the slope of the VTE trend is less than the slope limit, the trend is determined to be normal; otherwise, it is determined to be a suspected abnormal trend. Set the product bias constraint value; Construct an integral bias constraint. If the VTE integral bias is less than the integral bias constraint value, the offset is determined to be normal; otherwise, it is determined to be a suspected offset anomaly. If both the trend constraint and the product deviation constraint are satisfied simultaneously, then the fluctuation is determined to be normal. If only one constraint is satisfied, it is marked as to be observed, and high-density monitoring mode is started; If neither the trend constraint nor the product deviation constraint is satisfied, then a consistency verification is performed. If the continuous offset duration exceeds [a certain value], [the verification will be performed]. If the previous sliding window is the one to be observed, it is determined to be a continuous abnormal offset, and the temporal characteristics are recorded. Otherwise, it is judged as an abnormal instantaneous fluctuation, and the monitoring continues in the next sliding window.

6. The VTE real-time monitoring and intelligent prevention system according to claim 5, characterized in that, The specific steps for outputting the VTE risk probability include: For each parameter value within a sliding window, calculate the absolute deviation between the parameter value and the baseline mean, and combine this with the normal range of the adaptive individualized baseline to calculate the relative deviation percentage. When the relative deviation percentage exceeds the upper limit of the normal range of the adaptive individualized baseline or falls below the lower limit, an abnormal deviation is marked. By integrating the parameter values ​​of all marked abnormal deviations, a parameter deviation sequence is obtained; For a single parameter, the duration of a single missing event is less than the time. For missing data, the weighted average of adjacent valid data is used to fill in the missing data; For single parameter consecutive missing duration exceeding time The data segment is marked with a missing data identifier; The final bias time series is obtained; The deviation time series is segmented according to a preset physiological state cycle, and key dynamic fluctuation features of each time period are extracted. By aligning the key dynamic fluctuation features, the temporal features, and the static features in the time dimension, a spatiotemporal fusion feature matrix is ​​constructed.

7. The VTE real-time monitoring and intelligent prevention system according to claim 6, characterized in that, The specific steps for outputting the VTE risk probability also include: Historical VTE labeled data was used as the training set, weighted Focal Loss was used as the loss function, and the AdamW optimizer was used to train the time series risk prediction model. Input the spatiotemporal fusion feature matrix of the target patient into the trained temporal risk prediction model, and output the future... Hourly VTE initial risk probability; Dynamic calibration by calling upon historical clinical rule bases, outputting future results Hourly VTE risk probability.

8. The VTE real-time monitoring and intelligent prevention system according to claim 7, characterized in that, The specific steps for conducting dual-assessment intervention include: Calculate the deviation score of each parameter trend indicator from the preset normal trend range, and weight and merge the deviation scores of all parameter trend indicators to obtain the trend risk score; Set trend score weights and VTE risk probability weights, and calculate a dual risk score by weighted summation based on the VTE risk probability. ; Set the maximum threshold With minimum threshold ; Construct dual risk assessment conditions, if If so, it is classified as a Level 1 risk; if If so, it is judged as a level two risk; if If so, it is classified as a level three risk.

9. A VTE real-time monitoring and intelligent prevention system according to claim 8, characterized in that, The specific steps for conducting dual-assessment intervention also include: After the intervention measures are implemented at the nursing end, the implementation time, the content of the measures, and the implementer are recorded in real time. Healthcare workers can confirm the implementation status on their mobile devices and provide additional suggestions for adjustments. By comparing before and after the implementation of intervention measures The improvement rate is calculated based on the change in the dual risk score within one hour; If the improvement rate is greater than or equal to the preset improvement threshold, the prevention and control effect is deemed effective, and the intervention plan is recorded as an effective measure. Otherwise, if the risk is deemed not to have been effectively controlled, the intervention measures will be adjusted and the risk level will be reassessed. If the improvement rate still does not meet the standard after adjustment, a multidisciplinary assessment process will be triggered.

Images