Bleeding risk prediction method based on the fusion of stress feedback system and clinical variables

Abstract

This invention discloses a bleeding risk prediction method based on the fusion of a pressure feedback system and clinical variables, belonging to the field of medical information processing and intelligent prediction technology. It includes collecting dynamic physiological parameters of the puncture site using an intelligent pressure feedback system, acquiring multi-dimensional static clinical variable data, performing multi-source heterogeneous data preprocessing and spatiotemporal alignment on the dynamic physiological parameters and static clinical variables, constructing a standardized input vector, building a structurally optimized improved width-learning system bleeding risk prediction model, training the model, inputting the standardized input vector into the trained improved width-learning system bleeding risk prediction model, outputting the bleeding risk prediction probability, and performing risk stratification and early warning based on the bleeding risk prediction probability. This invention can effectively improve the accuracy of post-interventional bleeding risk prediction and is easy to use.

Description

Technical Field

[0001] This invention relates to the field of medical information processing and intelligent prediction technology, specifically to a bleeding risk prediction method based on the fusion of stress feedback systems and clinical variables. Background Technology

[0002] Post-interventional procedures often result in complications such as bleeding at the puncture site. These complications are caused by a variety of factors, including individual patient characteristics, intraoperative procedures, and postoperative management, leading to significant individual variability in the clinical process. Currently, standardized compression hemostasis and postoperative observation protocols are commonly used in clinical practice. However, there is a lack of forward-looking tools to predict individual bleeding risks, making it difficult to achieve precise prevention and control through early intervention.

[0003] Some studies have attempted to construct bleeding risk assessment tools, mainly based on traditional statistical methods such as logistic regression. The variables included are mostly static clinical indicators, such as age, coagulation function, and sheath size. However, these methods have the following shortcomings: First, the variable sources are singular, failing to integrate continuously monitored parameters that dynamically reflect the real-time physiological regulatory capacity under vascular-tissue compression, thus limiting the predictive dimensions; second, the model structures are mostly linear, making it difficult to effectively capture the complex nonlinear relationships that objectively exist between variables; and third, a modeling method that systematically integrates dynamic pressure and continuous perfusion data during the compression period with multi-scale static clinical variables has not yet been developed.

[0004] Furthermore, while existing intelligent pressure monitoring devices can collect local pressure and perfusion signals, their data is typically only used for manual assessment of pressure endpoints, lacking a channel for regular and calculable integration with clinical risk variables. Therefore, how to organically combine the high-frequency, multi-source dynamic signals collected by intelligent pressure feedback systems with clinical static information to construct a risk prediction model with nonlinear expression capabilities and adaptability to individual differences has become a crucial technical problem that urgently needs to be solved to improve postoperative bleeding control. Summary of the Invention

[0005] The purpose of this invention is to provide a bleeding risk prediction method based on the fusion of a pressure feedback system and clinical variables, which can effectively improve the accuracy of bleeding risk prediction after interventional procedures and is easy to use.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A bleeding risk prediction method based on the fusion of a stress feedback system and clinical variables includes: Step 1: Collect dynamic physiological parameters of the puncture site based on the intelligent pressure feedback system; Step 2: Obtain multidimensional static clinical variable data; Step 3: Perform multi-source heterogeneous data preprocessing and spatiotemporal alignment on dynamic physiological parameters and static clinical variables to construct standardized input vectors; Step 4: Construct and train an improved width learning system bleeding risk prediction model with optimized structure; Step 5: Input the standardized input vector into the trained improved width learning system bleeding risk prediction model, output the bleeding risk prediction probability, and perform risk stratification and early warning based on the bleeding risk prediction probability.

[0007] Furthermore, in step 1, the dynamic physiological parameters include the mean radial artery pressure and the dynamic change coefficient of tissue perfusion rate.

[0008] Furthermore, in step 1, the intelligent pressure feedback system consists of a wearable pressure sensor array, a tissue perfusion rate sensor module, a signal conditioning circuit, an ASIC chip, and a microcontroller. The ASIC chip integrates a feature extraction acceleration engine, which is used to perform on-chip real-time calculations on the pressure and perfusion signals after analog-to-digital conversion, extract the mean radial artery pressure and the dynamic change coefficient of tissue perfusion rate, and transmit the extraction results to the microcontroller through the SPI interface.

[0009] Furthermore, the mean radial artery pressure is calculated by the feature extraction acceleration engine of the ASIC chip using a sliding window averaging of the spatial mean values ​​of multiple wearable pressure sensor arrays.

[0010] Furthermore, the dynamic variation coefficient of the tissue perfusion rate is obtained by the feature extraction acceleration engine of the ASIC chip calculating the perfusion index ratio based on the dual-wavelength photoplethysmography signal of the tissue perfusion rate sensing module, and calculating the coefficient of variation of the perfusion index ratio within a sliding window.

[0011] Furthermore, in step 2, the multidimensional static clinical variable data includes general patient information, surgery-related information, and preoperative laboratory indicators.

[0012] Furthermore, in step 3, multi-source heterogeneous data preprocessing and spatiotemporal alignment are performed on dynamic physiological parameters and static clinical variables to construct standardized input vectors, specifically as follows: Missing values ​​in multidimensional static clinical variable data were imputed using multiple imputation methods. Moving average filtering is used to reduce noise in dynamic physiological parameters; Using the decompression time at the puncture point as the time origin, a unified observation window is set. Within the observation window, the dynamic physiological parameters after noise reduction are linearly interpolated and resampled at a fixed sampling frequency to obtain a dynamic feature vector of fixed length. The imputed multidimensional static clinical variable data is concatenated with the resampled dynamic physiological parameters, and the continuous variables are standardized to obtain a standardized input vector.

[0013] Further, in step 4, an improved width-learning system bleeding risk prediction model with optimized structure is constructed and trained, specifically as follows: Based on the standard width learning system, a channel attention mechanism is introduced between the feature node layer and the enhancement node layer to adaptively and dynamically weight each group of feature nodes. Furthermore, a random weight matrix that has undergone orthogonal normalization is used when generating enhancement nodes to construct an improved width learning system bleeding risk prediction model. During model training, the input vector of the preset training set is sequentially mapped through feature nodes, weighted by attention, and mapped through augmentation nodes to form an extended input matrix. Ridge regression with L2 regularization is used to solve the output weight matrix, resulting in a trained improved width learning system bleeding risk prediction model.

[0014] In summary, the beneficial technical effects of the present invention include: First, the ASIC chip integrated into the intelligent pressure feedback system enables high-frequency acquisition of multi-channel pressure and perfusion signals and on-chip feature extraction, ensuring the real-time and consistency of dynamic physiological parameter acquisition and providing high-precision dynamic input for the model.

[0015] Second, by spatiotemporally aligning and fusing continuous features that reflect the dynamic regulatory capacity of blood vessels and tissues during the compression period, such as mean radial artery pressure and dynamic change coefficient of tissue perfusion rate, with multidimensional static clinical variables, we have overcome the limitation of relying solely on baseline static indicators and made the model input information more complete.

[0016] Third, based on the standard width learning system, targeted structural optimization is carried out. A channel attention mechanism is introduced between the feature node layer and the enhancement node layer, which enables the model to automatically learn and recalibrate the importance of each group of feature nodes, suppress feature subspaces that contribute little to the prediction, and enhance the discriminative ability. At the same time, orthogonal normalization is used when generating enhancement nodes to effectively reduce redundancy and collinearity caused by random mapping and improve model stability.

[0017] Fourth, the output weights are analytically solved using ridge regression with L2 regularization, eliminating the need for iterations during training, resulting in high computational efficiency and facilitating rapid model deployment and subsequent incremental updates. Applying this improved width learning system to bleeding risk prediction yields more accurate individualized probability outputs and directly drives differentiated clinical intervention strategies through a hierarchical early warning mechanism, contributing to improved post-interventional management quality and patient safety. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] like Figure 1 As shown, this invention provides a bleeding risk prediction method based on the fusion of a pressure feedback system and clinical variables, comprising: Step 1: Collect dynamic physiological parameters of the puncture site based on the intelligent pressure feedback system; Step 2: Obtain multidimensional static clinical variable data; Step 3: Perform multi-source heterogeneous data preprocessing and spatiotemporal alignment on dynamic physiological parameters and static clinical variables to construct standardized input vectors; Step 4: Construct and train an improved width learning system bleeding risk prediction model with optimized structure; Step 5: Input the standardized input vector into the trained improved width learning system bleeding risk prediction model, output the bleeding risk prediction probability, and perform risk stratification and early warning based on the bleeding risk prediction probability.

[0021] Next, we will elaborate on the above steps in detail, taking into account the specific parameter settings.

[0022] In step 1, dynamic physiological parameters of the puncture site are collected based on the intelligent pressure feedback system, specifically: First, the intelligent pressure feedback system will be described. This system is used to acquire dynamic physiological parameters of the puncture site of patients after interventional procedures with high precision, and to perform on-chip preprocessing and feature extraction of the data. The system includes a wearable pressure sensor array, a tissue perfusion rate sensing module, a signal conditioning circuit, a performance-integrated ASIC chip, and a microcontroller. Each component will be described in detail below: The wearable pressure sensor array consists of multiple flexible sensing units based on the piezoresistive effect. These units are evenly spaced on a flexible polyimide substrate and encapsulated as a rectangular patch. The patch integrates [missing information - likely related to sensors or components]. One pressure sensing unit, The value is an even number, typically 8 or 16. When each pressure sensing unit is subjected to normal stress perpendicular to the patch plane transmitted from the blood vessel-tissue interface, its resistance changes approximately linearly. This resistance change is converted into an analog voltage signal output via a constant current source. The mathematical expression for the output signal of the pressure sensing unit is: ; in, For the first Each pressure sensing unit at time The output analog voltage value, To apply a constant bias current, This is the intrinsic resistance value of the sensing unit under zero pressure. This sensing unit is subjected to interstitial pressure transmitted by the radial artery. The resulting change in resistance satisfies , For the first The piezoresistive coefficient of each sensing unit.

[0023] The tissue perfusion rate sensing module uses a near-infrared spectral microcirculation perfusion measurement probe, which includes a pair of probes spaced at a distance of [missing information]. The light source and photodetector, the light source emission wavelengths are respectively Near-infrared light penetrates skin tissue; some of the light is absorbed by oxyhemoglobin and deoxyhemoglobin, while the rest is scattered back and received by a photodetector. The photodetector converts the intensity of the received scattered light into a photocurrent. The photocurrent is processed by subsequent circuitry to calculate tissue oxygen saturation and blood perfusion index. The tissue perfusion rate sensing module outputs two time-varying analog voltages, corresponding to the AC and DC components of the photoplethysmography signal at two wavelengths, respectively. The AC component... The direct current component reflects the time-varying blood volume caused by arterial pulsation. Reflects the non-pulsatile static components of the tissue. The raw perfusion rate signal can be expressed as: ; This ratio, after system calibration, can be converted into the standard perfusion rate unit PU, which characterizes the blood flow per minute per 100 grams of tissue in the puncture area.

[0024] The signal conditioning circuit is located between the pressure sensing array and the perfusion rate sensing module and the ASIC chip, and it includes... Each pressure sensing unit corresponds to The system includes a parallel-connected instrumentation amplifier channel and two transimpedance amplifier channels corresponding to the two outputs of the injection rate sensing module. For any pressure sensing unit, its output... First, connect an instrumentation amplifier constructed from a precision operational amplifier, which has a rated gain of [missing information]. Fixed at 100, this amplifies millivolt-level pressure signals to volt levels, while its high common-mode rejection ratio effectively suppresses power frequency interference. The amplified signal passes through a second-order Bessel low-pass filter with a cutoff frequency set to [value missing]. To eliminate muscle tremors and high-frequency electromagnetic noise above the cutoff frequency, a shaped pressure signal is obtained. For each photoplethysmography signal of the tissue perfusion rate sensing module, the photocurrent of the photodetector is first converted into a voltage signal using a transimpedance amplifier, and the gain resistor... A 100kΩ filter was selected to ensure a moderate output signal amplitude. The signal was then passed through a second-order Bessel low-pass filter, but with its cutoff frequency set to... This is to accommodate the low dynamic range of the injected signal. All the filtered analog signals are then sent to the analog front-end path of the ASIC chip in a multi-parallel manner.

[0025] The ASIC chip is the core processing and computing component of this intelligent pressure feedback system. It is fabricated using TSMC's 180nm CMOS process. Its internal architecture is functionally divided into five parts: analog front-end array, multi-channel analog-to-digital converter, digital control logic, feature extraction acceleration engine, and SPI interface.

[0026] The analog front-end array integrates a programmable gain amplifier and sample-and-hold circuitry, which can be matched to the signal received from the signal conditioning circuitry. Analog signals. During the initial power-up phase of the system, the digital control logic configures the gain of each PGA channel via internal registers to a uniform value. Furthermore, the sample-and-hold circuit is synchronously controlled to ensure that the signals of each channel are strictly captured synchronously in the time domain.

[0027] The analog-to-digital converter (ADC) employs a 12-bit successive approximation architecture, with a maximum sampling rate of 200kSPS. All analog channels share the ADC via an on-chip analog multiplexer. Channel switching timing is determined by a finite state machine in the digital control logic with a fixed polling period. The time point for managing and completing a round of channel conversion is... , This is a time-based indicator. For any given channel... At any moment The quantization output is: ; in This represents the analog voltage value maintained by the channel at this moment. and This is the reference voltage for the analog-to-digital converter. Indicates rounding to the nearest integer. The quantized numeric code. It is cached in the on-chip dual-port SRAM.

[0028] The feature extraction acceleration engine is a key module for realizing on-chip real-time computing. It uses a hard-wired state machine and a parallel multiply-accumulate unit array to extract the following two high-frequency features, avoiding the communication bandwidth bottleneck and power consumption caused by transmitting all the original high sampling rate data to the off-chip microcontroller.

[0029] The first characteristic is the mean radial artery pressure. This feature is based on The spatial domain fusion and time domain mean calculation of the road pressure channel are used to generate the signal. For each sampling time... First calculate Spatial mean of pressure sensing units: ; in, For the first Each pressure channel at time ADC output code, This is the zero-bit code measured when no pressure is applied to the channel. This is the sensitivity coefficient for that channel, expressed in codes per millimeter of mercury. This coefficient is pre-stored in the one-time programmable memory of the ASIC chip during factory calibration. Subsequently, the feature extraction acceleration engine processes the data in a [length missing] [area missing]. Calculation within the sliding window The moving average, as the value at the output update time. Characteristic values ​​of mean radial artery pressure: ; in, This is the feature output interval, which can be 1 second. For those contained within the window The number of sample points depends on the ADC polling cycle and Sure.

[0030] The second characteristic is the dynamic variation coefficient of tissue perfusion rate. Extract the two wavelength signals corresponding to the self-organized perfusion rate sensing module. The ASIC chip first... , , , Real-time calculation of perfusion index ratio based on ADC conversion results: ; in .

[0031] To avoid motion artifacts and transient interference, the feature extraction acceleration engine... The sequence is subjected to exponentially weighted moving filter, and its coefficient of variation within the observation window is calculated as the dynamic change coefficient of perfusion rate: ; in, and These represent the mean and standard deviation of the filtered PI sequence within the window, respectively, and the window length is... The time span of the sliding window used in the calculation remains consistent. One sampling point.

[0032] Based on the above calculations, at each feature output time... The feature extraction acceleration engine will calculate a set of dynamic physiological parameters. The corresponding timestamp is latched into the data output register.

[0033] As the central scheduling unit of the ASIC chip, the digital control logic (DCLogic) internally contains a precise timing sequence generator responsible for generating sample-and-hold signals, ADC start-up signals, RAM read / write control signals, and enable signals for the feature extraction acceleration engine, ensuring the orderly and coordinated operation of all modules. Simultaneously, the DCLogic responds to configuration commands from the microcontroller, receiving and parsing command words via the SPI interface to adjust PGA gain, polling cycle, and sliding window length. Dynamic adjustment of parameters such as...

[0034] The SPI interface module is used to realize full-duplex high-speed data exchange between the ASIC chip and the off-chip microcontroller. At feature output time, the ASIC chip converts the feature value in the data output register into an SPI data frame in parallel. The frame format includes a start bit, feature type identifier, data field, and cyclic redundancy check (CRC) code. The microcontroller, acting as the master, periodically reads this frame data through the SPI interface, packages it, and transmits it transparently to the host computer or local data processing terminal via USB or Bluetooth Low Energy.

[0035] The microcontroller uses a low-power chip based on the ARM Cortex-M4 architecture, and runs embedded software internally to perform the following functions: I. Obtaining the output of the ASIC chip via the SPI interface and The characteristic data is stored in the on-chip FIFO buffer. 2. Connect to the real-time clock module to add a system-wide timestamp accurate to milliseconds to each data record; Third, the time-stamped dynamic physiological parameter data stream is sent to the computing platform responsible for subsequent processing via USB virtual serial port or BLE 5.0 protocol, which integrates static clinical variables and executes subsequent steps.

[0036] In step 2, multidimensional static clinical variable data are obtained, specifically as follows: The data for static clinical variables come from two sources: one is objective records from the hospital's electronic medical record system, and the other is standardized data collection from a self-developed questionnaire on influencing factors of post-interventional hemorrhagic complications. The following provides a detailed explanation of the variable acquisition content, operating procedures, and data format for both sources: The first part, the variables extracted from the hospital's electronic medical record system, includes three main categories: The first category is general patient information, which specifically includes: The patient's date of birth is used to calculate the precise age; the age variable is denoted as... The unit is years, accurate to one decimal place.

[0037] Gender, denoted as Categorical variables, coded as 0 for female and 1 for male.

[0038] Height and weight are recorded as follows: and The Body Mass Index (BMI) is calculated using units of centimeters and kilograms, respectively. The formula is as follows: The result is rounded to one decimal place.

[0039] Smoking history, recorded as Categorical variables are coded as 0 for no smoking history and 1 for a smoking history.

[0040] Past medical history mainly extracts diagnostic information related to bleeding risk, including: ① history of hypertension, recorded as Classification code 0 for none, 1 for present; ② History of diabetes, recorded as Classification code 0 indicates none, 1 indicates yes; ③ History of coronary atherosclerotic heart disease, recorded as Classification code 0 indicates none, 1 indicates present; ④ History of chronic kidney disease, recorded as The classification code is 0 for none and 1 for yes.

[0041] Preoperative history of anticoagulation or antiplatelet drug use is recorded as follows: Categorical variables, where 0 represents no use, 1 represents use of antiplatelet drugs only, 2 represents use of anticoagulants only, and 3 represents use of both drugs in combination.

[0042] The second category is surgical-related information, specifically including: Interventional procedure type, denoted as Categorical variables, coded as follows: 1 for coronary angiography, 2 for coronary interventional therapy, 3 for neurovascular intervention, 4 for peripheral vascular intervention, and 5 for other.

[0043] Whether the first puncture was successful is recorded as follows: The classification code is 0 for failure and 1 for success.

[0044] Sheath size, denoted as The value is in French and refers to the nominal outer diameter of the sheath used.

[0045] The duration of the operation is recorded as follows: , defined as the time interval from successful arterial puncture to sheath removal, in minutes, accurate to the nearest whole minute.

[0046] The method of anesthesia is denoted as The classification code 0 represents local anesthesia and 1 represents local anesthesia combined with sedation.

[0047] Intraoperative anticoagulation regimen: The generic names, single doses, and frequency of administration of the anticoagulants used during the procedure are collected to calculate the intraoperative anticoagulant strength index. The index is defined as follows: using the anticoagulant activity of unfractionated heparin as the equivalent baseline, the total number of equivalent heparin units for all anticoagulants used during surgery is converted according to the "Conversion Table of Equivalent Doses of Anticoagulants," and then divided by the patient's weight to obtain the number of equivalent heparin units per kilogram of body weight, denoted as _____. The unit is IU / kg.

[0048] The third category is preoperative laboratory indicators, extracted from the most recent preoperative test report, specifically including: Prothrombin time, denoted as The unit is seconds.

[0049] Activation time of partial thromboplastin, denoted as The unit is seconds.

[0050] International Normalized Ratio (INR), denoted as , dimensionless.

[0051] Platelet count, denoted as The unit is 10^9 / L.

[0052] Glycated hemoglobin, denoted as , expressed as a percentage.

[0053] Serum creatinine value, denoted as The unit is μmol / L, used to estimate glomerular filtration rate, and its calculation formula follows the formula of the Chronic Kidney Disease Epidemiology Collaboration Group: ,in The value is 0.9 for males and 0.7 for females. The value is -0.411 for males and -0.329 for females.

[0054] All variables extracted from the electronic medical record system were independently retrieved by researcher A, who had received standardized training, through the hospital information system query interface, and recorded in the first sub-table of the standardized data extraction table. Access permissions and extraction periods for the medical record system must cover the entire perioperative period for each enrolled patient.

[0055] The second part, variables collected from the questionnaire on influencing factors of post-interventional hemorrhagic complications, were obtained through structured interviews and bedside observations, and specifically included the following variables: Postoperative bed rest compliance, denoted as Categorical variables were observed and recorded by researchers from immediately after surgery until 2 hours after decompression. Based on whether patients could strictly maintain immobilization of the operated limb and supine position, 0 was used to indicate complete compliance, 1 to indicate partial resistance, and 2 to indicate obvious defiance.

[0056] Postoperative agitation was assessed using the Richmond Agitation-Sedation Scale every 30 minutes within 6 hours postoperatively, and the highest value was recorded as the variable value. , is an integer-type hierarchical variable, ranging from -5 to +4. Positive postoperative agitation is defined as... Based on this, derived binary variables are generated. The code 0 represents no agitation and 1 represents agitation.

[0057] Postoperative pain was assessed hourly for the first 6 hours postoperatively using a numerical scoring method, and the highest score was recorded. , which is an integer from 0 to 10.

[0058] The duration of manual compression before decompression, and the intervention time difference from the start of sheath removal to the activation of the intelligent pressure feedback system, were recorded in real-time by the supervising nurse in the quality control form and denoted as follows: The unit is minutes.

[0059] The time interval from sheath removal to the first spontaneous urination after surgery is denoted as: The unit is minutes. If no urine is urinated within 6 hours after the operation, it is recorded as 360 minutes.

[0060] The questionnaire also included additional entries to correct for other variables, such as a consistency check mark between the actual intraoperative anticoagulant use details and the medical record, to improve data accuracy. The variable collection process was carried out by two researchers, B and C, with researcher B responsible for bedside assessment and interviews, and researcher C responsible for data entry and logical checks in a separate location, as well as reviewing the completeness of each questionnaire. Any missing items were immediately triggering a supplementary data collection process. Ultimately, all static clinical variables for each enrolled patient were integrated into a structured record, assigned a unique patient number, and stored as a comma-separated table file.

[0061] In step 3, multi-source heterogeneous data preprocessing and spatiotemporal alignment are performed on dynamic physiological parameters and static clinical variables to construct standardized input vectors, specifically as follows: The dynamic physiological parameter time series provided in Step 1 and the static clinical variables provided in Step 2 are integrated, cleaned, and structured to generate a standardized sample dataset. This step specifically includes five sub-steps: missing value handling and noise reduction, outcome label definition, unified resampling of dynamic parameters, input vector construction, and dataset partitioning. These are described in detail below: Step 301: Missing Value Handling and Noise Reduction First, construct a summary table of the original data. To accommodate all All variables for each enrolled patient. For the first... For each patient, the static clinical variables from step 2 constitute a dimension. row vectors If a certain static variable If a value is missing, multiple imputation is used to fill it in. Multiple imputation uses other complete variables as predictors and iteratively generates five imputation values ​​through a chain equation, taking the median as the final imputed value. The imputed static vector is denoted as . .

[0062] For the dynamic parameter sequence generated in step 1, the first... Two original equidistant time series were obtained from each patient: radial artery mean pressure sequence. With the dynamic change coefficient sequence of tissue perfusion rate ,in , Take from 1 , This represents the number of valid data collection points for this patient. To reduce high-frequency measurement noise and occasional motion artifacts, a moving average filter was used to smooth each sequence. The filter window width was set to 5 sampling points. The smoothed sequence is retained and recorded as follows: .

[0063] Step 302: Ending Tag Definition After decompression at the puncture point Whether a bleeding event occurred within an hour was used as a binary outcome label. Decompression time Defined as the timestamp corresponding to the first data frame recorded by the intelligent pressure feedback system at the start of applying controllable pressure. The observation window cutoff time is... ,Pick Hours. Bleeding events are strictly defined using a combined imaging and clinical standard: a pseudoaneurysm or hematoma at the puncture site confirmed by color Doppler ultrasound; or subcutaneous hematoma ≥3cm in diameter; or bleeding after decompression lasting >30 minutes or the appearance of fresh blood oozing. All bleeding events are assessed by an independent vascular surgeon unaware of the patient's dynamic parameter acquisition process within an observation window and are recorded in a structured manner in the electronic medical record system. If a bleeding event is determined to have occurred, then... ,otherwise .

[0064] Step 303: Dynamic Parameter Unified Resampling Because the monitoring duration and start and end times of data collection vary among different patients, the length of each raw sequence is... The sequences are inconsistent, and their start and end points are not strictly aligned to a uniform time zero. To construct a fixed-dimensional input vector, resampling and truncation operations are performed on the smoothed dynamic sequence for each patient.

[0065] Establish a unified reference time frame based on each patient's stress relief moments. Set the time origin. Select the observation window. As part of the analysis period, a uniform resampling frequency is set. ,For example That is, one sampling point per minute. Therefore, each dynamic feature is converted to a fixed length during the observation period. The sequence. In this invention minute, ,but One sampling point.

[0066] right Within this window Resampling is performed. If the time points of the original sequence are not perfectly aligned, linear interpolation is used to calculate the value at each resampled grid point. Finally, the... The mean radial artery pressure characteristics of each patient were converted into a length of [missing value]. vector ,in Indicates the number of decompression cycles. Mean radial artery pressure at each resampling time. Similarly, the sequence of dynamic change coefficients of tissue perfusion rate is transformed into a vector. .

[0067] Therefore, the first All dynamic characteristics of each patient were integrated into a single dimension. row vectors This dimension represents the size of the dynamic portion of the input, where... .

[0068] Step 304: Input Vector Construction and Data Standardization The preprocessed static vector With dynamic vectors Join them horizontally to form the first... The complete original input vector of each sample The total dimension of this vector is ,in, The final number of static clinical variables included in step 2 can be determined to be 22 based on the actual questionnaire items, but the count should be explicitly specified in the implementation; in this step, it is initially represented by symbols. express.

[0069] To eliminate the influence of variable dimensions on subsequent model training, continuous variables are standardized using z-scores, with the standardization operation based on the mean and standard deviation of the training set. Categorical and binary variables are retained unchanged using one-hot or numerical encoding. The final standardized input vector is denoted as... Its dimensions are still .

[0070] Labels of all samples With input vector Constructing a complete labeled dataset .

[0071] Step 305: Dataset Partitioning Stratified random sampling is used to sample the dataset. The training set is divided into two parts: 80% and 20%. With the validation set The stratification is based on the ending tag. To ensure that the proportion of positive bleeding events was similar in both groups, the partitioning process was reproducible by generating a fixed random seed.

[0072] In step 4, an improved width-learning system bleeding risk prediction model with optimized structure is constructed and trained, specifically as follows: Based on standardized training set and validation set A bleeding risk probability prediction model was constructed, with a width learning system incorporating an attention mechanism at its core. Hyperparameter optimization and final determination of model parameters were then completed. The number of samples in the training set. This is the number of samples in the validation set. Each input vector... Dimensions The vector concatenation and standardization have been completed in step 304. The following steps will be explained in detail: Step 401: Random mapping and sparsification generation of feature node layers Stack all input vectors in the training set row by row to form the training input matrix. , dimension To extract multiple linear spatial representations of the input data, the following steps are first performed: Group feature nodes, each group contains Each feature node unit, i.e., each mapping output dimension is [number] units. . No. The generation process of group feature nodes is achieved by a random weight matrix and a sparse activation function.

[0073] For the For each feature node, an initial weight matrix is ​​first randomly generated. , dimension Its elements are distributed according to the standard normal distribution. Independent sampling is performed. Bias vectors are also randomly generated. , length is Elements from a uniform distribution Medium sampling. Calculate the linear mapping input matrix: ; in, This represents a column vector with all elements equal to 1 and a dimension of 1. . Matrix dimension is .

[0074] Next, regarding The sparse activation function designed in this invention, based on the L1 norm, is applied to enhance the local discriminative power and anti-overfitting ability of feature nodes. This activation function is defined as a soft thresholding operator, specifically expressed as: ; in, For any element in the input matrix, For symbolic functions, This is a parameter for sparsity adjustment. This function works by considering values ​​less than 1 in the absolute value. The response is set to zero to achieve sparsity in the output, while the response greater than zero is used to achieve sparsity in the output. The portion then contracts towards zero. This preserves the amplitude ordering information of the input signal. The activation function is applied independently to each feature node group to obtain the... Output matrix of group feature nodes : ; All The output matrices of the group of feature nodes are concatenated in the column direction to form the total output matrix of the feature node layer. : ; Step 402: Dynamic weighting driven by channel attention mechanism In traditional width-based learning systems, feature nodes in each group are passed forward with equal contribution, making it difficult to highlight the mapping subspace most critical to the prediction task. Therefore, this invention embeds a channel attention mechanism between the feature node layer and the enhancement node layer, outputting a channel attention mechanism for each group of feature nodes. Implement weighted learning, emphasizing the importance of each subject in relation to its independent learning.

[0075] Attention weights are calculated for each set of feature node matrices. The global average pooling vector is used as input. First, the first... pooling vectors of groups , its first The elements are Mean of this column in the matrix: ; All The pooling vectors are concatenated in sequence to form the output vector of the pooling layer. Then, a lightweight attention network is used to calculate the weight coefficients for each group. The attention network is designed as a two-layer fully connected structure to avoid over-parameterization. Let... This is the first layer weight matrix, with size . ,in To compress dimensions, take To form a bottleneck structure, the activation function is ReLU. Let... This is the second layer weight matrix, with size [value missing]. The activation function is Sigmoid, and the output is normalized using Softmax to obtain the final weight coefficients for each group. .

[0076] Mathematically expressed as: compressed representation vector .

[0077] Original score vector .

[0078] No. Group attention weights The calculation formula is: ; The weights of this group satisfy . use Output matrix for feature node group Perform scalar multiplication and weighting to obtain the weighted eigenvalue matrix: ; The weighted feature matrices are then concatenated to form the weighted feature node layer output matrix. : ; Through the above operations, the model can learn the relative contribution of each group of feature nodes to the final bleeding prediction target end-to-end during the training process, suppress useless feature subspaces, and enhance the discriminative feature subspace.

[0079] Step 403: Generation of Orthogonal Normalization Enhanced Node Layer To significantly improve the nonlinear fitting ability of the model and maintain numerical stability, this invention is based on the weighted feature node layer output. Generate an augmentation node layer. When generating augmentation nodes, apply orthogonal normalization constraints to the random mapping matrix to ensure that each group of augmentation nodes can broadly cover nonlinear manifolds in different directions, avoiding feature redundancy and collinearity that may be caused by random projection.

[0080] For the Grouping enhancement nodes, firstly, an initial weight matrix is ​​randomly generated. , dimension ,in The number of augmentation node units in each group is determined by element sampling from... The matrix can be decomposed using QR decomposition or singular value decomposition to obtain an orthonormal matrix. , making The orthogonalization process is specifically expressed as: for Perform singular value decomposition Take the first left singular vector Column composition Bias vector Still from Random sampling.

[0081] Calculate the augmentation node's first Linear mapping input of the group: ; Subsequently, a nonlinear activation function was applied. This invention uses the hyperbolic tangent function. This ensures that the output values ​​are symmetrically distributed between -1 and 1, facilitating subsequent ridge regression solutions. The output matrix of the group enhancement node is: ; Repeat the above process to generate Group enhancement nodes. All The output matrices of the enhanced nodes are horizontally concatenated to obtain the total output matrix of the enhanced node layer. : ; Step 404: Construct an extended input layer and perform ridge regression. Output the weighted feature node layer With enhanced node layer output Concatenate them along the column direction to form the extended input layer matrix of the width learning system. : make To expand the total dimension of the features, the corresponding training label column vector is: .

[0082] The output weight matrix of this model The objective function is solved using a ridge regression objective function with L2 regularization to avoid the collinearity sensitivity problem encountered when solving for the spurious inverse. The objective function is: ; in, This is the hyperparameter for regularization strength. The problem has an analytical solution: ; in, for The identity matrix is ​​obtained. The solution can be completed in a single matrix operation, eliminating the need for iterative optimization based on gradient descent, thus significantly reducing training time.

[0083] Finally, for any sample input vector (dimension is) The weighted feature node row vectors are calculated using the same forward steps. and enhanced node row vectors The extended input row vector is obtained by concatenation. , dimension The model outputs the predicted probability value of bleeding risk: ; This is a continuous probability value, ranging from 0 to 1. The higher the value, the greater the risk of bleeding complications.

[0084] Step 405: Hyperparameter Joint Optimization To achieve optimal prediction performance, the following hyperparameters need to be jointly optimized: number of feature node groups. Number of feature node units per group Increase the number of node groups Number of enhancement node units per group Sparsity adjustment parameters and regularization strength A grid search strategy is employed on the validation set. The area under the receiver operating characteristic curve (AUC) was used as the evaluation metric, and the combination of hyperparameters that maximized the AUC of the validation set was selected.

[0085] The candidate sets for each hyperparameter in the grid search are set as follows: ; ; ; ; ; ; Traversing hyperparameter combinations For each combination, perform model training in steps 401 to 404, and... Calculate the AUC. Record the hyperparameter combination corresponding to the highest AUC value as the final model configuration.

[0086] With this optimal hyperparameter configuration, the entire training set is reused. Calculate the final output weight matrix The extended input matrix at this time and Parameters are serialized and saved.

[0087] In step 5, the standardized input vector is input into the trained improved width learning system bleeding risk prediction model, and the bleeding risk prediction probability is output. Risk stratification and early warning are then performed based on the bleeding risk prediction probability, specifically as follows: Based on the optimal model of the improved width learning system, trained and serialized in step 4, individualized real-time prediction of bleeding risk is implemented for newly admitted post-interventional patients, and hierarchical early warning and clinical intervention recommendations are triggered according to the predicted probability. This step specifically includes the following sub-steps: Step 501: Acquire dynamic and static data of new samples simultaneously For a newly admitted patient undergoing interventional treatment, after the sheath is removed and the intelligent pressure feedback system is activated, dynamic physiological parameters and static clinical variables are collected simultaneously according to the operating procedures established in steps 1 and 2.

[0088] The acquisition of dynamic physiological parameters follows the complete process of step 1: the wearable pressure sensor array and tissue perfusion rate sensor module continuously sense signals in the puncture area, which are then fed into the ASIC chip after signal conditioning circuitry. The on-chip feature extraction acceleration engine of the ASIC chip calculates the output at each time point online. radial artery mean pressure characteristics Dynamic variation coefficient of tissue perfusion rate It transmits timestamped data streams to the host computing platform one by one via the microcontroller.

[0089] The collection of static clinical variables was performed according to the standardized protocol in step 2: one trained researcher extracted the patient's general information, surgical-related information, and preoperative laboratory indicators from the hospital's electronic medical record system, while another researcher completed an influencing factor questionnaire at the bedside, recording variables such as postoperative positional compliance, agitation score, pain score, and duration of artificial compression before decompression, and then verified to form a complete static vector. .

[0090] Step 502: New Sample Data Preprocessing and Input Vector Construction For new patients, the data preprocessing pipeline defined in step 3 is followed in a completely consistent manner. First, missing fields in the static vector are imputed using multi-imputation model parameters already stored in the deployment environment, resulting in... Next, starting from the decompression moment... Starting from the observation window, the truncated length is [length missing]. Hourly dynamic data fragments, for and Both sequences were subjected to moving average filtering, with the window width remaining constant at 5 sampling points. Subsequently, a fixed frequency identical to that used in the training phase was applied. Linear interpolation resampling is performed on the filtered sequence to form a sequence of lengths... Dynamic feature vectors and .

[0091] The padded static vector and the dynamic vector are horizontally concatenated according to the variable order of the training phase to obtain the original input vector. Use the mean vectors of each continuous variable calculated from the training set in step 3 and which have been persistently stored. with standard deviation vector ,right For continuous variables, z-score standardization is applied, and the calculation formula is as follows: ; in Indexing is used for continuous variables. Standardization yields the final new sample input vector. The dimension remains the same. .

[0092] Step 503: Forward Inference of the Improved Width Learning System Model Will Input the data into the improved width learning system model trained under the optimal hyperparameter combination determined in step 4 to complete the forward computation. Its complete computation graph strictly follows the sequence: The first step involves random mapping and sparse activation of the feature node layer. This utilizes the stored... Group of random weight matrices With bias Calculate the linear mapping results for each group. Then, through the sparse activation function defined in step 4. Calculate the output row vector of the feature node unit The outputs of all feature node groups are concatenated as follows: , dimension .

[0093] The second step is dynamic weighting of channel attention. Using the corresponding position index used during global average pooling, the attention of each group is calculated. The mean of the values ​​forms the pooling vector. Using the parameters of the already trained attention network and Calculate the weights of each group The calculation process is as follows: ; ; ; Obtain the weighted feature node group output The weighted feature layer output is obtained by concatenation. , dimension .

[0094] The third step is to generate orthogonally normalized enhanced nodes. This involves calling the already stored... Group of orthogonal normalized matrices With bias Calculate the linear input of the augmented node The enhanced node output is obtained through the tanh activation function. The output of concatenating all enhanced node groups is , dimension .

[0095] The fourth step is to construct the expanded input and calculate the risk probability. The expanded row vector is obtained by concatenating the weighted features and the augmented nodes. , dimension ,in The final output weight matrix obtained during the training phase is called. (dimension is) ), calculate the predicted probability of bleeding risk: ; Because the model uses ridge regression to directly output real values. It may slightly exceed the [0, 1] interval; therefore, a truncation operation is performed for this: This is to ensure the rationality of the probabilistic interpretation.

[0096] Step 504: Determining the Risk Stratification Threshold Based on a pre-set stratification threshold, continuous probability values ​​are... The risk is mapped to three discrete risk levels. The thresholds are determined through retrospective analysis of the risk distribution on the training set, combined with clinical safety considerations, and are set as follows in this invention: High risk threshold Low risk threshold Both thresholds are determined after model development, based on a combination of the percentiles of the probability output from the training set and the Youden index, and are then incorporated into the inference process. The stratification determination rule is as follows: when It was determined to be at a high-risk level.

[0097] when It was determined to be at a medium risk level.

[0098] when It was determined to be at a low risk level.

[0099] Step 505: Early Warning Signal Generation and Personalized Intervention Plan Push Based on the risk level, the system automatically generates corresponding early warning signals and intervention strategy suggestions, which are then displayed visually and accompanied by voice prompts through the nursing terminal interface.

[0100] For high-risk levels, a red alert is triggered. The system immediately pushes a mandatory blocking pop-up window to the bedside nursing terminal, containing: a clear indication of "high risk of bleeding," the current predicted probability value, and abnormal pressure and perfusion characteristics. Recommended interventions include: immediately extending manual or mechanical pressure for at least 30 minutes; increasing the monitoring frequency of the puncture site after decompression from once every 30 minutes to once every 10 minutes, for at least 2 hours; notifying the attending physician to assess whether bedside ultrasound examination is necessary; and, based on the residual effect of intraoperative anticoagulants, suggesting the use of protamine sulfate or other reversal agents if necessary. After the relevant interventions are implemented, nursing staff confirm the implementation status item by item in the system, forming a closed-loop record.

[0101] For medium-risk levels, a yellow alert is triggered. The system inserts a warning label into the nursing task list and highlights it in yellow. The prompts include: closely monitoring for bleeding or expanding subcutaneous ecchymosis at the puncture site; maintaining routine compression and observing every 30 minutes; instructing the patient to keep the operated limb immobilized and reduce bending activities. If any signs of deterioration appear during subsequent monitoring, the alert can be manually escalated to a red alert.

[0102] For low-risk levels, no warning is triggered, and the system defaults to standardized post-interventional nursing routines, maintaining the puncture site observation frequency of once every hour and routine compression management.

[0103] Step 506: Data Recording and Model Iteration Mechanism The input vector for each new sample Predicting probability The output risk level and the actual outcome (whether a bleeding event occurred) are automatically appended to the system's backend database. When the number of newly accumulated samples reaches a preset update threshold (e.g., 300 cases), the incremental update process of the model can be triggered. Incremental updates do not involve complete retraining; instead, they utilize new data to rapidly expand the input layer of the existing width learning system. Add new feature nodes or augmentation node groups to the matrix and resolve the output weights. The mapping matrix between the original features and the enhanced nodes remains unchanged. This incremental learning mechanism ensures that the model continuously adapts to changes in patient characteristics and treatment patterns during clinical applications, maintaining the long-term stability of predictive performance.

[0104] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0105] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0106] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0107] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A bleeding risk prediction method based on the fusion of stress feedback system and clinical variables, characterized in that, include: Step 1: Collect dynamic physiological parameters of the puncture site based on the intelligent pressure feedback system; Step 2: Obtain multidimensional static clinical variable data; Step 3: Perform multi-source heterogeneous data preprocessing and spatiotemporal alignment on dynamic physiological parameters and static clinical variables to construct standardized input vectors; Step 4: Construct and train an improved width learning system bleeding risk prediction model with optimized structure; Step 5: Input the standardized input vector into the trained improved width learning system bleeding risk prediction model, output the bleeding risk prediction probability, and perform risk stratification and early warning based on the bleeding risk prediction probability.

2. The bleeding risk prediction method based on the fusion of pressure feedback system and clinical variables according to claim 1, characterized in that, In step 1, the dynamic physiological parameters include the mean radial artery pressure and the dynamic change coefficient of tissue perfusion rate.

3. The bleeding risk prediction method based on the fusion of pressure feedback system and clinical variables according to claim 2, characterized in that, In step 1, the intelligent pressure feedback system consists of a wearable pressure sensor array, a tissue perfusion rate sensor module, a signal conditioning circuit, an ASIC chip, and a microcontroller. The ASIC chip integrates a feature extraction acceleration engine, which is used to perform on-chip real-time calculations on the pressure and perfusion signals after analog-to-digital conversion, extract the mean radial artery pressure and the dynamic change coefficient of tissue perfusion rate, and transmit the extraction results to the microcontroller through the SPI interface.

4. The bleeding risk prediction method based on the fusion of pressure feedback system and clinical variables according to claim 3, characterized in that, The mean radial artery pressure is calculated by a sliding window averaging of the spatial mean values ​​of multiple wearable pressure sensor arrays using the feature extraction acceleration engine of the ASIC chip.

5. The bleeding risk prediction method based on the fusion of a pressure feedback system and clinical variables according to claim 4, characterized in that, The dynamic variation coefficient of the tissue perfusion rate is obtained by the feature extraction acceleration engine of the ASIC chip calculating the perfusion index ratio based on the dual-wavelength photoplethysmography signal of the tissue perfusion rate sensing module, and calculating the coefficient of variation of the perfusion index ratio within a sliding window.

6. The bleeding risk prediction method based on the fusion of pressure feedback system and clinical variables according to claim 5, characterized in that, In step 2, the multidimensional static clinical variable data includes general patient information, surgery-related information, and preoperative laboratory indicators.

7. The bleeding risk prediction method based on the fusion of pressure feedback system and clinical variables according to claim 6, characterized in that, In step 3, multi-source heterogeneous data preprocessing and spatiotemporal alignment are performed on dynamic physiological parameters and static clinical variables to construct standardized input vectors, specifically as follows: Missing values ​​in multidimensional static clinical variable data were imputed using multiple imputation methods. Moving average filtering is used to reduce noise in dynamic physiological parameters; Using the decompression time at the puncture point as the time origin, a unified observation window is set. Within the observation window, the dynamic physiological parameters after noise reduction are linearly interpolated and resampled at a fixed sampling frequency to obtain a dynamic feature vector of fixed length. The imputed multidimensional static clinical variable data is concatenated with the resampled dynamic physiological parameters, and the continuous variables are standardized to obtain a standardized input vector.

8. The bleeding risk prediction method based on the fusion of pressure feedback system and clinical variables according to claim 7, characterized in that, In step 4, an improved width-learning system bleeding risk prediction model with optimized structure is constructed and trained, specifically as follows: Based on the standard width learning system, a channel attention mechanism is introduced between the feature node layer and the enhancement node layer to adaptively and dynamically weight each group of feature nodes. Furthermore, a random weight matrix that has undergone orthogonal normalization is used when generating enhancement nodes to construct an improved width learning system bleeding risk prediction model. During model training, the input vector of the preset training set is sequentially mapped through feature nodes, weighted by attention, and mapped through augmentation nodes to form an extended input matrix. Ridge regression with L2 regularization is used to solve the output weight matrix, resulting in a trained improved width learning system bleeding risk prediction model.

Images